CCH 2023 Abstracts

Area 1 - Cognitive Computing for Healthcare
Full Papers Short Papers

Area 1 - Cognitive Computing for Healthcare

Full Papers
Paper Nr: 7
Title:

Knowledge Graph Based Trustworthy Medical Code Recommendations
Authors:

Mutahira Khalid, Asim Abbas, Hassan Sajjad, Hassan A. Khattak, Tahir Hameed and Syed C. Bukhari
Abstract: Medical coding is about assigning standardized alphanumeric codes to diagnoses, procedures, and interventions recorded in patients’ clinical notes. These codes are essential for correct medical claims and billing processes, which are critical in maintaining efficient revenue cycles. Computer-Assisted-Coding (CAC) employs AI models to automate medical coding hence cutting down human effort and errors. Despite their unrivalled performance, these models lack ‘explainability’. Explainability opens up the inner workings and results of black-box deep learning models. Attention mechanisms are the most common approach for ‘explainability’, but they leave some questions unanswered, for instance, the relationship between highlighted words and predictions. Where black-box models fail to answer such questions, ‘Symbolic AI’ such as ‘Knowledge Graphs’ provide a superior alternate approach. We consolidated the attention mechanism with Symbolic AI to help users understand the results of a deep-learning model for CAC. We evaluated its performance on the basis of strong and weak relationships on word-to-word and word-to-code levels by employing a semantically-enriched Knowledge Graph. We achieved 64% word-to-word and 53% word-to-code level accuracy. This paper is among the earliest ones on knowledge graphs for explainability in medical coding. It is also the deepest in applying attention-based mechanisms and knowledge graphs to any medical domain.
Download
Paper Nr: 9
Title:

Personalized Semantic Annotation Recommendations on Biomedical Content Through an Expanded Socio-Technical Approach
Authors:

Asim Abbas, Steve Mbouadeu, Tahir Hameed and Syed C. Bukhari
Abstract: There are huge on-going challenges to timely access of accurate online biomedical content due to exponential growth of unstructured biomedical data. Therefore, semantic annotations are essentially required with the biomedical content in order to improve search engines’ context-aware indexing, search efficiency, and precision of the retrieved results. In this study, we propose a personalized semantic annotation recommendations approach to biomedical content through an expanded socio-technical approach. Our layered architecture generates annotations on the users’ entered text in the first layer. To optimize the yielded annotations, users can seek help from professional experts by posing specific questions to them. The socio-technical system also connects help seekers (users) to help providers (experts) employing the pre-trained BERT embedding, which matches the profile similarity scores of users and experts at various levels and suggests a run-time compatible match (of the help seeker and the help provider). Our approach overcomes previous systems’ limitations as they are predominantly non-collaborative and laborious. While performing experiments, we analyzed the performance enhancements offered by our socio-technical approach in improving the semantic annotations in three scenarios in various contexts. Our results show overall achievement of 89.98% precision, 89.61% recall, and an 89.45% f1-score at the system level. Comparatively speaking, a high accuracy of 90% was achieved with the socio-technical approach whereas the traditional approach could only reach 87% accuracy. Our novel socio-technical approach produces apt annotation recommendations that would definitely be helpful for various secondary uses ranging from context-aware indexing to retrieval accuracy improvements.
Download
Paper Nr: 12
Title:

An NLP-Enhanced Approach to Test Comorbidities Risk Scoring Based on Unstructured Health Data for Hospital Readmissions Prediction
Authors:

Tahir Hameed, Haris J. Khan, Saad Khan, Mutahira Khalid, Asim Abbas and Syed C. Bukhari
Abstract: Hospital readmissions have emerged as a key healthcare quality indicator since the passing of the Affordable Care Act in 2010. It is easier to predict the readmission risk of patients without complications, but comorbidities, such as diabetes and cardiovascular diseases, make it difficult to accurately assess the readmission risk. 30-days hospital readmissions (30DRA) risk models typically rely on demographic, socioeconomic, and medical variables from structured data, such as diagnosis, vitals, lab reports, and comorbidities, etc. Comorbidity indices help in assessing overall disease burden by accounting for the disease codes in electronic health records (EHRs). With the advent of natural language processing (NLP), there is a potential to extract additional health related variables including the possibility of gleaning additional disease codes for comorbidities in unstructured portion of the EHRs, such as clinical notes, medical history, and discharge summaries. Whereas NLP has been applied heavily in healthcare information systems, to the best of our knowledge, there is no research that identifies comorbidities from unstructured clinical texts. This paper employs a Bidirectional Encoder Representation from Transfer (BERT) deep learning technique to predict additional comorbid conditions in the unstructured portions of EHRs and evaluates the effectiveness in comorbidity scoring. Comorbidity scores based on the NLP-predicted comorbidity codes (predicted) were compared against the scores calculated from codes identified by the health providers (diagnosed), and also against a combination of the two (diagnosed and predicted). We find NLP is effective in improving the accuracy of comorbidity calculations, that in turn could improve predictive power of AI models for hospital readmissions and mortality predictions. It is among the first papers employing NLP to predict ICD-10 codes from unstructured EHRs for comorbidity index calculations.
Download
Short Papers
Paper Nr: 5
Title:

Integration of a Deep Learning-Based Module for the Quantification of Imaging Features into the Filling-in Process of the Radiological Structured Report
Authors:

Camilla Scapicchio, Elena Ballante, Francesca Brero, Raffaella Fiamma Cabini, Andrea Chincarini, Maria Evelina Fantacci, Silvia Figini, Alessandro Lascialfari, Francesca Lizzi, Ian Postuma and Alessandra Retico
Abstract: The role of Computed Tomography (CT) in the characterization of COVID-19 pneumonia has been widely recognized. The aim of this work is to present the idea of integrating a Deep Learning (DL)-based software, able to automatically quantify qualitative information typically describing COVID-19 lesions on chest CT scans, into a structured report-filling pipeline. Different studies have highlighted the value of introducing the use of structured reports in clinical practice, as a reproducible instrument for diagnosis and follow-up rather than the commonly used free-text radiological report. Structured data are fundamental to helping clinical de- cision support systems and fostering precision medicine. We developed a Deep Learning based software that segments both the lungs and the lesions associated with COVID-19 pneumonia on chest CT scans and quan- tifies some indexes describing qualitative characteristics used to assess COVID-19 lesions clinically. Once assessed the robustness of the system by means of a multicenter clinical evaluation made by clinical experts, it can be used for the first stratification of patients, supporting radiologists with a computer-aided quantification, and the derived quantities, immediately intelligible for the clinicians, are suitable to be inserted in a structured report in COVID-19 pneumonia and then exploited as explainable features to build predictive models.
Download
Paper Nr: 6
Title:

A Convolutional Neural Network Model for Prediction of ICU Performance Metrics: Time Series and Image Transformation Approaches
Authors:

Ömer K. Karahan, Yasin Ulukuş and Çiğdem E. Erdem
Abstract: In our study we used Convolutional Neural Network (CNN) to predict Intensive Care Unit (ICU) performances of patients via images generated from patients’ Sequential Organ Failure Assessment (SOFA) scores which are used to assess the acute morbidity of intensive care unit patients. In our study we propose a novel method to predict ICU performances; mortality during the stay in ICU, mortality in one year after discharge from ICU, readmission and length of stay of ICU patients. We trained CNN models with images generated from multivariate time series data. Our model development process consists of two steps; converting SOFA scores of patients into an image and training the CNN with generated images to predict patients’ ICU performances. We search for the best performing image generation algorithm which has the highest AUROC value for each prediction. Our model gives us AUROC values for mortality in ICU, readmission after discharge from ICU and length of stay of patients in ICU as 0.83, 0.84, 0.87, 0.56 respectively. We compare our methods’ performance with random forest, support vector machine, Logistic regression and ensemble of these algorithms. The proposed image-based method in which we use the first day SOFA scores outperform the random forest, support vector machine and logistic regression algorithms. Our method performed similar to the studies in literature in terms of predicting mortality in ICU using first day data with an AUROC value of 0.83. Our model’s performance would be improved with further feature engineering.
Download
Paper Nr: 8
Title:

Short-Term and Long-Term Readmission Prediction in Uncontrolled Diabetic Patients using Machine Learning Techniques
Authors:

Monira Mahmoud, Mohamed Bader and James McNicholas
Abstract: Diabetes is a chronic disease and major health problem which leads to many complications if not managed probably. Hyperglycemia, or raised blood sugar, is a common effect of Uncontrolled diabetes that may leads overtime to serious complications, especially in the nerves and blood vessels. As well as leads to repeated hospital admission. The main purpose of this study is to help clinicians to improve healthcare of uncontrolled diabetic patients through using machine learning as a tool in decision making, consequently this will improve patient care and reduce the readmission which considered a medical quality measurement and cost reduction objective. This study aims to predict the hospital readmission of the uncontrolled diabetic patient who is considered more susceptible to developing life-threatening diabetes complications and based on the Diabetes 130-US hospitals dataset. Several machine learning employed to predict the short term (within 30 days), and both short and long-term readmission (within or after 30 days) of uncontrolled diabetic patient. As expected, the results are in line with other research in the literature. For the first scenario of whole readmission prediction, our model achieved a better accuracy of 64.5 % with SVM and attribute selection and for the second scenario, RF achieved the highest accuracy of 86.38 % which still come in context with other research in the literature.
Download
Paper Nr: 10
Title:

Attention-Based Explainability Approaches in Healthcare Natural Language Processing
Authors:

Haadia Amjad, Mohammad S. Ashraf, Syed A. Sherazi, Saad Khan, Muhammad M. Fraz, Tahir Hameed and Syed C. Bukhari
Abstract: Artificial intelligence (AI) systems are becoming common for decision support. However, the prevalence of the black-box approach in developing AI systems has been raised as a significant concern. It becomes very crucial to understand how an AI system makes decisions, especially in healthcare, since it directly impacts human life. Clinical decision support systems (CDSS) frequently use Natural Language Processing (NLP) techniques to extract information from textual data including Electronic Health Records (EHRs). In contrast to the prevalent black box approaches, emerging ’Explainability’ research has improved our comprehension of the decision-making processes in CDSS using EHR data. Many researches use ’attention’ mechanisms and ’graph’ techniques to explain the ’causability’ of machine learning models for solving text-related problems. In this paper, we conduct a survey of the latest research on explainability and its application in CDSS and healthcare AI systems using NLP. For our work, we searched through medical databases to find explainability components used for NLP tasks in healthcare. We extracted 26 papers that we found relevant for this review based on their main approach to develop explainable NLP models. We excluded some papers since they did not possess components for inherent explainability in architectures or they included explanations directly from the medical experts for the explainability of their work, leaving us with 16 studies in this review. We found attention mechanisms are the most dominant approach for explainability in healthcare AI and CDSS systems. There is an emerging trend using graphing and hybrid techniques for explainability, but most of the projects we studied employed attention mechanisms in different ways. The paper discusses the inner working, merits and issues in the underlying architectures. To the best of our knowledge, this is among the few papers summing up latest explainability research in the healthcare domain mainly to support future work on NLP-based AI models in healthcare.
Download