HeFDI Data Week 2025: From Simulation to Prediction. AI and Synthetic Data in Biomedicine

HeFDI Data Week 2025 Abstract: Synthetic data, such as electrocardiograms (ECGs) generated through electrophysiological simulations, can be used as a powerful tool to overcome the challenges of accessing large, high-quality datasets in healthcare, which are often difficult to obtain due to privacy concerns, variability and challenges in annotation. This talk will look at how synthetic data can complement real-world datasets, improve model generalisation, facilitate learning in rare or underrepresented scenarios as well as improve the accuracy and reliability of machine learning models in biomedical applications. Further, examples of using such data to train deep learning models for various biomedical applications will be discussed together with potentials and key challenges. Ultimately, the talk explores how synthetic data can help to bridge the gap between computational modeling and real-world clinical AI applications. About theHeFDI Data Week: The HeFDI Data Week 2025 is a multi-day online event series which is offered in the context of the nationwide "Digitaltag". The series is aimed at researchers, teachers, students and anyone who wants to learn more about data management, FAIR data and code. Over the course of the week, various topics, developments and challenges related to research data will be covered, such as tools and offers for disciplines from NFDI consortia, legal aspects of research data management as well as research data management and artificial intelligence. The HeFDI Data Week is an offer of the federal state initiative HeFDI - Hessian Research Data Infrastructures, which is funded by the Hessian Ministry of Higher Education, Research, Science and the Arts (HMWK). DOI-Link: https://doi.org/10.5281/zenodo.15756218. Licence information: Creative Commons Attribution 4.0 International (CC BY 4.0) Date Topic Presenter 27. June 2025 From Simulation to Prediction: AI and Synthetic Data in Biomedicine Silvia Becker (Institute of Biomedical Engineering (IBT)) gefördert durch From Simulation to Prediction: AI and Synthetic Data in Biomedicine Silvia Becker Institute of Biomedical Engineering (IBT) Content 1. Opening Example 2. Challenges of AI in Biomedicine 3. Background ▪Cardiac Modeling ▪Digital Twins & Chimera 4. Solution: Synthetic Data ▪Potentials ▪Challenges 5. Clinical Applications 6. Take Home Messages Opening Example 27/06/25 Silvia Becker - AI and Synthetic Data in Biomedicine4 Patient with rare condition x ECG Analysis Patient with treated condition x Patient with rare condition x Can we also diagnose this person? How can AI help us? Challenges of AI in Biomedicine Missing ground truth datasets with adequate size and high quality labels Big data but small data Biased, unbalanced datasets Do we capture the relevant range? Fairness Intransparent methodology FAIRness, partly due to data privacy Explainability 27/06/255 Silvia Becker - AI and Synthetic Data in Biomedicine www.biomeris.it Challenges of AI in Biomedicine 27/06/256 Silvia Becker - AI and Synthetic Data in Biomedicine Atrial fibrillation (AF) is the most common sustained arrhythmia Healthy case: Sinus rhythm Pathology: Atrial fibrillation Challenges of AI in Biomedicine 27/06/257 Silvia Becker - AI and Synthetic Data in Biomedicine Ablation is a standard procedure to treat AF Specifically pulmonary vein isolation (PVI) But: high recurrence rates after PVI Who benefits from PVI? →AI for the rescue? Is there enough data available? Atrial fibrillation Atrial fibrillation ? ECG PVI Hopkins, CC BY-NC-ND Challenges of AI in Biomedicine 27/06/258 Silvia Becker - AI and Synthetic Data in Biomedicine Luongo et al., 2021, J Cardiovasc Digital Health PVI Simulations to create 1,128 synthetic ECGs Simulations with AF drivers in different locations Ground truth available (within PVs = PVI success) Classifier was trained on this synthetic data And tested on a clinical dataset 82.6% specificity and 73.9% sensitivity Cardiac Modeling 27/06/259 body surfacemyocyteion channel tissue / organ nanometer micrometer centimeter biological level of integration Silvia Becker - AI and Synthetic Data in Biomedicine 27/06/2516 MRI/CT Electroanatomical mapping Electrical behavior on the patient’s heart’s surface ECGs that represent this patient’s heart activity on the body surface EHR Age, sex, BMI, diagnosis, … Image Output Statistical/mathematical models OR Anatomical model Functional model Virtual chimera A B C Digital Twins Silvia Becker - AI and Synthetic Data in Biomedicine Simulations Digital Twins 27/06/2517 Nagel et al., 2021, Medical Image Analysis Azzolin et al., 2023, Computerized Medical Imaging and Graphics Anatomical personalization Silvia Becker - AI and Synthetic Data in Biomedicine 27/06/2518 MRI/CT Electroanatomical mapping Electrical behavior on the patient’s heart’s surface ECGs that represent this patient’s heart activity on the body surface EHR Age, sex, BMI, diagnosis, … Image Output Statistical/mathematical models OR Anatomical model Functional model Virtual chimera A B C Digital Twins Silvia Becker - AI and Synthetic Data in Biomedicine Simulations 27/06/2519 MRI/CT Electroanatomical mapping Electrical behavior on the patient’s heart’s surface ECGs that represent this patient’s heart activity on the body surface EHR Age, sex, BMI, diagnosis, … Image Output Statistical/mathematical models OR Anatomical model Functional model Virtual chimera A B C Digital Twins Silvia Becker - AI and Synthetic Data in Biomedicine Simulations Digital Twins 27/06/2520 Functional personalization Hopkins, CC BY-NC-ND Silvia Becker - AI and Synthetic Data in Biomedicine Healthy tissue Fibrotic tissue 27/06/2521 MRI/CT Electroanatomical mapping Electrical behavior on the patient’s heart’s surface ECGs that represent this patient’s heart activity on the body surface EHR Age, sex, BMI, diagnosis, … Image Output Statistical/mathematical models OR Anatomical model Functional model Virtual chimera A B C Digital Twins Silvia Becker - AI and Synthetic Data in Biomedicine Simulations Synthetic ECGs 27/06/2522 Atrial fibrillation ECGElectric potentials on the body surface Luongo et al., 2021, J Cardiovasc Digital Health Silvia Becker - AI and Synthetic Data in Biomedicine 27/06/2523 Cardiac Modeling Controlled research environment to understand fundamental physiology and pathomechanisms to evaluate diagnostic and therapeutic approaches (in silico studies) Scalable research environment to generate big, qualitycontrolled datasets for machine learning Personalized medicine through digital twins Silvia Becker - AI and Synthetic Data in Biomedicine Digital Chimera 27/06/2524 anatomical envelope Digital Chimera Adapted from Zeike Taylor, University of Leeds functional envelope Silvia Becker - AI and Synthetic Data in Biomedicine Digital Chimera 27/06/2525 Anatomical personalization Silvia Becker - AI and Synthetic Data in Biomedicine Statistical shape models (SSM) describe the shape variability of an object among a population. Clinical Applications 27/06/2532 Silvia Becker - AI and Synthetic Data in Biomedicine Covering relevant variability Shape models Clinical Applications 27/06/2533 Silvia Becker - AI and Synthetic Data in Biomedicine Closing the domain gap MedalCare-XL Gillette et al., 2023, Sci Data Clinical Applications 27/06/2534 Silvia Becker - AI and Synthetic Data in Biomedicine Closing the domain gap MedalCare-XL Gillette et al., 2023, Sci Data Clinical Applications 27/06/2535 Silvia Becker - AI and Synthetic Data in Biomedicine Closing the domain gap MedalCare-XL Gillette et al., 2023, Sci Data Clinical Applications 27/06/2536 Martinez Diaz al., 2024, Computing in Cardiology Hospital 10 90 (mV) 12.8±4.3 hours Vulnerability Assessment 1-2 hours Biatrial Model Generation 2-3 hours Electrophysiological Study >15 hours Personalized Digital Twin Assessment XSubstantial time XHPC dependent Silvia Becker - AI and Synthetic Data in Biomedicine Reducing computational costs Clinical Applications 27/06/2537 Martinez Diaz al., 2024, Computing in Cardiology Model Gen. Simulation MLFeatures Explainable Data 3. ML Data Analysis Random Forest Combination of simulations and machine learning reduces computational costs Silvia Becker - AI and Synthetic Data in Biomedicine Clinical Applications 27/06/2538 Martinez Diaz al., 2024, Computing in Cardiology Silvia Becker - AI and Synthetic Data in Biomedicine True Positive Rate 1.0 0.8 0.6 0.4 0.2 0.0 0.2 0.4 0.6 0.8 1.0 Heat Map for Inducibility Prediction 1 0 Inducibility Anterior Posterior Receiver Operating Characteristic (ROC) Curve Do NOT simulate Clinical Applications 27/06/2539 Silvia Becker - AI and Synthetic Data in Biomedicine EHRA (European Heart Rhythm Association) checklist for AI in clinical electrophysiology Data collection and source of data Information about training and test data and data split Definition of gold standard and ground truth Data preparation, data issues Balanced groups Performance (+compared to classic statistical methods) Generalizability … Svennberg et al., 2025, Europace Rare condition ▪ECG features to diagnose might be unclear Use AI Few data available Use synthetic data to increase dataset ▪Usage of AI enables screening ▪Objective, scalable, reproducible 27/06/2540 Opening Example Patient with rare condition x Can we also diagnose this person? How can AI help us? Silvia Becker - AI and Synthetic Data in Biomedicine Take Home Messages 27/06/2541 AI is transforming biomedicine –but data is the fuel Synthetic data is a promising tool To increase training data and balance it To understand pathomechanisms To reduce animal experiments & clinical trials To live up to the FAIR principles Silvia Becker - AI and Synthetic Data in Biomedicine