Abstracts Track 2025

Nr:	26
Title:	Gelatin-Encapsulated HepG2 Spheroids: A Novel 3D Platform for Enhanced Hepatotoxicity Assessment
Authors:	Hyun Jong Lee
Abstract:	Drug-induced liver injury remains a significant challenge in pharmaceutical development, often leading to late-stage drug failures. Conventional two-dimensional (2D) cell culture models frequently fail to accurately predict clinical outcomes, highlighting the need for more physiologically relevant in vitro systems. This study presents an innovative approach to hepatotoxicity assessment using HepG2 spheroids encapsulated in gelatin methacrylate (GelMA) hydrogels with varying mechanical properties. We developed a 3D culture platform by encapsulating HepG2 spheroids in GelMA hydrogels of 5%, 10%, and 15% concentrations. The mechanical properties of the hydrogels were characterized using rheometry and scanning electron microscopy. The encapsulated spheroids exhibited sustained viability and liver-specific functionality for up to 21 days, significantly outperforming 2D cultures. Gene expression analysis revealed enhanced levels of drug-metabolizing enzymes (CYP3A4, CYP2E1, UGT1A1, and SULT2A1) in the encapsulated spheroids compared to 2D cultures. Notably, the 5% GelMA hydrogel environment showed the most significant increase in gene expression, with up to 8-fold higher levels for some enzymes. The platform's sensitivity to drug-induced toxicity was evaluated using six compounds with varying hepatotoxic potential. The encapsulated spheroids demonstrated increased sensitivity to drug-induced changes in viability compared to 2D cultures, with differences in IC50 values ranging from 1.3-fold to over 13-fold. Importantly, the model successfully identified compounds known to cause drug-induced liver injury in clinical settings, including cases where 2D cultures failed to detect toxicity. This novel 3D spheroid-hydrogel platform offers several advantages over traditional 2D culture systems, including enhanced maintenance of liver-specific functions, improved expression of drug-metabolizing enzymes, increased sensitivity to drug-induced toxicity, and better correlation with known clinical hepatotoxicity. Our findings suggest that this platform could significantly improve the predictive power of preclinical hepatotoxicity screening, potentially reducing late-stage drug failures and animal testing. Future work will focus on incorporating primary human hepatocytes and exploring the platform's potential for personalized toxicity screening. This innovative approach to hepatotoxicity assessment represents a significant step towards more accurate and physiologically relevant in vitro liver models, with the potential to accelerate drug development and improve patient safety.

Nr:	34
Title:	Understanding the Effects of Respiration, Heart Rate, and Blood Pressure on Spinal CSF Flow in a Rodent Model of CSF Obstruction Using in vivo Near-Infrared Imaging
Authors:	Joel Berliner, Zac Penprase, Donald Ogolo, Shinuo Liu, Lynne Bilston, Marcus Stoodley and Sarah Hemley
Abstract:	Cerebrospinal fluid (CSF) obstruction can result from scar tissue and inflammation (arachnoiditis) after spinal cord injury. These changes to the spinal subarachnoid space alter fluid homeostasis within the spinal cord and precipitate the formation of fluid-filled cysts in the parenchyma (post-traumatic syringomyelia). People who live with spinal cord injury often experience dysregulation to normal breathing and require ventilation, as well as changes to heart rate and blood pressure. Recent work has shown that over the normal respiratory cycle shifts in intrathoracic pressure drive CSF along and into the spinal cord, with arterial pulsations exerting less effect. The present study aimed to determine the effects of these variable on spinal CSF flow when there is a CSF obstruction. In Sprague-Dawley rats, physiological parameters were manipulated such that the effects of spontaneous breathing (generating alternating positive and negative intrathoracic pressures), mechanical ventilation (positive intrathoracic pressure only), tachycardia, and hypertension were separately studied. Animals were divided into control or CSF obstruction groups. Both groups received a C7 – T1 laminectomy. The obstruction group additionally received an extradural suture tied around the spinal cord to obstruct CSF flow. To investigate spinal CSF hydrodynamics, intracisternally-infused indocyanine green was recorded in vivo using a Pentero surgical microscope with a near-infrared camera. Tracer deposition along microscopic pathways was assessed by ex vivo epifluorescence imaging using fluorescent ovalbumin. Irrespective of physiological parameter, a CSF obstruction significantly decreased CSF flow in the spinal subarachnoid space. Tachycardia was the only parameter that reduced flow of CSF along the spinal cord in the presence of an obstruction, however this difference abated over time. Ex vivo analysis indicated that in hypertensive rats, a CSF obstruction reduced ovalbumin deposition in the spinal cord. Hypertension was observed to reduce ovalbumin deposition in the spinal cord above, and increase deposition below, the CSF obstruction. Together, these results indicate that acute changes to breathing, heart rate, and blood pressure have spatial- and temporal-dependent effects on flow of CSF. For people with spinal cord injury, tachycardia may initially reduce spinal CSF flow whilst hypertension reduces CSF inflow above and increases inflow below the site of arachnoiditis. Future investigations should examine how these parameters influence CSF flow pathways in post-traumatic syringomyelia.

Nr:	72
Title:	Capturing Stigmatization Language in Clinical Notes of Patients with Opioid Use Disorder in the Emergency Department
Authors:	Samah Jarad, Gail D'Onofrio and Katherine Hawk
Abstract:	Background: Stigma in healthcare can adversely impact quality of care and contribute to health disparities. Stigma can be particularly detrimental for patients with opioid use disorder (OUD) and often implicated as a significant barrier to quality and effective evidence-based care (EBC) including medications for OUD (MOUD) such as buprenorphine or methadone, harm reduction strategies of overdose prevention education, dispensing or providing prescriptions for naloxone, and direct referrals. MOUD have established benefits in reducing mortality and morbidity of OUD, but less than 20% of individuals with OUD accessed these treatments in 2019.9 Our research shows that perceived stigma among ED patients with OUD is an enormous barrier to seeking care, including after an opioid overdose. Importantly, many providers are unaware of their own stigma as well as their personal use of stigmatizing language (SL) both with patients and within the electronic health records (EHR). Data to identify stigma in the EHR is difficult to obtain and tools that might capture the use of SL within healthcare remain limited. The persistent use of stigmatizing terms like addict/abuser continues to widen the treatment gap of OUD.7 The lack of natural language processing (NLP) and informatics tools to quantify SL and EBC language (EBCL), defined as the documentation of offering harm reduction strategies, mentions of MOUD, and referrals, limits research to characterize the extent and effect of stigma in OUD care. These tools will support future targeted interventions to reduce stigma and the use of SL. In our abstract we use standard NLP methods to addressing challenges and capturing stigma related language within ED provider notes. Study Data: We collected retrospective EHR data including 91,107 ED clinical notes pertaining to 22,681 patients who presented to the ED between 2013-2021. An average of 5 notes per patient. In the cohort, 56% patients are males and 44% are females. 70% are White and 15% are Black. A total of 1,979 providers (55% male and 45% female) cared for these patients. 89% were attending physicians and 7% advanced physician assistants. On average, a provider authored 46 ED notes. Methods: We manually composed a dictionary of words used to express stigma and harm reduction language as well as utilizing the National Institute of Drug Abuse (NIDA) published list of recommended terms. Clinicians on the team including Drs. Hawk, and D’Onofrio contributed their expertise contributed to this task. We also gathered language from different publications investigating Stigma. Our NLP Approach focused on using Spacy V.3.0.7, an NLP toolkit. We used string matching and regular expressions to find words/terms of a given concept. We ran our code on a sample of 10,000 ED notes. Results: We estimated that 6% of ED notes contained at least one word/term associated with the Stigma concept and 3% of the notes has presence of EBCL (e.g., naloxone; count=117, buprenorphine; count=83) and offering appropriate follow-up programs such as the methadone program which was present in 106 notes. We acknowledge that this preliminary analysis adopts a standard string-matching approach, but it is sufficient to obtain crude estimates that provide evidence of the presence of stigma in ED notes. We will build more robust methods by employing advanced approaches utilizing deep learning-based NLP to incorporate the context around the target concepts.

Nr:	262
Title:	Video Autofocus Using an Adjustable-Focus Air-Lens and Sharpness Analysis of an Inclined Image Plane
Authors:	Thomas Lynch, Thierry Savin and Paul Meyer
Abstract:	The continuous preservation of focus is a challenge for all forms of imaging, but it is especially so with video microscopy, where the shallow depth of field means imaging is confounded by even the smallest disturbance in the imaging target. Existing active video autofocus solutions are unsuitable in common scenarios, such as when range-finding radiation is reflected by features other than the object of interest, or when the imaging target lacks a steady offset to a discernible surface. Here, we present a device that locates the image plane on a sensor inclined with respect to the image beam. This information informs the movement of an adjustable-focus air-lens, allowing continuous re-focusing of the image to accommodate variation of the object distance. We demonstrate the capability of this device by producing 30 to 60-second in-focus haemoglobin video imaging studies of conjunctival and episcleral microvasculature in the human eye. This is a location in which microcirculations can be reliably sampled and re-visited. Stable, focused videos of the smallest blood vessels in the body offer wide-ranging research and diagnostic opportunities, as these vessels are impacted by many disorders including diabetes, hypertension, large and small vessel vasculitis, haematological disorders, and glaucoma. We also propose that this modular and easy-to-use device is valuable beyond conjunctival imaging, and even beyond video microscopy. This method can be applied to other imaging modalities confounded by axial movement, including still and time-lapse photography, slow motion, and other forms of videography. It is also applicable in other optical imaging systems, such as ophthalmoscopy and telescopy, and to imaging outside the visible spectrum.

Nr:	278
Title:	Ensemble of Self-Supervised Learning Methods for Robust Skin Disease Image Diagnosis Leveraging Unlabeled Data
Authors:	Kosuke Shido
Abstract:	This study investigates advancements in skin disease diagnosis by leveraging self-supervised learning (SSL) techniques applied to the National Skin Disease Dataset of Japan (NSDD). NSDD comprises approximately 400,000 dermatological images covering over 1,500 skin disease types, collected from Japanese patients. Out of these, around 330,000 images remain unlabeled, presenting a unique opportunity to explore SSL methods that utilize large-scale unlabeled datasets for model pre-training. The study specifically employs three cutting-edge SSL methods—SimCLR, BYOL, and DINO—to pre-train models on the NSDD-unlabeled subset, consisting of 100,000 randomly selected images. These methods enable models to learn effective representations without reliance on annotated data, which is often resource-intensive to obtain. Subsequent comparisons against a baseline model pre-trained on ImageNet revealed consistent performance improvements. The SSL-pretrained models demonstrated a 1–2% accuracy gain when fine-tuned and evaluated on the NSDD-labeled subset of 70,196 images. When further tested on external dermatological datasets, including SD-198, Fitzpatrick 17k, Derm7pt, and HAM10000, the SSL models exhibited superior generalization, especially when hybrid pre-training (combining SSL and supervised pre-training) was utilized. Notably, the ensemble approach of diverse SSL models yielded the most robust results, achieving accuracy gains of up to 8.3% on SD-198 and significant improvements across other datasets. This research highlights the potential of SSL in medical image analysis, particularly its capability to leverage ethnically specific data (e.g., Japanese patient images) to enhance model generalizability. The findings also underscore the importance of tailoring pre-training strategies to dataset characteristics, as evidenced by varying performance improvements across different datasets and skin disease categories. Challenges such as biases inherent in the NSDD-unlabeled dataset and the need for task-specific optimization are also discussed, providing a roadmap for future improvements in SSL-based medical imaging frameworks. The study's implications extend beyond dermatology, offering a scalable, open-access methodology for pre-training deep learning models on large, unlabeled medical datasets.

Nr:	284
Title:	Development of a Vessel-on-a-Chip Platform to Study Endothelial Dysfunction
Authors:	Xenia Kraus, David Wörle, Steffen Winkler, Manuel Sirch, Christoph Westerhausen and Janina Bahnemann
Abstract:	Three-dimensional (3D) cell culture has become a popular in vitro method for modelling realistic, in vivo-like systems. Such systems offer significant advantages over traditional two-dimensional (2D) cell culture models. The cultivation of appropriate cell types in a three-dimensional configuration enables the creation of model systems that mimic the structural and functional characteristics of blood vessels. Such systems offer significant potential for investigating pathological alterations in the vasculature, encompassing a range of processes from inflammation to atherosclerosis. This project presents two 3D-printed microfluidic vessel-on-a-chip designs for the co-culturing of endothelial and smooth muscle cells for the purpose of blood vessel reconstruction. The systems comprise a flow chamber, which enables the replication of physiological culture conditions under well-defined shear stress, and a second chamber for static culture. A semi-permeable, track-etched polyethylene terephthalate (PET) membrane is situated between the two chambers, allowing the exchange of substances while maintaining spatial separation. Furthermore, an investigation was conducted to ascertain the suitability of these membranes in terms of their capacity to facilitate cell adhesion and proliferation, as well as their resistance to steam sterilisation. The successful cultivation of human umbilical vein endothelial cells (HUVEC) on the chip for several days under flow demonstrates the potential and potential applications of the 3D-printed system.

Nr:	318
Title:	Leveraging Large Language Models and Embedding-Based Similarities for Patient Journey Analysis: Integrating Visual Analytics with AI-Driven Techniques
Authors:	Rahel Lüthy, Jan Azzati and Dominique Brodbeck
Abstract:	In the practice of medicine, doctors constantly leverage their experience by mentally comparing a patient's journey with those of previously treated patients. By identifying which past histories led to successful outcomes, doctors can learn valuable lessons and adjust the treatment plan of the current patient. However, rather than having to base judgement on few patients, digital medical information allows the inspection of data from thousands of patients at the same time. Achieving this goal requires a robust mechanism to quantify patient similarity, a cornerstone for comparing and learning from vast amounts of patient data. In addition, there is a need for tools that enable exploration of this data, in order to understand the various patient journeys and make sense of the patterns found. Traditionally, machine learning models have been employed to develop algorithms that compute similarity scores, utilizing structured data. The result has then been fed into interactive visual analytics tools, that allow users to obtain an overview, filter the data of interest, and access details in context. There are two problems with this approach. Training machine learning models to compute similarities for sparse heterogeneous data is expensive and difficult. And, visual analytics tools are powerful, but difficult to use for non-experts due to the multitude of functionalities. The advent of large language models (LLMs) has made it possible to address these problems. The concept of LLM embeddings allows for a new approach to comparing patient journeys: patient journeys can be transformed into and treated as textual narratives, which can then be "embedded", into high-dimensional vectors that can easily be evaluated for similarity. The visual analytics tool was augmented with an agent architecture, incorporating an LLM as the reasoning engine. This agent integrates into the tool, knows its capabilities, and the available data sources. It is responsible for: - Interpreting user intent - Translating user intentions into specific queries and tasks - Identifying the most appropriate data sources for these tasks - Executing tasks using the application's tools - Orchestrating user interaction for an efficient analysis process Thus, our tool capitalizes on proven paradigms of classical data analysis applications, optimized to explore embedding-based patient similarities and providing a modern user experience. Users can interact with the data through a combination of traditional "command-based" interfaces and AI-supported "intent-based" experiences. For instance, structured data can be explored in tabular format, while a range of visualizations — such as timelines, histograms, and scatter plots — highlight specific data aspects and leverage similarity scores. AI prompts enhance this experience by enabling users to query the data via a chat window, with the AI responding not only with text but also by orchestrating the user interface, such as highlighting specific patients. This integration of classical data analysis methods and AI-driven techniques offers a powerful and intuitive way to analyze patient journeys, leveraging both data visualization and intelligent assistance to enhance the user experience. Our early experiments were performed on the MIMIC IV dataset, with patient journeys based on little medical context (i.e. sequences of clinics visited by a patient), yet the results produced insights of surprising depth.

Nr:	329
Title:	Evaluating the Potential of Large Language Models for Automating Systematic Reviews: A Study Based on Cochrane Review Data
Authors:	Siun Kim, Yujin Park, Hahyun You and Hyung-Jin Yoon
Abstract:	The rapid accumulation of medical literature has created an urgent need to evaluate the potential of Large Language Models (LLMs) in automating systematic review processes. This study assessed LLM performance in abstract screening using real-world Cochrane review cases. We analyzed 12 Cochrane drug intervention reviews from diverse therapeutic areas (June 2023-2024), processing 1,112 references with standardized DOIs from major medical databases (Ovid MEDLINE, Embase, and CENTRAL). We evaluated performance by comparing LLM screening results against papers included after both abstract screening and full-text review in the original Cochrane reviews, emphasizing recall given the auxiliary nature of automated screening. Among the compared architectures, GPT-4o achieved the best performance with an abstract screening recall of 0.750 (precision: 0.750) and full-text review recall of 0.894 (precision: 0.492), performing marginally better than GPT-4o-mini (abstract screening recall: 0.750, precision: 0.588; full-text review recall: 0.894, precision: 0.386) and substantially outperforming Llama-3.1 8B (abstract screening recall: 0.333, precision: 0.136; full-text review recall: 0.379, precision: 0.085). Using GPT-4o, we reduced 1,112 references to 120 candidates while missing only 7 of the 66 references ultimately included in the original reviews. Furthermore, our qualitative error analysis revealed that these 7 missed cases were primarily those that deviated from explicit free-text guidelines or were included based on arbitrary decisions in the original reviews, suggesting that LLMs can enhance the tractability and reproducibility of the selection process. Future work will focus on: 1) testing newer open-source models such as Phi-4 and Llama-3.3, given the limited performance of Llama-3.1, 2) developing adjustable parameters to control the screening threshold of LLMs, and 3) expanding the evaluation to other systematic review tasks including full-text review, risk of bias assessment, and PICO (Population, Intervention, Comparison, Outcome) element extraction. To our best knowledge, this is the first study to evaluate LLM capabilities using high-quality Cochrane systematic review data, demonstrating promising potential for automating and standardizing the systematic review process while maintaining the rigor of evidence-based medicine practice.

Nr:	334
Title:	Improving In-Context Learning Interpretability with Free-Text Explanations for Complex Tasks
Authors:	Prasan Yapa and Zilu Liang
Abstract:	While large language models (LLMs) have shown impressive performance across various downstream tasks, their ‘black-box’ nature can result in harmful content generation, including hallucinations and potential bias. Drawing inspiration from the in-context learning (ICL) capabilities of LLMs and recent progress in leveraging natural language explanations for improved generalization, we introduce a novel approach that combines free-text explanations E with the in-context examples, also called demonstrations. This method enables LLMs to adapt to unseen tasks, improving the interpretability of demonstrations, while keeping the majority of their parameters frozen. To this end, text-chat conversational (TCC) analysis is used as the pre-training source task, with depression detection (DD) as the previously unseen target task. Our methodology consists of three components to improve the interpretability of LLMs for depression detection. Initially, a demonstration retriever ⅅꭆ identifies the demonstrations from the Reddit Self-reported Depression Diagnosis (RSDD) corpus which are most semantically similar to the input TCC query. These demonstrations D are then utilized to produce synthetic E for the query using LLMs, including Mistral-7B, to improve the interpretability of D for DD. Finally, these E are ranked to determine the most suitable ones through external validation. The validated E are then paired with the TCC source embeddings and their D for downstream DD tasks, utilizing multiple prompt templates and verbalizers within the OpenPrompt framework to address the previously unseen DD. As the key contribution, we propose a method to rank E based on the semantic relevance and diversity of input TCCs and their D. DSM-5 criteria is also incorporated to ensure E are clinically accurate. The proposed methods are evaluated on benchmark datasets constructed from TCC data. The IMHI corpus is used to validate the generated top-E. Mistral-7B, LLaMA-2-7B, and MentaLLaMA-7B were used to evaluate the top-3 generated E and the results are shown in Table 1. For the evaluations of DD shown in Table 2, MentalBERT and MentaLLaMA-7B were used as open-source LLMs to evaluate DD on RSDD corpus. The logic was implemented using Python libraries, including PyTorch and OpenPrompt. KATE was used as ⅅꭆ to obtain top-5 D and three prompt lengths (l) were applied. We will release our code on GitHub for reproducibility.

Nr:	346
Title:	Polarization-Resolved Third Harmonic Generation in Starch
Authors:	Maria Kefalogianni, Sotiris Psilodimitrakopoulos, Leonidas Mouchliadis and Emmanuel Stratakis
Abstract:	Starch serves as the primary energy storage polysaccharide in plants. It is extensively studied due to its complex structure and significant industrial applications in food, materials, and biofuels[1]. It consists primarily of amylose and amylopectin molecules[2]. Based on the amylose composition, starch is categorised to one of the three different crystalline structures. Corn starches contain ~ 20 - 28 % amylose and belong to A-type and orthorhombic symmetry[3,4]. In this study, we present a deeper insight into the polarization dependent third harmonic generation (THG) signals produced by a hydrated corn starch granule. By varying the direction of the excitation linear polarization, we observe two different THG signal modulations, one-peaked and double-peaked which correspond to different regions of the granule, i.e. outer shell, and bulk inside[5]. We employ a theoretical model based on nonlinear optics to describe this response and extract susceptibility ratios and molecular angle orientation. We define the contribution of x orientation compared to y for an orthorhombic starch structure symmetry on the derived THG signal as anisotropy ratio χxxxx/χyyyy. The outer surface of the granule reveals anisotropy ratio approximately 1,2 which means χxxxx is bigger than χyyyy, while in bulk of starch is almost equal to 1. This technique provides a non-invasive way to study and discriminate outer shell from inner regions of granule, with potential applications in agricultural science, food technology, and biopolymer research. [1] E. Agama-Acevedo, et al. Chapter 3 - Cereal Starch Production for Food Applications, Starches for Food Application, 71–102 (Academic Press 2019). [2] D. FRENCH, et al. Starch: Chemistry and Technology, 183–247 (1984). [3] B. Zhang, et al. Food Hydrocoll 31, 68–73 (2013). [4] A. Buléon, et al. Int J Biol Macromol 23, 85–112 (1998). [5] J. Morizet, et al. Optica 6, 385-388 (2019).

Nr:	351
Title:	Corrected Thermal Imaging and VAE Anomaly Detection for non-Contact Health Monitoring of Field Workers
Authors:	Masahito Takano, Kent Nagumo, Lamsal Bikash, Kosuke Oiwa and Akio Nozawa
Abstract:	Effective health monitoring is essential for preventing health-related incidents and accidents, especially in high-risk environments like construction sites. Traditional health assessments, usually conducted by supervisors face several challenges, including false health declarations, unrecognized symptoms, and communication barriers. To overcome these challenges, we have developed a non-contact health monitoring method using facial thermography based on infrared imaging. Facial thermal images reflect skin blood flow patterns regulated by the autonomic nervous system, providing potential as a non-invasive health indicator. This study aims to develop a health condition classification model using thermal images captured in field environments. Over a 10-month period, we conducted a longitudinal investigation involving 279 construction workers, collecting 2,235 facial thermal images along with corresponding health condition data. The analysis revealed significant variations in facial skin temperature caused by ambient temperature and circadian fluctuations. To address these effects, we applied a previously developed thermal image correction method. We then used a Variational Autoencoder (VAE) to model facial thermal patterns under healthy conditions and developed an anomaly detection model to identify signs of unhealthy conditions. Performance evaluation with unseen data yielded an area under the curve (AUC) of 0.66. These findings highlight the potential of facial thermography as a non-invasive tool for health monitoring in occupational settings, contributing to improved worker safety and well-being.

Nr:	353
Title:	Evaluation of Feature Selection for Classification of BOLD Signal and Artifacts in fMRI Data
Authors:	Tereza Švestková, Michal Mikl, Martin Gajdoš and Radek Mareček
Abstract:	Multi-echo fMRI data can provide optimal sensitivity to BOLD signals (blood oxygenation level-dependent) and offer various possibilities to suppress some of artifacts. One approach is based on independent component analysis (ME-ICA), where the use of multiple echo measurements enables the evaluation of metrics capable of distinguishing the presence of BOLD effects in the data. This method is available in the Tedana software. The aim of this study was to evaluate the ability of automated classification of the independent components using machine learning methods, both based on metrics implemented in the Tedana toolbox and on newly proposed metrics. A dataset containing data from 42 healthy subjects performing 3 task types (block, event-related, and resting-state) and two measurement types (TR = 800 ms and TR = 1800 ms), i.e. 6 fMRI runs, was used. Data were processed in MATLAB using the SPM12 toolbox. ME-ICA metrics were implemented according to the Tedana tool, and new metrics based on data quality assessment (tSNR, SNS) were added (total 15 features). Three raters provided expert descriptions of all independent components in the dataset to create ground truth for experimenting with feature selection and evaluating classification accuracy with different classifiers. Initially, metrics with pairwise correlation coefficients above 0.8 were removed to mitigate multicollinearity. The feature selection process was based on the information gain metric, which assesses the importance of individual features in relation to the target variable. Information gain values were first calculated for all available features in the dataset. Then, based on the standard deviation of these values, a threshold was established to classify a feature as important. Features with information gain values exceeding this threshold were labeled as important and included in the final model. The classification models used were Random Forest, Support Vector Machine, and Gradient Boosting. Model performance was evaluated using cross-validation. Final models were assessed using standard metrics, including F1-score, sensitivity, and specificity. Out of the original 14 features (after removing highly correlated features), 10 features were identified using the aforementioned method. Among the most significant features for classification were selected: SNS, kappa, tSNR, signal_noise_t, and cluster ration (CR). The results of classification achieves a high F1-score of 96–98%. The sampling period (TR) did not significantly affect classification results. Differences across individual task types were negligible. Our study shows that automatic classification in the ME-ICA method is sufficiently robust and the metrics based on data quality assessment can be effectively combined with original ME-ICA metrics applicable for preprocessing multi-echo fMRI data. This work was supported by the Czech Science Foundation grant 23-06957S.

Nr:	355
Title:	Towards Metric Learning Applications with Common Spatial Patterns for Improving the Motor Imagery Classification Accuracy
Authors:	Tugçe Balli and Emrullah Fatih Yetkin
Abstract:	BCI illiteracy is a common problem where the users have difficulty effectively controlling a Brain-Computer Interface (BCI) system. Besides, BCI users should be highly motivated during data collection, which is not always possible. Thus, one should consider the accuracy problems arising from the participant-related problems during data collection to build robust BCI applications. Our study introduces a novel pipeline that employed a well-known metric learning approach, Large Margin Nearest Neighbor (LMNN), as a post-pre-processor to enhance the classification accuracy, explicitly targeting the subjects with low accuracy (lower than 45%) in EEG-based MI classification. From the dimensionality reduction perspective, the Common Spatial Patterns (CSP) belongs to the supervised dimensionality reduction family. Consequently, like most comparable approaches, it suffers from a lack of local projections. In the existing literature, there have been a few attempts to improve the efficiency of CSP, especially with multimodal datasets. In this study, instead of changing the mathematical model of the CSP method, we combined LMNN with the CSP algorithm for a better adaptation to local data behaviors. The classification accuracy in BCI studies relies on various factors, such as the selection and extraction of the features, optimization of the classifier parameters, and user proficiency. We used CSP for feature extraction and applied metric learning to enhance the classification accuracy of subjects using the Adaboost algorithm. The proposed approach utilizes the well-known LMNN method to enhance the discriminability of the data before the classification. We used two well-known covariance estimators for CSP: a) classical and b) Ledoit-wolf approach, and compared the results with and without the metric learning approach. Overall results are evaluated regarding test set performances where we kept 70% of each subject's data for training and the remaining 30% for testing. This study used the EEG Motor Movement/Imagery data set recorded using the BCI2000 platform accessible via Physionet. The dataset includes EEG recordings from 109 participants performing various motor and imagery tasks. EEG data were recorded using 64 electrodes placed on the scalp according to the extended international 10-20 system, with a sampling frequency of 160 Hz. Experimental findings suggest that using the LMNN as a post-pre-processor improves the overall accuracy of MI classification. The average testing accuracy for the participants whose classification test accuracy is 45% is increased to 62% after the LMNN application. Moreover, there is no negative impact of LMNN application for highly accurate participants test accuracy values. The results suggest that applying LMNN is preferable, especially for the datasets containing a large number of participants, to improve the classifier accuracy regardless of subject proficiency. For future studies, we plan to extend the experimental studies with different datasets and reveal the connection between the CSP and LMNN (and metric learning in general) for proposing a computationally efficient way of enhancing the performance of modern BCI applications. This work was supported by the Scientific and Technological Research Council of Turkey (TÜBITAK 1001) - Project number: 124E057.

Nr:	356
Title:	In Vitro Validation of Brain Metabolite Detection Using 0.3-Tesla Magnetic Resonance Spectroscopy
Authors:	Ryo Enari, Hiroyuki Ueda and Yosuke Ito
Abstract:	Currently, one in eight people worldwide suffers from mental disorders, highlighting the severity of their social impact and the importance of early diagnosis and treatment. However, the underlying causes of many mental disorders remain unclear, leading to issues such as misdiagnosis during medical interviews, inappropriate pharmacological treatments, and difficulties in monitoring therapeutic progress. Efforts toward quantitative evaluation employ both invasive and non-invasive methods. Focusing on the latter for their safety, techniques such as functional magnetic resonance imaging (fMRI) and magnetoencephalography (MEG), which primarily target neural activity, have become mainstream. In contrast, magnetic resonance spectroscopy (MRS) has gained attention for its potential to directly quantify metabolites in the brain. However, its clinical application is limited due to the high costs associated with the superconducting magnets and the insufficient amount of available data. This study aims to explore the feasibility of low-field MRS using low-cost MR systems equipped with permanent magnets that do not require a superconducting state. Such systems have the potential to improve access to diagnostic tools, significantly increase data availability, and contribute to quantitative evaluations. However, low-field MRS faces challenges such as low resolution and reduced SNR, necessitating improved signal acquisition methods. In this research, in vitro measurements of metabolites such as N-acetylaspartate (NAA), glutamine (Gln), and glutamate (Glu) were conducted using a 0.3-T MR system. These metabolites were selected based on their clinical relevance and concentration levels in the occipital lobe. To suppress water signals, which are approximately 3000 times stronger than brain metabolite signals, the Inversion Recovery (IR) sequence was employed for signal acquisition. To address the challenges of reduced signal strength in low-field environments, simulations were conducted using optically pumped magnetometers (OPMs), high-sensitivity magnetic sensors known for their exceptional performance in low magnetic fields. A flux-transformer with an RLC circuit was implemented to adapt magnetic field signals for OPM measurement, effectively reducing thermal noise. These simulations provided insights into the magnetic field ranges where OPMs demonstrate superior sensitivity, complementing the findings from the experiment. The experimental results demonstrated quantified noise levels for MRS at 0.3-T, providing a foundation for practical low-field MRS applications. Simulations further revealed the potential of OPMs as high-sensitivity detectors in low-field environments, suggesting that advancements in noise reduction and signal processing could pave the way for 1H-MRS in low-field systems.

Nr:	360
Title:	Causal Machine Learning Approach for Quantifying Heterogeneous Effects of Maternal Renal Disorders on Adverse Pregnancy Outcomes
Authors:	Zio Kim and Hyung-Jin Yoon
Abstract:	Background: Maternal renal disorders pose significant risks to maternal and neonatal health, yet the heterogeneity of these effects remains poorly understood. This study employs a causal machine learning technique to assess the varied effects of maternal renal disorders on adverse pregnancy outcomes (APO). By utilizing both traditional statistical method and advanced causal machine learning model, this study aimed to discover patterns in the relationship between maternal renal health and APO, and aim to inform more personalized caring strategies. Methods: This retrospective study was conducted using data from the National Health Insurance Service (NHIS) Database of South Korea from January 1st, 2008, through December 31st, 2017. The study population included women who have given birth within the study period with and without renal disorders, and their offspring. To assess effect heterogeneity, Causal Forest (CF) model was implemented, which allows for the estimation of individual treatment effects (ITE). The CF model achieves this by combining machine learning techniques with causal inference principles, using a modified random forest algorithm to predict treatment assignments. Multivariable logistic regression was also employed to estimate average treatment effects. Subgroup analyses were conducted based on the heterogeneity identified by the CF model, and Cohen’s D metric was utilized to quantify the size of differences between subgroups. The results of traditional method were compared with those of the CF approach. Findings: A total of 164,647 mothers were enrolled, of whom 130,481 (79.2%) were classified as healthy controls and 34,166 (20.8%) were identified as having renal disorders. Traditional statistical method revealed significant associations between maternal renal disorders and APO (OR: 1.15, 95% CI: 1.11–1.19). The CF models discovered substantial heterogeneity in this effect. The treatment effect was varied based on maternal BMI, delivery type, and pre-existing comorbidities, significantly modulating the effect of renal disorders (Estimated ITE difference between underweight group and overweight group: ΔITE = -0.0313, Cohen’s D: -0.77 (medium); Normal BMI group and overweight group: ΔITE = -0.0311, Cohen’s D: -0.83 (large); Dyslipidemia group and Non– Dyslipidemia group: ΔITE = -0.0109, Cohen’s D: -0.28 (small); Caesarean section group and Natural birth group: ΔITE = 0.0112, Cohen’s D: 0.29 (small)). The t-test results for all subgroups were statistically significant under 0.05 level. Cohen’s D metric and t-test results both indicated significant difference between covariate subgroups and inferred to underlying heterogeneity in treatment effect of maternal renal disorder. The CF model outperformed traditional method in predicting individual-level treatment effects and identifying high-risk subgroups. Interpretation: Findings discovered from this study highlight the complex and varied relationship between maternal renal disorders and APO. The heterogeneity identified by CF model suggests that a one-size-fits-all approach to managing pregnancies complicated by renal disorders may be inaccurate. Instead, results support the need for more personalized risk assessment and tailored interventions based on individual maternal characteristics. This study demonstrates the potential of causal machine learning approaches in advancing our understanding of complex health relationships and informing precision medicine in prenatal care.

Nr:	362
Title:	UV Light-Triggered Nanoparticles for Drug Delivery Systems in Skin Environments
Authors:	Maria Marin, Javier Sendros, Mari Carmen Gonzalez, Lola Mulero, Jose Antonio Llamas, Saioa Medizuri, Jaume Boix and Mireia Torres
Abstract:	Introduction: Drug delivery systems have been evolving in the last decades due to improvements in materials, allowing to control the release under controlled conditions. The following work shows the development of a release system controlled by ultraviolet light. A nanoparticle system has been developed using solvent evaporation methodology. These nanoparticles are made of an amphiphilic material, which in the apolar tail has a chemical group that reacts upon interaction with ultraviolet light. To study the release of these active ingredients, 3D skins models have been used to show both the permeation of these nanoparticles and the release of an active ingredient marked with an oil red. The study was carried out by tracking the nanoparticles, both irradiated with ultraviolet light and non-irradiated, with an LSM980-AiryScan2 confocal microscope. Methodology: The skin model was surface treated with an aqueous particle dispersion. The non-active particles fluoresce green due to the chemical group in the apolar tail. In contrast, when the skin is treated with particles loaded with oil red, there is a superposition of fluorescent signals, red + green being yellow/orange. In addition, when the skin treated with oil red nanoparticles is irradiated with ultraviolet light, oil red is released. These results could be shown by confocal microscopy. Results & Conclusions: High-resolution confocal microscopy has been effective in characterizing the UV-sensitive nanoparticles. It has been possible to show visually how the release system is sensitive to this irradiation. The physico-chemical properties of the particles allow their permeation through the different layers of the skin, maintaining their release integrity

Nr:	365
Title:	Novel Neuron-Targeted Dendriplexes for siRNA Delivery: Insights from a PNS-CNS-on-Chip Model
Authors:	Ana Paula Pêgo, Ana P. Spencer, Miguel Xavier, Sofia C. Guimarães, Adriana Vilaça, Rafael Santos, María Lázaro, Eran Perlson, Victoria Leiro and Ben Maoz
Abstract:	Neurological disorders, a leading global cause of death and disabilities, encompass conditions affecting both the peripheral and central nervous systems (PNS and CNS, respectively). Limited axon regeneration is a significant challenge in these disorders, and it has been linked to proteins like PTEN. RNA-based therapeutics, particularly siRNAs, hold the potential for silencing these inhibitory pathways, but their clinical application is hindered by poor stability and cellular uptake. Our study addressed this challenge with the development of novel, fully biodegradable nanoparticles (NPs) designed specifically for neuron targeting upon a peripheral and minimally invasive intramuscular administration. Both, a novel proprietary family of fully biodegradable dendrimers and thiolated trimethyl chitosan (TMC-SH) have been explored as carriers of siRNA targeting PTEN. To attain neurotropism, our nanoparticles were functionalized with the non-toxic C-terminal fragment of the tetanus neurotoxin (TeNT) heavy chain (HC), enhancing selective neuronal targeting and cellular internalization. In prior studies, we explored trimethyl chitosan (TMC) NPs that were functionalized with the HC fragment to specifically target PNS neurons and undergo retrograde transportation to the cell body in vivo. However, HC’s full potential in CNS-targeting and transcytosis remained unexplored. To explore this, three advanced microfluidic platforms were developed to explore the features of our fully biodegradable and targeted NPs, focusing specifically on intra-neuronal migration, biological effects, and transcytosis between neurons. For this, it was crucial to create a novel microfluidic-based smart model, the PNS-CNS-on-Chip, integrating microelectrode arrays (MEAs), and engineered to simulate various aspects of the complex PNS-CNS interface, allowing the validation of the targeted delivery capabilities of our NPs. The retrograde transport of our nanoparticles along the axon to the cell body of neurons was characterized through STED super resolution microscopy and spinning disk confocal microscopy. We demonstrated that both nanosystems efficiently mediate siRNA delivery in neuronal cultures without causing any cytotoxic effects. Axonal growth promoted by the delivery of siRNA anti-PETN was further confirmed in microfluidic models. In a groundbreaking PNS-CNS-on-Chip we were able to mimic the bio-interaction between the dendriplexes and PNS neurons and monitor the intracellular trafficking of these nanosystems. Dendriplexes exhibited effective migration from PNS to CNS neurons, highlighting their potential for targeted therapeutic delivery via a minimally invasive administration. This study pioneers the application of microfluidics to demonstrate the CNS targeting of nanosystems, paving the way for innovative treatments in the field of nanomedicine. Acknowledgments: This work was supported by Portuguese funds through Fundação para a Ciência e a Tecnologia, I. P. (FCT) in the framework of the project PTDC/BTM-MAT/4156/2021.

Nr:	366
Title:	Machine Learning Based Analysis for Radiomics Features Robustness in Real-World Scenarios
Authors:	Sarmad Ahmad Khan, Zahra Moslehi and Florian Buettner
Abstract:	Radiomics enables quantitative analysis of radiological images, allowing for comprehensive analysis within a region of interest (ROI). However, variability in imaging techniques, image acquisition and segmentation can affect model robustness. To study the effect of these variations on radiomics features selection, this study investigates the robustness of different radiomics feature sets under different conditions. Building on a previous study that identified robust features across multiple MRI sequences out of all the radiomics features extracted using Pyradiomics software, this research quantifies the robustness of machine learning models building on such feature sets. We used a gradient boosting based machine learning model used for the features-based training and prediction. To control the experimental setup, four fruits (lime, kiwi, apple, onion) were imaged using five MRI sequences. T2 map, T2 HASTE, T2 FLAIR, T2 TSE, and T1 TSE. The MRI images were imported into 3D Slicer, where a VOI was defined using the "Paint" tool. The segmentation was refined with the "Grow from Seeds" algorithm and manually corrected. We then set up a machine learning task of classifying fruit types based on radiomics features derived from the segmented objects. More specifically, we employed two distinct test sets—one with rotated and another with partially segmented features—alongside training and validation sets from normally segmented phantoms. In addition, we integrated different MRI sequences, which have different data distribution from each other. This approach allows us to evaluate model performance in out-of-domain (using data that differ significantly from the training set) scenario, simulating real-world domain drift. Our methodology compares model robustness when using (1) consistent features across sequences, (2) sequence-specific robust features, and (3) the full set of features extracted from Pyradiomics. The results show that consistent features improve model performance across different scenarios, followed by sequence-specific robust features, with the use of all features yielding the lowest performance. Model accuracy decreased for datasets with limited similarity to the training set, with F1 scores of ~55% for partial segmentation datasets compared to >80% for rotated phantom datasets. These results highlight the value of feature robustness in optimising machine learning models in radiomics.

Nr:	367
Title:	Efficient Automatic IMT Region Extraction Using MobileNetV3 and SegNet for Mobile Carotid Ultrasound Imaging
Authors:	Hiroyasu Usami, Natsuki Nakayama and Naoko Arakawa
Abstract:	[Background] Carotid Intima-Media Thickness (IMT) measurement in ultrasound imaging is essential for cardiovascular risk assessment, offering insights into vascular health. However, manual methods are time-consuming and prone to variability, posing challenges for reliability and standardization. Portable ultrasound devices offer opportunities for automated IMT measurement, particularly in resource-limited settings. However, deploying existing methods on mobile platforms faces challenges due to computational power, memory, and energy constraints. Lightweight, efficient deep-learning architectures are needed to achieve high accuracy within these constraints. [Methods] A two-stage deep learning architecture optimized for mobile ultrasound devices was developed. The first stage employs MobileNetV3, a lightweight convolutional neural network, to classify ultrasound images by relevance for IMT measurement, reducing computational load by filtering irrelevant images. The second stage uses SegNet, a memory-efficient encoder-decoder model, to segment IMT regions accurately. SegNet leverages pooling indices during decoding to minimize memory usage while maintaining precision, making it suitable for mobile platforms. The system was evaluated on 190 annotated carotid ultrasound images. Three optimization techniques enhanced mobile deployment: 1. Model Quantization: Reduced weights and activations to 8-bit precision, decreasing memory use and inference time. 2. Pruning: Removed less significant parameters, reducing model size while retaining accuracy. 3. Knowledge Distillation: Transferred knowledge from a larger model to the lightweight architecture for competitive performance. [Results] The architecture achieved a 43.9% Intersection over Union (IoU) for IMT segmentation and 91.3% classification accuracy, outperforming conventional single-stage methods. The pre-segmentation of blood vessel regions in the two-stage design minimized false positives. Optimization techniques reduced memory usage by 45%, enabling real-time processing with an average inference time under 120 milliseconds per image on standard mobile GPUs. [Conclusions] The automated pipeline for IMT measurement addresses the need for scalable cardiovascular diagnostics on mobile devices. By integrating MobileNetV3 and SegNet, the system achieves high accuracy while adhering to mobile platform constraints. Optimization techniques ensure a lightweight, robust architecture for real-time use in resource-limited environments. This system can democratize cardiovascular diagnostics by providing an affordable, portable solution for underserved regions. Applications extend to non-clinical settings such as community health programs, enabling early detection of atherosclerosis and improving patient outcomes. Future work includes validation on larger datasets, development of user-friendly applications, enhancement of accuracy through advanced techniques, and compliance with regulatory standards for clinical adoption. This system marks a significant step forward in mobile health technologies, paving the way for improved cardiovascular diagnostics and disease management.

Nr:	370
Title:	Digital Twin Based Personal Health Coaching Framework Using on-Device AI
Authors:	Yeong-Tae Song
Abstract:	When a person comes down with a disease, it is important to get proper treatment at right time to get it cured but due to various reasons, it may not be always possible. What is more important is to try not to be in such situation by monitoring their own health and take necessary actions when needed. Clinical institutions such as hospitals are perfect place to get treatments but when it comes to maintaining personal health, they may not be the right place to be as health monitoring requires constant assessment and guidelines that may differ by person. With the advancement of health information technology, such personalized and constant guidance that can ensure health information privacy, security, and sharing may be feasible. In this paper, we propose a novel approach in personal health monitoring framework using digital twin and on-device AI. Digital twin is used to maintain user’s most up-to-date health condition through clinical sensors and observed symptoms. On-device AI is used to provide secure personalized health coaching capabilities as it does not consult external AI such as ChatGPT.

Nr:	371
Title:	Game On for Better Sleep: Designing Serious Games to Promote Sleep Hygiene in University Students
Authors:	Zilu Liang, Edward Melcer, Daeun Hwang, Nhung Huyen Hoang and Kingkarn Khotchasing
Abstract:	University students often face sleep disturbances, negatively impacting their academic performance, mental health, and well-being. Existing sleep education programs often fail to engage this population effectively. This study addresses this gap by developing engaging sleep hygiene promotion systems using serious games and smartwatches. The study progressed in two phases. Phase one involved three empathy workshops with 51 university students aged 18 to 35. These workshops identified key factors contributing to sleep disturbances, including excessive screen time, stress, anxiety, irregular sleep schedules, poor eating habits, and uncomfortable sleep environments. These findings align with previous research. However, an interesting insight was that participants did not always view late-night disruptions as negative. For some, staying up late was a choice for creativity or productivity, not just procrastination. This challenges the conventional view that sleep disruptions should always be avoided, emphasizing the value of the co-design approach. The workshops generated design ideas to address these sleep disturbances and improve existing technologies. Proposed solutions included automated sleep environment optimization, screen time regulation, relaxation techniques, AI-driven sleep induction, and smoother transitions from sleep to wakefulness. These ideas emphasized actionable interventions targeting behavioral and psychological changes. Participants also suggested integrating IoT, VR, and AI (e.g., ChatGPT) into sleep apps. In addition to improving sleep, students desired tools that supported academic and personal goals, such as learning retention, skill development, and managing personal projects. Gamification and creative features were also proposed to enhance engagement and provide long-term benefits. Based on these insights, the research team developed three sleep game prototypes. In phase two, a design workshop collected feedback on the three prototypes: Hero’s Sleep Journey, Sleep Tamagotchi, and Sleepland. The workshop began with an introduction to the study goals, serious games, wearable sleep-tracking technologies, and sleep hygiene. After an icebreaker activity, participants evaluated the game prototypes using storyboard cards, which included questions such as “How useful is this feature for improving sleep?” and “Would you use this feature?” Participants rated features on a Likert scale and provided open-ended feedback. They also expressed preferences for the most appealing and effective games. Hero’s Sleep Journey was the most popular game, but Sleep Tamagotchi was considered the most effective for improving sleep health. Qualitative analysis revealed three dimensions of relevance: psychological (personal connection), logical (alignment with sleep health goals), and situational (contextual fit). These dimensions align with the autonomy and relatedness constructs in self-determination theory, leading to three design recommendations. The findings suggest that serious sleep games can engage university students in promoting healthy sleep hygiene. Future designs should create games that resonate with users’ personal lives, health goals, and situational contexts. A variety of game genres and features should be offered to cater to diverse needs. Moreover, allowing users to customize their experience and set technology boundaries will help foster a sense of control and autonomy in improving sleep hygiene.

Nr:	373
Title:	Evaluating the Reliability of Multi-Echo T2*-Weighted fMRI Across Sessions and Sites: A Multi-Center Study
Authors:	Anezka Kovarova, Michal Mikl and Petr Hlustik
Abstract:	In fMRI research, ensuring the reliability of data across sessions and sites is one of the crucial aspects for advancing robust and reproducible findings, particularly in multi-center studies. Multi-echo (ME) fMRI, which acquires data at multiple echo times, has shown promise in enhancing data quality by improving the signal-to-noise ratio and separating neural signals from physiological and scanner noise. Despite these advantages, the reliability of ME fMRI—and even ME T2-weighted signal components—remains not thoroughly explored in multi-center settings. This study is a pilot insight by evaluating the session-to-session and site-to-site reliability of ME T2-weighted fMRI data relative to conventional ME data. Human participants were scanned at two separate sites, each undergoing one session per site, to provide comprehension into the robustness of ME T2-weighted fMRI. Our findings aim to give a bit of insight into strategies for harmonizing data in multi-center neuroimaging research, thereby enhancing the reliability of pooled datasets and the generalizability of the results. Data was collected from 24 healthy volunteers (8 women, 16 men, aged 21-58) with ethics approval. Scans were conducted at Brno (Siemens Prisma 3T) and Olomouc (Siemens Vida 3T) using high-resolution anatomical imaging and three multi-echo runs: Visual Oddball (VOB): Reaction to the letter "X"; Motor Task (SFO): Finger tapping in response to "START"/"STOP"; Verbal Fluency Task (VFT): Inventing words starting with a given letter. Imaging parameters included echo times of 14, 34, and 54 ms, voxel size 3x3x3 mm, repetition time 700 ms, and multi-band factor 6. Data were processed using two approaches: a composite ME model optimized for contrast-to-noise and a T2-weighted estimation. Reliability metrics included cosine similarity, Dice similarity, and Pearson similarity, analyzed globally and within task-specific regions of interest (ROIs) using the AAL atlas. The comparison of the pair ME acquisitions and the pair ME T2* acquisitions data was done using several metrics. When we compared contrast maps with cosine and Pearson similarity, we found out that the ME model performed better in reliability results in both metrics. However, in cosine similarity, Pearson similarity and dice similarity computed from t-statistics maps, the T2* model yielded as high reliability values as ME model in all three metrics. In Pearson similarity, we can even say that both models perfomed equally good. According to the task activation, ROIs across the brain were chosen from the AAL atlas. We checked the reliability in areas associated with voluntary motor movement (precentral gyrus), visual function (middle occipital gyrus), also supplementary motor area and other task-related regions. The results for cosine similarity were equally strong for ME and ME T2* models for both motor and visual areas, reaching over 93% for SFO and VFT tasks. This study demonstrates that ME fMRI data exhibit robust reliability across sessions and sites, with the T2-weighted model performing comparably to the conventional ME model in key metrics such as Pearson and cosine similarity. These findings confirm the potential of ME T2-weighted acquisitions for keeping high level of reliability in multi-center data and maintaining task-relevant signal integrity.

Nr:	19
Title:	A Real-Time Approach to the Detection and Phenotyping of Cell Vibrations
Authors:	Ali Al-khaz'Aly, Salim Ghandourah, Jared James Topham, Nasir Mohamed Osman, Taye Louie, Farshad Farshadfar and Matthias Walter Amrein
Abstract:	Background: All living cells produce unique vibrations due to internal metabolic processes which have previously been studied using Atomic Force Microscopy (AFM). These studies have established proof of concept for the potential of cell vibrational profiling in biomedicine. However, few, if any, cell-specific frequencies were witnessed due to the lack of sensitivity of AFM for this application, precluding detailed phenotyping. In this proposal, we use the Optical Tweezers (OT), a non-invasive tool, to provide improved resolution of vibratory signals in a novel approach to cell vibrational profiling. We hypothesize that due to metabolic processes, every cell releases a unique vibrational frequency pattern (VFP) which contains sufficient information to phenotype the cell using OT. Rationale: The OT can detect piconewton-scale forces acting on molecules or cells, with megahertz temporal resolution. To test the utility of OT for VFP detection, we differentiated VFPs of glioblastoma cells (i.e., BT48, BT53, and U251), and analyzed the effects of different conditions on VFPs over time. We utilized multivariate statistical analysis (MVSA) techniques to statistically establish differentiation. The adopted MVSA techniques of Principal Component Analysis (PCA) and Orthogonal Partial Least Squares Discriminant Analysis (OPLS-DA) group or separate data samples on the basis of variance inherent in the data, then train supervised learning models on the dataset and corresponding cell types. This ongoing work, while already demonstrating that closely related and unrelated cells may be successfully differentiated at the single-cell level, exposed problems that we hereby seek to address. Cell cycle, media viscosity, external noise or cell health, among other factors, have been found to affect results but are only discovered after batch processing of large data sets days or weeks after the experiment. The lack of an intuitive real-time assessment of an experiment result in an insufficient dataset for MVSA models to differentiate and profile cells. Objectives: First, a real-time analysis system will be developed. We propose to progress the methodology by incorporating real time analysis of signals using a spectrum analyzer, running Fast Fourier Transform (FFT) and Autocorrelation techniques for real-time data evaluation. This provides a real time assessment of cells as they vibrate, granting an immediate diagnosis of cell condition and type. Second, we employ quantitative data processing in a thorough offline analysis using MVSA. This offline analysis of high-pass filtering, FFT, and peak tabulation will establish the statistical procedures that allow for differentiating phenotypes. This includes training OPLS-DA models to predict cell phenotypes based on the vibrational spectrum. Significance: Once complete, VFP detection has application in metabolic disorders and cancer treatment, benefitting intraoperative diagnostics. In a use scenario for VFP detection, it would be performed alongside surgery for cancer excision. Cells would be obtained directly from surgical suction lines, injected into OT and analyzed for type and stage. Reaching the surgical resection margin would become clear with the appearance of VFPs of healthy cells in the sample. This proposal strides towards making these applications a reality. This proof-of-concept project is significant, building upon previous work, and brings the OT and VFP detection closer to medical application.

Nr:	31
Title:	Architectural Design for a Web3-Based Healthcare System Associated with the MyHealthWay Platform in the Republic of Korea
Authors:	Sungjae Jung, Dongjae Shin and Hyung-Jin Yoon
Abstract:	Introduction: Current Internet services are structured in a centralized manner, where service providers own customer data, making it difficult for users to manage or exercise rights over their data. However, with the recent advancements in blockchain technology and its applications, the practical implementation of Web3, where individuals own and control their data, has become feasible. In the healthcare sector, blockchain technology can be utilized to share data securely through smart contracts. Additionally, security and control over data can be enhanced by leveraging decentralized data storage solutions, such as DID-based identity verification and IPFS. In 2023, the government of the Republic of Korea launched the MyHealthWay platform, enabling individuals to access and utilize their health information. In this paper, we propose a structural design for a Web3-based healthcare system that allows individuals to directly own, manage, and grant third-party access to their data when necessary. System Design: Our system replaces traditional centralized verification methods with blockchain-based IDs by managing personal digital identities through DIDs. Data access control is handled using DIDs and smart contracts. Data is stored on blockchain and distributed storage systems. It is encrypted, hashed, and linked to an individual’s DID to ensure controlled access. Healthcare data exchange is compatible with third-party services by adhering to the HL7 FHIR standard, which is also followed by the MyHealthWay platform to ensure seamless interoperability. Implementation: To verify the effectiveness of our system design, we developed the HealthCube platform in collaboration with PARAMETA Corp., a Web3 specialist company, and CRYPTOLAB Inc., a specialist in homomorphic encryption. Individual users interact with the system through the following steps: (1) User authentication and health record collection: The user authenticates their identity and retrieves their health records, such as medical treatment history, prescription records, and health checkup results, from the MyHealthWay platform. (2) Secure storage of health records: The user’s health records are securely stored in a distributed data storage system through the HealthCube platform, with their data linked to their DID, ensuring full control over access permissions. (3) Granting access to healthcare providers: When the user uses healthcare services, they grant the provider’s DID access to their health data stored in the decentralized storage. This process ensures that the user controls which third parties can access their data. (4) Data privacy and protection: Even when the user shares access to their data, the original data remains protected by homomorphic encryption. This ensures that the healthcare provider can process the necessary information without compromising the privacy or security of the user’s original data. Conclusion: Web3-based healthcare systems have the potential to revolutionize personal data management and healthcare services. However, challenges such as regulatory compliance and technical complexities must be addressed before these systems can be implemented as operational services. Acknowledgement: This research was supported by a grant of the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health & Welfare, Republic of Korea (grant number: RS-2023-KH136520).

Nr:	180
Title:	Combination of pMUT and Microelectrodes for the Development of Neural Stimulators
Authors:	Sahil Sharma, Bruno Fain and Clement Hebert
Abstract:	Piezoelectric micromachined ultrasonic transducers (pMUTs) have drawn of attention in recent years due to the potential for their application in biomedical applications, particularly neural stimulation, and bioelectronics. The beneficial characteristics of pMUTs comprise miniaturization, low power consumption, and wireless capabilities. Recent work highlights their usage in a variety of applications, including drug delivery systems to neural stimulation, where precise control of electrical signals is crucial for effective neural therapy. Wireless powering via ultrasound has emerged as a transformative solution compared to traditional batteries and wire-based microelectrodes in implantable medical devices. Wired systems not only restrict patient mobility but also pose risks of infection and require complex surgeries for placement and maintenance. The present study intended to evaluate the capacity of a single bimorph pMUT resonating at 140kHz to be used in a wireless stimulator device by measuring the electrical current that flows between microelectrodes connected at its terminals in phosphate-buffered saline (PBS) solution. Preliminary results showed that the pMUT coupled to the microelectrode is able to generate an electrical sinusoidal current with an amplitude of up to 10µA in PBS which is the first step for the development of a neural stimulator. Further optimization of the pMUTs with integrated micro-electrode array architectures is expected to significantly enhance their performance. The results emphasize the future potential of pMUT-based wireless neural electrical stimulation, which could be applied to neural interfaces for minimally invasive therapies.

Nr:	194
Title:	A Novel Robotic Driving Simulator for Post-Stroke Bilateral Rehabilitation of Upper Limbs
Authors:	Rocco Adduci, Francesca Alvaro, Michele Perrelli, Francesco Cosco, Nicola Marotta, Gionata Fragomeni, Antonio Ammendolia and Domenico Mundo
Abstract:	Stroke is one of the main causes of disability worldwide with a variety of sensory, motor, cognitive, and psychological symptoms that affect the ability to perform basic activities of daily living. The individual rehabilitation project is drawn up in order to allow the patient a personalized therapeutic path aimed at avoiding the loss of residual abilities and obtaining a progressive recovery of lost abilities. This therapeutic path requires specific skills and advanced technological tools, which can support therapists in implementing an efficient and effective recovery strategy where patient motivation plays a crucial role. In recent times, the use of robots is generating significant improvements in common rehabilitation practices thanks to their versatility in being used over different phases of the therapy. Typical robot-based rehabilitation modes can be summarized as: (i) passive, when the robot performs the movement entirely and the patient exerts no effort; (ii) assisted, in which the device detects the patient’s intention and provides assistance in executing the task; (iii) active-assisted, in which the robot completes the execution of the movement when the limb does not exert enough force; (iv) active, in which the robot is used as measurement device and exerts no force; and (v) resistive, where the robot exerts a resisting force that opposes the movement imparted by the patient. A variety of robots are commercially available for upper limb rehabilitation, ranging from simple manipulators, with one or two degrees of freedom, to industrial manipulators with five and more degrees of freedom, to wearable exoskeletons. For patients with peripheral nerve injury, driving simulators are largely used for the rehabilitation of the upper limbs and/or to assess the ability of individuals with neurological and/or motor disorders to resume driving activities. However, driving simulators commonly used today suffer from inherent limitations, e.g., they present lack of differentiated virtual scenarios, tasks and rehabilitation modes. In present work, concept and working principle of a novel driving simulator for motorial and neurological rehabilitation is presented to overcome the above-mentioned limitations. It consists of: two kinematically independent steerable handles configured to be gripped by a left hand and a right hand, respectively, of a patient with motor deficits in at least one upper limb and/or cognitive impairment; a pair of servomotors respectively connected to the handles and configured to control their independent angular positions; an electronic control unit connected to the servomotors including an acquisition and control module configured to handle a plurality of operating modes, to assess rehabilitation progresses and the patient's cognitive status. The proposed device combines the advantages of standard rehabilitation robots but differs in its ability to control and impart tasks to the patient's upper limbs independently for bilateral rehabilitation. The device allows simultaneous administration of rehabilitative and visuo-cognitive tasks that can emulate realistic driving conditions. The nature of the tasks is highly motivating in view of the patient's future return to driving. The motivational aspect suggests that rehabilitation therapy is more effective than other devices with similar purposes. It is expected that an affordable product can be developed and therefore used for home rehabilitation.

Nr:	288
Title:	Low-Cost All-Polymeric Microfluidic Sensor for Detection of Streptococcus pyogenes and Lactate in Saliva: Steps Towards Space Application
Authors:	Ali Doostmohammadi and Pouya Rezai
Abstract:	Long-duration space missions present unique health challenges, requiring innovative diagnostic solutions that can perform reliably under extreme environmental conditions. This study introduces an imprinted polymer (IP)-based microfluidic sensor, designed in a pacifier-style format, for the non-invasive, multiplex detection of salivary biomarkers critical to astronaut health. By integrating microfluidic and molecular/cell imprinting technologies, the all-polymeric sensor offers a compact, low-cost, and robust diagnostic platform for the demands of space exploration and astronaut health monitoring. The device incorporates four parallel microchannels, two for Streptococcus (S.) pyogenes bacteria and two for lactate detection, each pair containing an IP sensing and a non-imprinted polymer (NIP) control test area. The targeted analytes are key biomarkers for assessing immune function and metabolic activity under space conditions. The four channels use capillary-driven flow, enabling efficient saliva collection and transport, and fluorescence intensity attenuations before and after capturing target analytes (S. pyogenes or lactate) on IPs as the sensing mechanism. NIP fluorescent changes showed nonspecific binding while IP fluorescent changes demonstrated specific analyte capturing and detection. To optimize the sensor’s performance, sample injection flow rates ranging from 0.01 to 0.5 mL/min were tested. The results indicated a flow rate of 0.01 mL/min as the most effective condition for analyte-IP interaction and detection sensitivity. This target flow rate was achieved under capillary-driven conditions by designing microchannels with specific dimensions: a width of 500 µm, a height of 126 µm, and a length of 20 mm. These dimensions ensured efficient sample transport and maximized interaction time between the target analyte and the IP-coated surfaces, significantly improving detection efficiency. The sensor's performance was validated through dose-response experiments. For bacterial detection, the device demonstrated a dose-dependent response to S. pyogenes at concentrations up to 10^8 CFU/mL on bacteria-IP sensing areas compared to NIP areas. Similarly, lactate detection within the physiologically relevant range (0–2 mM) showed high sensitivity, with the lactate-IP channels displaying significantly higher responses compared to NIP control. Selectivity tests confirmed the sensor’s accuracy, showing very little cross-reactivity with non-target species like Salmonella and E. coli in the bacterial channels. Likewise, the lactate channels showed almost no interference from ascorbic acid or uric acid, demonstrating the specificity of the IP coatings. The suitability for space applications was tested under simulated space conditions, including thermal fluctuations (4–35°C) and mechanical vibrations. Both the bacteria and lactate imprinted channels maintained consistent fluorescence intensity responses across all tested conditions. This microfluidic sensor represents a significant step forward in diagnostic technologies for space exploration. Its compact and lightweight design, coupled with its ability to provide real-time, non-invasive diagnostics, has the potential to address the gaps in astronaut healthcare monitoring. Beyond its application in space, the sensor holds potential for use in resource-limited and remote environments, offering a portable, rapid, and sensitive tool for health monitoring.

Nr:	289
Title:	Wearable Single Channel EEG Using Qvar and Artificial Intelligence
Authors:	Michele Antonio Gazzanti Pugliese di Cotrone, Marco Balsi, Nicola Picozzi, Alessandro Zampogna, Soufyane Bouchelaghem, Antonio Suppa, Leonardo Davì, Denise Fabeni, Alessandro Gumiero, Ludovica Ferri, Luigi Della Torre and Fernanda Irrera
Abstract:	Wearable technologies enable non-invasive health monitoring in domestic and outdoor contexts. This paper explores the use of Qvar, a front-end for signal conditioning and A/D conversion, in a wearable single-channel device for the long-term, real-time acquisition of ElectroEncephaloGram (EEG). The Qvar, originally developed for automotive applications, stands out for its low power consumption, small size and ease of integration. We integrated the Qvar into a compact platform suitable for easy and friendly placement, as, for example, in the arm of smart glasses. The study validates the effectiveness of the Qvar for EEG acquisition by comparison with a hospital gold standard (Micromed Brain Quick), demonstrating encouraging results in terms of signal accuracy and reliability. Validation was performed acquiring the EEG signal from a single channel in F7, with wet electrodes of the two devices positioned on the temple very close to each other and referenced to the mastoid. It included both time analyses and the Power Spectral Density (PSD), as well as a comparison of key EEG features, showcasing robustness and sensitivity of Qvar. When supported by proper signal processing the application of Qvar in a comfortable, non-invasive wearable system for EEG recording may facilitate monitoring of neurological conditions in non-clinical environments. So, embedding Machine Learning (ML) in the Qvar platform can realize an accessible, cost-effective solution for applications where ubiquity, real-time, long-term EEG recording are key requirements. As a first application, we focused on the detection of drowsiness condition in healthy subjects. To the aim, we used an online available dataset containing EEG traces from F7 in many subjects performing all the same task (drive simulation with sudden trajectory deviations) in sleep deprivation. We considered Reaction Times (RT) in response to the visual stimulus. RTs were labelled either drowsy or alert states based on statistical analysis (alert: RT<5th percentile x 1.5; drowsy: RT>5th percentile x 2.5). Two ML algorithms, Linear Discriminant Analysis (LDA) and Support Vector Machine (SVM), were adopted for a preliminary study. The labelled data were split into 70% for training and 30% for testing algorithms. Twenty features were extracted from the power spectrum and Higuchi's Fractal Dimension (HFD) was also calculated. Additionally, subsets of features were created and selection was performed using either the highest correlation (2 features) or the Minimum Redundancy Maximum Relevance method (5,10,15 features). Results: As an example, we consider a subject with 90 minutes acquisition and 631 drowsy episodes, and calculate True Positive Rate (TPR), True Negative Rate (TNR), Accuracy, Cohen's Kappa (K). Performances are similarly satisfactory for the two algorithms and all the studied parameters (accuracy 78% at best, K=0.45 at best), with SVM generally outperforming LDA. Selecting a reduced number of features improves performances, as less significant features are removed. In particular, using 15 features from mRMR leads to the best result for the SVM, while the 5 mRMR seems the best compromise between classification performance and computational cost. Results obtained from this preliminary study are very promising: they indicate that the Qvar and artificial intelligence can effectively support wearable single-channel EEG recording in non-clinical settings.

Nr:	291
Title:	Leveraging ICT for Pediatric Diabetes Management: A Systematic Review and Recommendations for Developing Gamified mHealth Solutions
Authors:	Tahmineh Aldaghi, Karel Hana, Pavel Trnka and Jan Muzik
Abstract:	Background: Managing type 1 diabetes in children and adolescents presents challenges for parents, healthcare providers, and patients. Advances in information and communication technologies (ICT) have significantly improved the quality of diabetes care, particularly for young populations. Objective: This study aims to systematically document the spectrum of ICT tools utilized in pediatric diabetes management, highlighting the most frequently adopted technologies. Furthermore, it explores methodologies for developing diabetes-focused digital solutions and provides actionable recommendations for developers to enhance design processes. Methods: A systematic literature review was conducted using MEDLINE (PubMed), Web of Science, and Google Scholar. Keywords such as "type 1 diabetes," "adolescents," "kids," "mHealth," "children," and "coaching" were used. Inclusion criteria included open-access, English-language studies published between 2012 and 2023, focusing on patients younger than 18 years and aligned with our objectives. Following the PRISMA method, 33 papers met the inclusion criteria after rigorous review. Results: The analysis identified two main key areas, ICT types and methodology. Mobile health (mHealth) applications emerged as the dominant ICT type, cited 27 times, and user-centered design was the most frequently mentioned methodology (22 times). Conclusions: The literature review shows that while user-centered design remains the most widely adopted methodology for developing ICT solutions, gamification emerges as a more practical approach for designing tools tailored to adolescents and pediatric populations. Considering the target group's unique needs and engagement challenges, we propose developing a gamified mHealth application to revolutionize pediatric diabetes management. This proposed application would incorporate essential functionalities such as blood glucose monitoring, strategies for optimizing hemoglobin A1c levels, carbohydrate tracking, and comprehensive educational content designed for children and caregivers. By integrating gamification techniques, the tool can foster greater user engagement and adherence, making diabetes management more intuitive and motivating for young patients. This proposal aligns seamlessly with the conference's mission to advance healthcare through innovative ICT solutions.

Nr:	328
Title:	Introduction of Nanodiamonds into Meningeal Macrophages and Realization of in vivo Quantum Sensing
Authors:	Manami Takahashi, Kiichi Kaminaga, Yuta Masuyama, Chihiro Suzuki, Ayaka Takada, Hiroshi Abe, Takeshi Ohshima, Ryuji Igarashi and Hiroyuki Takuwa
Abstract:	Neurons, glial cells, and immune cells exhibit dynamic interactions in the brain and surrounding regions, leading to various changes in the intracellular and extracellular environments based on their activity. Multiparameter analysis of cells in vivo is a powerful approach to comprehensively understand cell functions and roles. For example, measurement of intracellular temperature and reactive oxygen species concentrations associated with metabolism, magnetic fields generated by neural currents, and pH dynamics in the cellular microenvironment can provide insights into brain homeostasis and pathogeneses of various conditions, such as neurodegenerative disorders and cancer. However, multiparameter measurement at single-cell level is challenging owing to the technical limitations of existing fluorescent tracers. Nanodiamonds containing nitrogen-vacancy (NV) centers have emerged as promising quantum sensors, sensitively and quantitatively measuring various physicochemical parameters in living cells owing to their high biocompatibility and sensitivity to magnetic and electric fields, temperature, and pH. Nanodiamond quantum sensors have been widely studied in in vitro environments, such as cultured cells; however, their application in vivo remains challenging. In this study, we aimed to improve the quantum-sensing technology for in vivo imaging and establish a system for multiparameter analysis of living brain cells. We developed a method to introduce nanodiamonds into mouse brain cells and optimized a microscopic system for in vivo quantum sensing. Nanodiamonds modified with hyperbranched polyglycerol to inhibit aggregation were injected into the cerebrospinal fluid of mice and their uptake by meningeal macrophages was investigated. In quantum sensing, visible light is used to excite the sensor, and fluorescence signals emitted by the sensor are subsequently detected. Mouse scalp was incised at the measurement site, the skull was removed, and an observation window was attached. The mice were secured under isoflurane anesthesia beneath the objective lens of a wide-field microscope, and rectal temperature was monitored to maintain constant body temperature. Immunostaining of the excised brains of nanodiamond-injected mice revealed that the hyperbranched polyglycerol-modified nanodiamonds were well-dispersed in the brain and intracellularly taken up by meningeal macrophages. Subsequently, temperature in the macrophages of the living mouse brain was measured by optically detected magnetic resonance (ODMR). By carefully optimizing the excitation light and microwave intensity, which significantly affect the detection sensitivity, temperature at the single-cell level was measured with an error margin of ±0.8 °C. To the best of our knowledge, this study is the first to monitor the localized temperature distribution in the living mouse brain with high sensitivity using nanodiamond quantum sensors. Additionally, non-radiative transitions enhanced by spin relaxation (T1 relaxation time) were successfully measured using the same nanodiamonds in cells. Measurement of T1 relaxation time can aid in free-radical detection. Overall, this study successfully developed an in vivo nano quantum sensor measurement system for the simultaneous analysis of multiple physicochemical changes in the brain. The developed system shows great potential for advancing life science and medical research.

Nr:	338
Title:	Optimization of Nanodiamond Size and Its Surface Functionalization for Intracellular Uptake by Glial Cells in the Living Brain
Authors:	Ayaka Takada
Abstract:	Nanodiamonds (NDs) with nitrogen vacancy (NV) centers are used as fluorescent probes in life science research owing to their high photostability and low cytotoxicity. Recently, NDs have gained attention as promising nanoscale sensors, often referred to as nano-quantum sensors, for the measurement of multiple physicochemical parameters, including intracellular temperature, radicals, and magnetic and electric fields. Their applications have been extensively studied in vitro; however, their specific effects on living tissues in vivo remain unclear. Efficient uptake of NDs by specific cell types is necessary to measure different parameters in the individual cells of living tissues. Therefore, effective methodologies are needed to advance the life science applications of these nano-quantum sensors. In this study, we aimed to optimize the intracellular uptake of NDs by glial cells (microglia and astrocytes) in the brain. Twelve types of NDs, with four different sizes (50, 100, 200, and 500 nm) and three surface functionalizations, were locally injected into the brain parenchyma of mice. Subsequently, intracellular uptake of NDs was evaluated via immunohistochemistry and confocal microscopy. Surface functionalization was performed using compound A, which enhances dispersion, compound B, which promotes cellular binding, or their combination (compound A + B). Microglia preferentially internalized the compound A-functionalized NDs, especially small NDs (50 nm), and showed the highest uptake rate. Compound A-functionalized NDs exhibited high diffusivity and effectively interacted with the surrounding microglia, which enhanced their uptake. In contrast, astrocytes exhibited significantly lower ND uptake rate than microglia, regardless of ND size and surface functionalization. Interestingly, compound B-functionalized NDs showed higher uptake than compound A-functionalized NDs in astrocytes, possibly due to their localization at the injection site, where phagocytic activity is increased by tissue damage-induced inflammation. Overall, our findings indicate that ND uptake characteristics differ depending on the cell type. Compound A-functionalized NDs, with high diffusivity, were more suitable for microglia, whereas compound B-functionalized NDs, which remained localized at the injection site, are more suitable for astrocytes. This study provides a foundation to further optimize the applications of nano-quantum sensors using NDs in the living brain. Future studies should explore other functionalizations, such as cell-specific antibody conjugation, for the selective targeting of specific glial cell types.

Nr:	345
Title:	End-to-End Neural Network for Direct Estimation of Each Foot's Ground Reaction Forces During Double Support from Kinematic Data Without a Decomposition Step
Authors:	Jungkeun Lee and Chang June Lee
Abstract:	Estimating ground reaction forces (GRFs) during walking is essential for kinetic gait analysis and inverse dynamics, enabling the calculation of joint reaction forces and moments. Recent studies have explored GRF estimation using kinematic data obtained from optical or inertial motion capture systems, removing the dependency on force plate measurements. While GRFs for each foot can be computed using Newton's equations of motion during single support, the process becomes underdetermined during double support. Traditionally, this challenge has been addressed through a two-step process: (i) estimating the combined GRF of both feet, and (ii) decomposing this combined GRF into individual GRFs for each foot, often relying on assumptions such as smooth transitions. In this study, we propose an end-to-end neural network for directly estimating each foot's GRF during double support from kinematic data, eliminating the need for a separate decomposition step. The proposed model is applicable to both level and inclined walking. To train the neural network, treadmill walking experiments were conducted with seven male participants (age: 24.7 ± 1.7 years, height: 1.75 ± 0.05 m, mass: 72.9 ± 6.9 kg) at two speeds (4 and 6 km/h) and three inclines (0%, 3%, and 6%). Full-body kinematics were captured using an optical motion capture system, and GRF data were recorded from an instrumented treadmill with a built-in force plate. This study focuses on optimizing model inputs and architecture to enable accurate GRF estimation for each foot during double support across both level and inclined walking conditions. The model’s performance will be evaluated using leave-one-subject-out cross-validation and compared with conventional GRF estimation methods.

Nr:	347
Title:	Cardiovascular Abnormality in Fibulin-4 Mutant Mice with Ascending Aortic Aneurysms
Authors:	Jungsil Kim, Woori Jo and Hee-Young Yang
Abstract:	Elastic fibers play a key role in maintaining normal cardiovascular function. Many genes are associated with the assembly of elastic fibers. In genetically engineered fibulin-4 mutant mice, elastic fibers are fragmented and disorganized. In particular, deletion of the fibulin-4 gene in smooth muscle cells (SMCs) leads to the postnatal development of an ascending aortic aneurysm within one month. Aneurysms of the ascending aorta are stabilized by 2-3 months of age with thicker aortic wall and cardiac hypertrophy. Therefore, in this study, we investigated the cardiovascular function of fibulin-4 SMC-specific deletion mice (Fbln4SMKO). The ascending aorta of Fbln4SMKO mice had a significantly larger diameter than that of wild-type mice (controls) over the entire pressure range from 0 to 160 mmHg. Ejection fraction (EF) and fractional shortening (FS) values, which indicate the contractile capacity of the heart, were significantly different in Fbln4SMKO mice compared to controls, but relaxation capacity, as indicated by E' and E/E' values, showed no significant difference between the two groups. Acknowledgement: This research was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (2021R1I1A3059740).

Nr:	348
Title:	Vascular Endothelial Cell Alignment by Micropatterning and Elastin in Microfluidic Vessels on a Chip
Authors:	Iksong Byun, Hoon Seonwoo and Jungsil Kim
Abstract:	Although traditional cell culture methods have provided valuable insights, they often fail to capture the organized architecture and mechanical properties of blood vessels. The development of in vitro blood vessel models that accurately replicate the complex structure and function of native vessels remains a significant challenge in vascular biology research. This study presents a novel in vitro blood vessel model on a chip that aligns human umbilical vein endothelial cells (HUVECs) using micropatterns and promotes endothelial differentiation by elastin coating. A PDMS-based microfluidic device was fabricated using 3D-printed molds and coated with a gelatin-elastin mixture that mimics the extracellular matrix of blood vessels. A computational fluid dynamics (CFD) analysis was performed to optimize the flow conditions within the chip. The viability and metabolic activity of HUVECs were evaluated using the LIVE/DEAD and WST-1 assays, respectively, after cultivation in the device. Evaluation of cell alignment and protein expression was performed by immunocytochemistry. The results showed high cell viability (>93%) with optimal performance observed in the 0.05% gelatin-elastin coating condition. The combination of micropatterns and protein coating significantly enhanced HUVEC alignment and cell-cell junction formation as demonstrated by immunofluorescence analysis. The aligned endothelial cells showed enhanced expression of key markers CD31 and eNOS, indicating a more in vivo phenotype. This model may provide a promising platform for drug screening and investigation of various mechanisms of vascular disease. Acknowledgement: This research was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (2021R1I1A3059740).

Nr:	354
Title:	Evaluation of the Effect of RNA Integrity on De Novo Assembly and Gene Expression of a Non-Model Plant Species Endemic to the Mongolian Gobi
Authors:	Batdelger Erdenetsetseg, Josephine Galipon, Munkhbat Uuganbayar, Kazuharu Arakawa, Sainbileg Undrakhbold, Shinji Fukuda and Bazartseren Boldgiv
Abstract:	The Gobi, located in the southern part of Mongolia, is the coldest Gobi in the world and is home to rare wild animals and plants such as khulan, argali, ibex, and leopard, as well as a rich biodiversity. Gobi plants also have the unique ability to adapt and resist extreme arid environments. A detailed study of the characteristics of Gobi plants is of great importance for the restoration and breeding of these plants. The target species of this study, Potaninia mongolica Maxim, is an endemic plant of Mongolia. It has been assessed as endangered species in the Red List of plants of Mongolia and listed as “Rare” The Law on Natural Plants of Mongolia and the Red Book of Mongolia are listed as endangered. The species is the only species of genus Potaninia in the Rosaceae family monotypic plant genus. The species has poor regeneration ability and germination in nature. The main goal of this study is to investigate how Gobi shrubs respond to irrigation and moisture treatment at the transcriptome level. In this work, we conducted irrigation experiments on P. mongolica, and extracted plant RNA samples over a 4-point time course, and did poly(A)-enriched transcriptome analysis for 3 replicates per time point. RNA integrity numbers (RIN) varied from 2.2 to 6.8 (cf. video). The proprietary RIN number (Agilent Technologies, Inc.) is the current default for evaluation of RNA quality for transcriptomics, but it was originally developed based on mammalian RNA from human, rat, and mouse [1]. Plant RIN is known to vary widely due to the presence of additional rRNA bands from chloroplasts. A previous study showed substantial variation in RIN between plant species, tissue types, and age, albeit concluding that plant transcriptomes are relatively robust to variations in RNA quality [2]. In addition, poly(A)-enriched RNA-seq data with lower RINs tend to introduce a bias with higher coverage of the 3'-ends, which leads to inaccuracies in quantification which are transcript-specific [3]. Previously, DegNorm was developed to compensate for this degradation effect based on human transcriptome and simulation data, but to the best of our knowledge, it has not been properly evaluated in non-model plants. [4]. In this conference, we will present a thorough evaluation of the effectiveness of (1) including or excluding RIN < 4 samples from the consensus de novo assembly, and (2) normalizing for RNA degradation on the accuracy of gene expression quantification in this plant. The results of this work will provide insights into how Gobi shrub plants respond to moisture. Additionally, it will provide other researchers with practical information on the challenges of de novo assembly and transcript quantification using RNA of variable quality, by providing a thorough evaluation of the usability of these normalizing tools in the case of a non-model plant.

Nr:	361
Title:	Proposal of Important Variables with Flux Sampling to Reduce Experimental Burden
Authors:	Yuki Kuriya, Masahiro Murata, Masaki Yamamoto, Naoki Watanabe and Michihiro Araki
Abstract:	Flux sampling using genome-scale metabolic models (GSMMs) has become more prevalent for comparing and analyzing the metabolic states of microorganisms and cells. Therefore, we proposed important variables for predicting metabolic states by flux sampling using GSMMs, and aimed to reduce experimental burden by measuring these variables. First, we used Escherichia coli GSMM iJO1366 as an example and performed flux sampling on acetate production from glucose by optGpSampler. Variables were limited to extracellular fluxes, and the importance of variables was evaluated by their ability to narrow down the solution. Among the important variables obtained, the CO2 flux was selected, and the CO2 flux value in 13C metabolic flux analysis (13C-MFA) was obtained from the literature. Using this value as a query, samples with values close to it were extracted from the flux sampling results. These samples were compared to the 13C-MFA result, suggesting that their intracellular flux distributions were similar. For Corynebacterium glutamicum, several common important variables with E. coli were obtained, but no appropriate 13C-MFA data were found for comparison. However, the important variables identified in this study are useful for predicting metabolic state, suggesting that their measurement may also be useful for validation with less experimental data.

Nr:	363
Title:	Characterization and Optimization of Cell Seeding in Electrospun Scaffolds by Mean of a Micro-Optofluidic Bioreactor
Authors:	Lorena Saitta, Emanuela Cutuli, Francesca Guarino, Gianluca Cicala and Maide Bucolo
Abstract:	Nowadays the demand for tissue regeneration is growing due to a scarcity of donors and biocompatibility issue in transplant immune rejection. For these reasons, scientists are investigating artificial tissues as alternative to regenerate damaged ones by combining cells with nanostructured scaffolds (Park et al. 2016). Scaffolds play a crucial role in tissue engineering, working as temporary skeleton for cells adhesion and proliferation when a tissue is damaged. Thus, in this context, the first step of cell culture within scaffolds is the cell seeding (Luyten et al. 2001). From the latter depends the initial cells quantity and their spatial arrangement, affecting in turn cell proliferation, migration, and the differentiation of the engineered implant (Schliephake et al. 2009). Hence, the cell seeding represents a crucial step in the development of efficient construct for regeneration purposes. In this context arise the need for the development of reliable seeding techniques. Among these, according to the literature, have been used static, rotational and perfusion protocols. The former is the most used one, consisting of passively introducing concentrated cell suspension into a scaffold placed in a well. However, some drawbacks are associated to this approach. First, depending on the scaffold’s density, these can float or move within wells during culture or media changes, making it difficult to maintain a consistent microenvironment. Second, if the used material is hydrophobic, providing the cells by using a pipette, the cell adhesion can be hindered. Third, the 3D porous constructs can interfere with conventional imaging techniques, so complicating quantitative analysis (cell counting). In this context, this study aims to introduce a micro-optofluidic (MoF) bioreactor, characterized by an encapsulated polyethersulfone scaffold, for cell seeding performed in a dynamic way to achieve good cell seeding efficiency (CSE) and cell spatial distribution (CSD). Thus, the latter responses were investigated by mean of a Design of Experiments (DoE) approach, where as influential factors were selected the microparticles type, the flow rate, imposed to feed the scaffold with the concentrated cell suspension, and the seeding time. The selected factors belong them affecting the quality of the seeded scaffold (Impens et al. 2010). In detail, the scaffold was fabricated via electrospinning, because it is a simple and cost-effective technique to realize non-woven polymeric architecture of nanofibers morphologically similar to the native extra-cellular matrix (Keshvardoostchokami et al. 2020), with a high surface-to-volume ratio and microporous morphology enabling cells adhesion. Next, the MoF device was realized using a 3D printing-based master/slave approach (Cutuli et al. 2023). It was made of polydimethylsiloxane for two reasons: it is characterized by nontoxicity and good gas permeability, which makes it suitable for biomedical applications as bioreactor in cell culture; it is transparent, so enabling its use for the optical detection. This feature is crucial since the two investigated responses, i.e., CSE and CSD were monitored, while varying in a systematic way all the selected design factors, by using an optical approach relying on the optical absorption (using optical fibers) and imaging techniques. The former relied on the investigation of acquired optical signals, whilst the latter on an image analysis approach.

Nr:	364
Title:	Biocompatible 3D-Printed Micro-Optofluidic Devices for Cell Concentration Monitoring
Authors:	Emanuela Cutuli, Lorena Saitta, Gianluca Cicala, Francesca Guarino and Maide Bucolo
Abstract:	The increasing demand for precision, efficiency, and scalability in life sciences and biomedical research has driven the development of advanced microfluidic and lab-on-a-chip (LOC) devices. Accurate cell concentration measurement is critical for applications like cell culture, drug development, diagnostics, and tissue engineering. Microfluidic devices utilize techniques such as flow cytometry, impedance spectroscopy, digital microfluidics, and image analysis methods, though challenges like invasiveness, complexity, and cost persist. Micro-optofluidic (MoF) devices address these limitations offering non-invasive, label-free, real-time monitoring by combining fluid manipulation with optical components, enabling light-based analytical capabilities. This work addresses issues of similar geometries previously fabricated using polydimethylsiloxane (PDMS) and soft lithography techniques, such as surface deformation due to swelling, lack of permanent bonding, low dimensional accuracy from PDMS' softness and thermal expansion, high fabrication costs, time-intensive procedures, and the ongoing investigation into the feasibility of single-step PDMS device printing (Saitta et al., 2022). In this context, this study aims to design, fabricate, and validate two MoF devices, 3D printed via Projection Micro-Stereolithography (PµSL). It enables the production of monolithic microstructures integrating optical and fluidic functionalities with micron-level precision (10µm), rapid turn-around times, and the ability to print complex shapes as single piece without the need for assembly or support materials. By using a novel biocompatible resin, the precision, optical transparency, and material compatibility for biomedical applications are achieved. The devices feature a T-junction microchannel for two-phase flow generation and micrometric dead-end slots for optical fiber insertion, enabling light-based sensing. They were designed at two different microscales: the first device has microchannels and optical fiber slots with 500µm square cross-sections, while for the second these components sizes are scaled down to 200µm. The working principles allow for two-phase flow monitoring and cell concentration detection, using optical absorption, where light transmission variations are correlated with fluid or cell interactions within the microchannels. Experimental validation demonstrated the devices' ability to detect air-water slug flows and measure yeast cell concentrations suspended in phosphate-buffered saline solutions. Key operational parameters, i.e., flow rates and laser input power, were optimized using a Design of Experiments methodology to ensure repeatability. A comparative study of the two MoF devices evaluated the effects of scaling down microchannel dimensions on optical and fluidic performance. Results showed that both devices were capable of distinguishing biological fluids with varying cell concentrations. The smaller showed superior sensitivity in detecting lower cell concentrations, highlighting the potential for developing devices with microchannels comparable to individual cells, enabling real-time cell counting based on optical signals. This study demonstrates the potential of PµSL-fabricated MoF devices for real-time, label-free, and non-invasive biosensing applications. It advances the integration of micro-optofluidic and biomaterials in LOC technologies, offering solutions for next-generation diagnostics and biomedical research.

Nr:	368
Title:	Sequence-Based Deep Learning Prediction of Nuclear Subcompartment-Associated Genome Architecture
Authors:	Shae McLaughlin, Sajad Ahanger and Daniel Lim
Abstract:	The spatial organization of the genome within the nucleus is partially determined by its interactions with distinct nuclear subcompartments, such as the nuclear lamina and nuclear speckles, which play key roles in gene regulation during development. However, whether these genome-nuclear subcompartment interactions are encoded in the underlying DNA sequence remains poorly understood. The mechanisms for gene regulation are primarily encoded in noncoding DNA sequences, but deciphering how these sequence features control gene expression remains a significant challenge in genomics. Here, we present Nucleotide GPT, a transformer-based model that predicts genomic associations with spatially distinct, physical nuclear subcompartments from DNA sequence alone. Pre-trained on a diverse set of multi-species genomes, we demonstrate Nucleotide GPT’s genomic understanding through evaluation on diverse prediction tasks, including histone modifications, promoter detection, and transcription factor binding sites. When finetuned to predict genome interactions with two separate nuclear subcompartments – the lamina of the inner nuclear membrane and nuclear speckles that lie more interior – Nucleotide GPT achieves an average accuracy of 73.6% for lamina-associated domains (LADs) and 79.4% accuracy for speckle-associated domains (SPADs), averaged across three cortical development cell types. Analysis of the model’s learned representations through Uniform Manifold Approximation and Projection (UMAP) reveals that Nucleotide GPT develops internal embeddings that effectively distinguish LADs from inter-LADs, with predicted probabilities closely corresponding to experimentally determined LAD classifications. When examining these representations in the context of cell type-invariant constitutive LADs (cLADs) compared to cell type-specific LADs, the model assigns lower confidence scores to cell type-specific LADs compared to cLADs that are conserved across neuronal differentiation, suggesting sequence features may play a stronger role in maintaining cLAD associations. Examination of the model’s attention patterns at correctly classified regions suggests that specific sequence elements govern model decision making about nuclear subcompartment associations. To further understand what sequence features the model learns during pre-training, we trained sparse autoencoders on the model's internal representations, decomposing them into human-interpretable latent features. Initial analysis reveals features that reliably detect specific DNA sequence motifs, including conserved elements like Alu sequences and transcription factor binding sites. Our results demonstrate the utility of transformer architectures for studying three-dimensional (3D) genome organization and substantiate a role for DNA sequence in determining nuclear subcompartment associations.

Nr:	372
Title:	Alkaline Phosphatase-Based ELISA and Particle-Based Microfluidic Test for Biotin Detection
Authors:	Naila Nasirova
Abstract:	Biotin is a naturally occurring vitamin and it also plays an important role in biochemical studies due to its strong and stable non-covalent bond with the protein avidin (dissociation constant KD = ~10-15 M) (Jain & Cheng, 2017). Estimation of biotin is important because biotin interference has become a reoccurring problem in clinical analysis. If the sample contains a critical amount of biotin, it will cause false positive and false negative test results on wide array of immunoassays (Li et al., 2020). We propose a new approach to utilize a well-known biochemical detection technique, the enzyme-linked immunosorbent assay (ELISA), by putting it on the developed particle-based microfluidic test to achieve high sensitivity. A new approach to develop particle-based microfluidic systems by using screen printing (SP) method to create patterned porous materials was utilized in our research group previously (Evard et al., 2021). In this study, we aim to simplify, reduce the analysis time, and spare the reagents by combining microfluidics with ELISA and develop a test for biotin (vitamin H/B7) detection. Alkaline phosphatase (AP) was chosen as the enzyme for ELISA. The developed test binds biotin with avidin and uses biotinylated alkaline phosphatase (bAP) for sensitive ELISA detection. Colorimetric detection is achieved by using a combination of coloring substrates for AP, nitro blue tetrazolium (NBT), and 5-bromo- 4-chloro-3-indolyl phosphate (BCIP). Evard, H., Priks, H., Saar, I., Aavola, H., Tamm, T., & Leito, I. (2021). A New Direction in Microfluidics: Printed Porous Materials. Micromachines, 12(6), Article 6. https://doi.org/10.3390/mi12060671 Jain, A., & Cheng, K. (2017). The Principles and Applications of Avidin-Based Nanoparticles in Drug Delivery and Diagnosis. Journal of Controlled Release : Official Journal of the Controlled Release Society, 245, 27–40. https://doi.org/10.1016/j.jconrel.2016.11.016 Li, D., Ferguson, A., Cervinski, M. A., Lynch, K. L., & Kyle, P. B. (2020). AACC Guidance Document on Biotin Interference in Laboratory Tests. The Journal of Applied Laboratory Medicine, 5(3), 575–587. https://doi.org/10.1093/jalm/jfz010.

Nr:	376
Title:	Multi-Branch CNN-Based Computer-Aided Diagnosis System for Major Depression Disorder Using Resting-State EEG
Authors:	Sanggyu Kim, Sihoon Hwang, Hosang Moon, Sungtaek Chung and Miseon Shim
Abstract:	The Major Depressive Disorder (MDD) is a mental disease that requires early diagnosis to ensure timely treatment and improve therapeutic outcomes. Previous studies have developed electroencephalography (EEG) based diagnostic tool using machine learning algorithms; however, their feasibility has been limited due to the need for hand-crafted feature extraction for these algorithms. Therefore, in the present study, we developed a practical EEG-based diagnostic tool to assist in the diagnosis of MDD patients using deep learning. To this end, eyes-closed resting-state EEG data were collected from 90 MDD patients and 90 gender- and age matched health controls. To reject artifacts, such as eye blinks, muscle artifacts, and cardiac signals, independent component analysis (ICA) was performed. The artifact-free EEG data were epoched into 60 seconds, which used as input data for deep learning model. In this study, a novel Convolutional Neural Network (CNN) model was proposed which can effectively extract the spatial and temporal features of EEG signals. Specifically, the model employs a multi-branch structure to extract frequency components, temporal features, and spatial features of EEG signals in parallel, integrating them to quantitatively classify MDD. To evaluate classification performance, the Leave-One-Out Cross-Validation (LOOCV) method was applied (a validation-to-training ratio of 1:11). As a result, the proposed model achieved an average classification accuracy of approximately 92%, demonstrating excellent generalization performance. This study represents significant step toward diagnosing MDD patients using resting-state EEG data to improve both efficiency and diagnostic performance. Moreover, this study highlights new possibilities for EEG-based depression diagnosis research.

Nr:	377
Title:	Development of Machine-Learning-Based Multi-Class Hand Gesture Recognition System Using Electromyography (EMG)
Authors:	Jongjin Lee, Sungtaek Chung, Hosang Moon, Han-Jeong Hwang and Miseon Shim
Abstract:	Non-verbal communication methods can serve as a way to facilitate communication for individuals with impaired language function or limited mobility. In particular, the physiological data-based communication systems can offer the more accurate and reliable communication performances than the behavioral communication methods. Previously developed physiological-data based communication systems primarily used electromyography (EMG) data recorded from larger arm muscles and did not address the detailed gestures made by smaller muscles in the hand. Therefore, in the present study, we developed the machine learning-based classification system to differentiate the multiple hand gestures using EMG signals. 16 participants (10 males, 6 females) were recruited to record EMG for this study. An Arduino Mega 2560 board-based EMG measurement device was used, with four sensors attached to the back of the hand. Participants were performed the nine distinct hand gestures and each gesture was repeated 20 times, using an inter-stimulus interval of about 1.5 seconds to minimize muscle fatigue. To minimize artifacts, both a 60-Hz notch filter and a 1-55 Hz band-pass filter were applied to the EMG data. Then, four time-domain features (root mean square, variance, and mean absolute value) were computed using artifact-free EMG data for each gesture, resulting in a total of 36 features. Linear Discriminant Analysis (LDA) was utilized to classify nine different gesture with a 5x5-fold cross-validation. We achieved the average classification accuracy of 80.09% across the nine hand gestures. This result demonstrates that EMG signals recorded from small muscles in the back of the hand can classify various hand gestures with relatively high accuracy. However, the classification accuracy should be improved to enhance practical feasibility by applying deep learning models as a future study.

Nr:	378
Title:	An AI-Driven Real-Time Fall Detection and Monitoring System for Elderly Assistance Using 4D Imaging Radar
Authors:	Museong Choi, Jo En Sang, Wang Eun Young, Hosang Moon, Miseon Shim and Sungtaek Chung
Abstract:	The global aging population and the increase in elderly individuals living alone present societal and economic challenges, including rising healthcare costs, vulnerability to accidents, and the need for effective emergency response systems. Falls are a significant concern, leading to severe injuries, reduced mobility, and psychological issues. Existing solutions like wearable devices and camera-based systems are limited by user discomfort, privacy concerns, and environmental dependencies. This study introduces a radar-based monitoring system that non-invasively collects Point Cloud data using 4D imaging radar. The radar operates within the 77–81 GHz frequency band, capturing spatial data (x, y, z), velocity, and time within a 12-meter range. Preprocessing steps, including zero-padding for short frames and K-means clustering for longer ones, standardize input length while preserving spatial structure. Sorted data enhances consistency for the AI model. The convolutional neural network (CNN) classifies three postures: standing, sitting, and lying. Using convolutional layers with Leaky ReLU activation, max-pooling, and fully connected layers, the CNN extracts spatial features. A SoftMax function classifies the postures, while dropout layers prevent overfitting. For fall detection, the system analyzes abrupt changes in the height (z max) and velocity of the body’s highest point, distinguishing falls from controlled lying actions. The system achieved a validation accuracy of 96.77% for posture classification, with precision, recall, and F1-scores exceeding 96%. The fall detection algorithm demonstrated strong reliability, identifying 45 out of 50 simulated falls with a recall of 90% and a precision of 91.84%. By ensuring privacy protection and accurate monitoring, this system effectively addresses the limitations of traditional methods. It enables rapid fall detection and emergency response, particularly for elderly individuals living alone, supporting independent living and enhancing their safety and quality of life.

Area 1 - Scale-IT-up

Nr:	380
Title:	Fostering Healthy Longevity Through Active Blended Learning: Testing an “Innovation and Entrepreneurship” Course Designed for Older Adults
Authors:	Burcu Demiray, Kathrin Inerle, Sina Berger, Miriam Wallimann and Zeynep Erden
Abstract:	Background: Lifelong learning plays a pivotal role in promoting healthy longevity by improving physical, mental, and cognitive health, as well as fostering social integration. However, educational programs for older adults rely on passive learning models, underutilizing older adults’ potential as active learners and contributors. Our applied project WiseLearn at the Healthy Longevity Center (University of Zurich) addresses this gap by offering older learners an age-friendly e-learning platform with blended courses on contemporary topics that allow them to apply what they learn. In collaboration with innovation management experts at Zurich University of Applied Sciences, we have tested the feasibility of a blended “design thinking” course for older learners (funded by DIZH; Digitalization Initiative of Zurich Higher Education Institutions). Findings from this feasibility study with five participants (Mean age = 64, one woman) showed that integrating active methodologies such as design thinking into lifelong learning not only improves older adults’ educational experience, but also empowers them to innovate and contribute meaningfully to society (publication under review). Participants valued the course’s modern topic, focus on innovation, and collaborative approach. They developed original ideas and built social-impact projects out of these ideas by using design thinking methods and tools (e.g., Podcast for senior citizens). Building on these promising preliminary results, we have been developing a full-scale course on “innovation and entrepreneurship”. Methods: This new course will be tested with a new group of older learners in February 2025. The course modules combine theory acquisition (online; e-learning) with practical application (on site) in a blended format. Both qualitative (e.g., interviews) and quantitative research methods (e.g., questionnaires) will be used to assess the accessibility and content qualities of the course, as well as learning experiences, satisfaction and performance. Furthermore, pre- and post-course measures will be used to examine psychological variables such as learners’ well-being, sense of belonging, and perception of their societal impact. By focusing on these psychological and social dimensions, the study aims to uncover the broader effects of such learning interventions on healthy longevity. Implications: This study seeks to establish the feasibility and impact of blended learning interventions tailored for older adults. Beyond assessing the technical and content accessibility, it explores the psychological benefits of active learning approaches. This research will provide critical insights into the potential for scaling up these interventions via a platform to foster healthy longevity. By empowering older adults as active learners and innovators, this approach holds the promise of transforming lifelong learning into a tool for prevention of decline, personal growth, and societal contribution.

Nr:	381
Title:	Promoting Healthy Aging by Co-Developing an Educational Digital Platform Against Ageism with Older Adults: A Use Case from Switzerland
Authors:	Miriam Wallimann, Kathrin Inerle, Andrea Ferrario, Erica Benz-Steffen and Burcu Demiray
Abstract:	Ageism encompasses stereotyping, prejudice, and discrimination based on an individual’s age, predominantly affecting older adults (Iversen et al., 2009). It has detrimental effects on physical, psychological and social health, contributing to premature mortality, decreased life quality, as well as increased healthcare costs (Levy et al., 2002, Chang et al., 2020). Despite urgent calls for action to counter ageism (WHO, 2021), research-based initiatives in Switzerland are still scarce. To close this gap, we are launching an anti-ageism campaign to promote healthy aging by (1) raising awareness and providing educational materials about ageism, (2) collecting real-life experiences from older adults, and (3) fostering a community to combat age-related discrimination. Key to achieve the goals of our campaign is an online digital platform—envisioned as a multi-purpose website—which is planned to be launched in Spring 2025. We are co-developing the digital platform with older adults, leveraging advantages of scalability and accessibility of digital technology. Design processes promoting user-centricity and co-development foster inclusion of real-life experiences of older adults, the identification of their needs and barriers, resulting in more robust and sustainable digital solutions (Duque et al., 2019). To co-develop the digital platform, we created a sounding board of N = 7 retired adults (M age = 69.71; four women) providing advisory input throughout the project. We collected their inputs in four focus group meetings and an online survey. Qualitative analyses identified major themes for each of the digital platform’s aims: empowerment of older adults and avoidance of victimization was recommended. Moreover, linking educational resources to actionable guidelines was highlighted. Then, motivators for sharing ageism experiences, such as anonymous reporting were brought up. Difficulty of recognizing ageism in everyday life was named a key challenge. Finally, the value of various online and offline interaction opportunities and active involvement in working towards educational and social objectives, including fostering intergenerational exchanges was emphasized. These preliminary findings highlight the importance of incorporating older adults’ perspectives in developing digital educational interventions to address ageism. They will inform real-life data collection from older adults and the design of educational interventions that will be promoted by the digital platform.

Nr:	382
Title:	Study Protocol: Integrating Non-Invasive Sweat Lactate Monitoring and Digital Health Support for Sepsis Management
Authors:	Karmen Markov, Sophie Vervullens, Maarten Gijssel and Andres Mellik
Abstract:	Background: Sepsis causes 11 million deaths annually and is challenging to diagnose, especially in resource-limited ICUs. Non-invasive sweat lactate monitoring offers a promising approach for early detection and management. Objective: This study evaluates the clinical effectiveness, feasibility, and safety of the IDRO sweat lactate biosensor integrated with the CoNurse digital platform for managing sepsis in ICUs and supporting post-discharge care. Secondary objectives include validating a predictive machine learning model for sepsis onset using multimodal inputs like sweat lactate and electronic health records. Methods: A prospective, multi-center cohort study will include 120 ICU patients at risk of sepsis and post-discharge care for recovering patients. Sweat lactate will be monitored twice daily using the IDRO device and compared with blood lactate levels. Clinical data, including SOFA scores, biomarkers, and vital signs, will validate a predictive model for sepsis onset. Post-discharge, CoNurse will support personalized care plans and caregiver engagement. Usability of both platforms will be assessed using standardized tools. Expected Outcomes: The study aims to validate sweat lactate monitoring and digital health integration for early sepsis detection and streamlined care. Findings will inform scalable, cost-effective strategies to enhance sepsis management in critical care settings.

Area 2 - SyntBioGen

Nr:	344
Title:	pyMDMA: An Open-Source Multimodal Framework for Enhanced Auditing of Real and Synthetic Data
Authors:	Ivo Façoco, Joana Rebelo, Pedro Matias, Nuno Bento, Ana Morgado, Ana Sampaio, Luís Rosado and Marília Barandas
Abstract:	Data auditing is essential for ensuring the reliability of machine learning models, as it safeguards the quality and integrity of the datasets used in their development. With the rising adoption of synthetic data to tackle challenges like data scarcity and privacy concerns, the need for a comprehensive and robust data auditing framework has become increasingly important. In this talk, we will present pyMDMA - Multimodal Data Metrics for Auditing real and synthetic data. It is an open-source Python library (https://github.com/fraunhoferportugal/pymdma) that provides metrics for evaluating both real and synthetic datasets across image, tabular, and time-series data modalities. It was developed to address gaps in existing evaluation frameworks which often lack metrics for specific data modalities, do not include certain state-of-the-art metrics, and do not provide a comprehensive categorization. pyMDMA provides a standard code base throughout all modalities, to make the integration and usage of metrics easier. The library is organized according to a new proposed taxonomy, that categorizes more clearly and intuitively the existing methods according to specific auditing goals for input (e.g., perceptual quality, uniqueness, correlation, etc.) and synthetic data (e.g. fidelity, diversity, authenticity, etc.). In particular, each metric class is organized based on the data modality (image, tabular, and time-series), validation domain (input and synthesis), metric category (data-based, annotation-based, and feature-based), and group (quality, privacy, validity, utility). We provide additional statistics for each metric result to help the user reach more concrete conclusions about the audited data. For each data modality, we will present practical use cases, demonstrating how to utilize this framework and prepare the data for evaluation.

Nr:	350
Title:	Deep Generative Models for Privacy-Preserving Clinical Data: Advancing Early Detection of Cardiac Decompensation and Pulmonary Exacerbations
Authors:	Aníbal Silva, Pedro Matias, César Gálvez-Barrón, Carlos Pérez-López, André Restivo, Moisés Santos, Carlos Soares and Marília Barandas
Abstract:	Heart failure (HF) and chronic obstructive pulmonary disease (COPD) are chronic conditions that significantly affect the general population, requiring early detection of decompensation or exacerbation to preserve individual health and mitigate disruptions to daily life (Boult et al., 1996). Machine learning (ML) models offer a promising alternative for early symptom detection but demand extensive training data, a critical challenge in the medical domain (Gálvez-Barrón et al., 2023). The collection of such data is labor-intensive, requiring efforts from patients and healthcare institutions. To address this challenge, Deep Learning (DL) generative models have emerged as a solution, synthesizing data with similar statistical properties to original datasets and creating artificial datasets that can be used for downstream tasks such as Machine Learning inference. Synthetic data not only aids in overcoming data scarcity but also addresses privacy concerns by retaining statistical properties while preventing individual identification (Hernandez et. al, 2022), assuring secure data sharing across clinical institutions. In this work, we investigate the potential of DL generative models to synthesize tabular data tailored to the specified conditions, focusing on generating privacy-preserving data to enable the early detection of exacerbation or decompensation phases. We consider two known families of DL generative models, specifically adapted versions of a Variational Autoencoder (VAE) (Fu et al., 2019) and a Generative Adversarial Network (GAN) (Arjovsky et al., 2017). Baseline approaches such as SMOTE (Chawla et al., 2011), and Probabilistic Sampling (PS) are considered for comparison. A small-sized dataset (with 252 samples) containing demographics and signal features from oximetry (SpO2) and heart rate (HR) measures collected from HF and COPD patients (Gálvez-Barrón et al., 2023) is chosen for evaluation. The prediction task targets the detection of decompensated heart failure (exacerbated), or compensated heart failure (stable) phases based on the signals’ traits. The synthetic data produced is evaluated using three families of metrics: 1) Statistical Fidelity (Kynkäänniemi et al., 2019) 2) Privacy-Preserving (Liu et al., 2024), and 3) ML Utility using a recently released data auditing library, pyMDMA (https://github.com/fraunhoferportugal/pymdma). Results show that the GAN and VAE architectures achieve a reasonable privacy-fidelity tradeoff while maintaining a stable ML utility. Specifically, the GAN achieved privacy and fidelity scores of 91.2% and 87.0%, respectively, while the VAE was the top performer, with privacy and fidelity scores of 95.5% and 93.0%. Comparisons with baselines show that while SMOTE achieved a fidelity of 100%, the data generated from this model resulted in a low privacy score of 5.5%. In contrast, while PS retained a high privacy score of 99.7%, the fidelity of the generated data w.r.t. the real one is relatively low (27%). In terms of ML Utility, SMOTE was the top performer, followed by the DL methods. To conclude, these findings highlight the potential of the GAN and VAE architectures as robust models for generating synthetic clinical data that balance privacy, fidelity, and utility, promoting efforts for safer and more effective use of synthetic data in healthcare applications.

Nr:	374
Title:	Designing for Qualitative Evaluation of Synthetic Health Data
Authors:	Isabella Silva, Elsa Oliveira and Ricardo Melo
Abstract:	Synthetic data has gained attention for its ability to meet the demand for large datasets to improve model performance, particularly in healthcare, where access to data is often limited due to privacy concerns, the scarcity of data for rare events, and high acquisition costs. However, synthetic data in healthcare still presents several challenges, such as the risk of amplifying bias, low interpretability and transparency, overgeneralization, and the misrepresentation of demographic and medical profiles. Additionally, there is a lack of robust methods for auditing data quality and a standardized framework for validating synthetic data [1][2]. A promising solution to these issues is the doctors-in-the-loop (DITL) approach, where doctors apply their knowledge to identify gaps and manage complex data [3]. This human interaction also enhances explainability reducing the black-box effect in ML systems [3]. This study aims to inform the design of an interface (DITL) approach for evaluating synthetic health data. By exploring strategies and challenges for integrating doctor feedback, we seek to enhance the assessment of synthetic data in healthcare. Our methodology included a literature review on human and doctor-in-the-loop methods, synthetic data in healthcare, and interactive machine learning, as well as workshops with 18 participants (7 doctors and 11 ML engineers). This paper highlights key findings from both the literature and workshop discussions, summarized as follows: 1. Ensure ethical and privacy standards. Although synthetic data can mitigate privacy risks, the potential for real data leakage persists and requires careful management [2]. A “privacy-by-design” approach can be adopted to ensure that there is no risk of re-identification of individuals [1]. 2. Expect healthcare stakeholders’ skepticism. Due to risk concerns doctors may question the reliability of diagnoses from models using synthetic data, especially given the high stakes for patients and providers [1]. These can affect their decisions when acting as evaluators. 3. Visual comparison between real and synthetic data. Human visual inspection frequently plays a key role in validating generative models. However, this approach can compromise privacy-by-design principles, as real data may reveal identifiable details or lead to re-identification [1]. 4. Sampling strategies to reduce human burden. Sample review presents model outputs for user evaluation [5]. This can involve manual selection by users or automatic suggestions based on model properties. A key challenge is balancing user effort by reducing queries while gathering enough feedback to improve the model. Effective strategies should focus on selecting diverse and representative samples for review, avoiding redundancy [5]. 5. Support understanding of model uncertainty. Machine learning models naturally involve uncertainty, which can be hard to convey in user interfaces [5]. Non-expert users may struggle to grasp the variability and potential errors in model outputs, so it is essential to clearly communicate uncertainty to help users set realistic expectations and make informed decisions. The insights from this study will inform the design of an interface for doctors to qualitatively assess synthetic health data. These findings will help improve interface strategies, strengthen design decisions, and highlight potential weaknesses or alternative approaches for further consideration.

Area 3 - uOrganChip

Nr:	379
Title:	A Bioelectronic Mucus Model Capable of Electrochemical Control and Monitoring Mucin’s Oxidative State
Authors:	Daniel Kaufman, Alexander Snezhko, William E. Bentley, Gregory Francis Payne and Hadar Ben-Yoav
Abstract:	The gut mucosa, composed mainly of mucin proteins, serves as a critical interface between the host and the microbiome. Mucin proteins possess sulfur-based functional groups that dynamically transition between thiol (–SH, reduced, "open" state) and disulfide (–S–S, oxidized, "closed" state) forms, modulating their structural properties. These thiol groups act as molecular switches, responding to the oxidative environment in the gut, influenced by reactive oxygen species (oxidants) and antioxidants (reductants). The redox state of mucins, regulated by both host and microbial factors, significantly impacts the mucus layer's barrier function and physiological role. Despite its importance, there is a lack of effective models to study the redox dynamics of the gut mucosa and the factors that influence it, leaving the interplay between host, microbiota, and external agents largely unexplored. To address this gap, we developed a novel bioelectronic system to control and monitor the thiol-based redox state of mucins electrochemically. This model employs a mucus-like structure formed by electrodepositing mucin-2 proteins encapsulated in a calcium alginate porous hydrogel onto gold electrodes. A redox conduit comprising oxidative and reductive species facilitates electron transfer between the electrode and the sulfur groups within mucin proteins, enabling real-time monitoring and precise redox control. Oxidative species with high standard reduction potentials promote disulfide bond formation by accepting electrons from thiol groups, transferring the electrons to the electrode, and generating an oxidative charge. Conversely, reductive species with low reduction potentials donate electrons to disulfide bonds, converting them back to thiols and altering the mucin's structural state. To validate this approach, we prepared electrodes coated with mucin-2 proteins (0.05 mg/ml in alginate) and performed cyclic voltammetry using 200 µM of the oxidative agent iridate (potential range: 0.1V to 0.8V vs. Ag/AgCl; scan rate: 0.1 V/s; 30 cycles). After every 10 cycles, the electrodes were treated with 50 µM Tris(2-carboxyethyl) phosphine hydrochloride (TCEP), a reductive agent, to reduce disulfide bonds. Initial voltammograms showed a decrease in oxidative charge over the first 10 cycles, indicating progressive thiol oxidation. However, upon TCEP treatment, a significant increase in oxidative charge was observed, reflecting the reduction of disulfide bonds back to thiols. This behavior was repeatable across subsequent cycles, demonstrating the model's reliability and controllability in regulating mucin redox states. These findings confirm that mucin biostructures can be electrochemically manipulated to monitor and control their redox states. This innovative gut model provides a robust platform to study how diet, inflammation, and disease influence the oxidative state of the mucus layer, paving the way for novel insights into gut physiology and redox-dependent processes in health and disease. [1] J. Li et al., Free Radic. Biol. Med., vol. 105, pp. 110–131, Apr. 2017.