SyntBioGen 2025 Abstracts

Area 1 - SyntBioGen
Full Papers

Area 1 - SyntBioGen

Full Papers
Paper Nr: 5
Title:

Intraoperative Electrocorticography Signal Synthesis to Improve the Classification of Epileptiform Tissue
Authors:

Leonor Almeida, Sem Hoogteijling, Inês Silveira, Dania Furk, Irene Heijink, Maryse Van’T. Klooster, Hugo Gamboa, Luís Silva and Maeike Zijlmans
Abstract: Epilepsy surgery is a viable option for treating drug-resistant cases where anti-seizure medications fail, but accurately localizing epileptic tissue remains challenging. This process can be guided by the visual assessment of intraoperative electrocorticography (ioECoG). Data scarcity limits developing machine learning (ML) models for automatic epileptic tissue classification. To address this, we propose a generative model based on Generative Adversarial Networks (GANs) to synthesize realistic ioECoG signals. Our approach identified three distinct ioECoG patterns using Agglomerative Clustering, which guided training individual Deep Convolutional Wasserstein GANs with Gradient Penalty (DCwGAN-GP). Synthetic data (SD) was evaluated across multiple dimensions: fidelity using temporal (e.g., Wasserstein distance (WD)), frequency and time-frequency metrics; diversity through dimensionality reduction; and utility by comparing ML performance with and without SD. It replicated temporal and frequency characteristics of real signals (fidelity), though lacked variability (diversity) due to potential data misclassifications. Specifically, the WD between real and synthetic signals outperformed literature benchmarks (i.e., 0.043 ± 0.025 vs. 0.078). Classifiers trained on a combination of real and SD achieved 88% accuracy, compared to 85% with real data alone. These results demonstrate the potential of SD to replicate real signals, address data scarcity, augment ioECoG datasets, and advance ML-based epilepsy surgery research.
Download
Paper Nr: 6
Title:

Towards High-Fidelity ECG Generation: Evaluation via Quality Metrics and Human Feedback
Authors:

Maria Russo, Joana Rebelo, Nuno Bento and Hugo Gamboa
Abstract: Access to medical data, such as electrocardiograms (ECGs), is often restricted due to privacy concerns and data scarcity, posing challenges for research and development. Synthetic data offers a promising solution to these limitations. However, ensuring that synthetic medical data is both realistic and clinically relevant requires evaluation methods that go beyond general quality metrics. This study aims to overcome such challenges by advancing high-fidelity ECG data generation and evaluation, presenting an approach for generating realistic ECG signals using a diffusion model and introducing a novel evaluation metric based on a deep learning evaluator model. The state-of-the-art Structured State Space Diffusion (SSSD-ECG) model was refined through hyperparameter optimization, and the fidelity of the generated signals was assessed using quantitative metrics and expert feedback. Complementary evaluations of diversity and utility ensured a comprehensive assessment. The evaluator model was developed to classify individual synthetic ECG signals into four quality classes and was trained on a custom-developed quality dataset designed for the generation of 12-lead ECG signals. Results demonstrated the success in generating high-fidelity ECG data, validated by evaluation metrics and expert feedback. Correlation studies confirmed an alignment between the evaluator model and fidelity metrics, highlighting its potential as a valid tool for quality assessment.
Download
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.