Analysis of a Deep Learning Model Quantifying Aging Effects in the Electrocardiogram: Is Data From > 1.5 Million Patients Enough to Train a Fair Model?

Abstract

Deep learning (DL) models show remarkable results which also holds true for the field of biomedical time series. These time series are measured via techniques such as electrocardiography (ECG) or phoplethsymography and are a cornestorner of cardiovascular health assesment. Traditionally, these signals were analyzed by medical experts using visual inspection following rules defined in clinical guidelines. In recent years, DL models have begun to demonstrate a level of accuracy and reliability that rivals that of human experts in a variety of tasks such as diagnosis and risk assessment.

Additionally, models have recently been published that detect novel biomarkers which clinicans are not trained to detect or are not aware of. One of these biomakers is based on aging effects within the ECG and is linked to cardiovascular risk and mortality. Focusing this biomaker, we used the longitudinal and population-based data of the Study of Health in Pomerania (SHIP) to perform an in-depth evaluation of a state-of-the-art AI model trained on millions of ECGs signals w.r.t. aspects such as fairness, robustness, and interpretability.

Our results demonstrate that the model aligns with clinical knowledge by focusing on ECG features known to reflect aging and generalizes well on population-based study data in terms of accuracy. However, in special edge cases the predictions are biased with the risk of discrimative predictions. Therefore, although these AI models contain the potential of eventually being used on population level to identify patients at risk early, they must be carefully evaluated to ensure fair predictions.

Short CV

Nicolai Spicher is an Associate Professor at the Department of Health Technology, Technical University of Denmark since 2025. Before that, he was a Junior Research Group Leader at the University Medical Center Göttingen (2022-2024) and completed postdoctoral training at the Peter L. Reichertz Institute for Medical Informatics at TU Brauschweig and Hannover Medical School (2020-2022). He has a background in computer science and holds a Ph.D. degree from University Duisburg/Essen where he was a member of the Erwin L. Hahn Institute for Magnetic Resonance Imaging (2015-2020). His research interests include signal processing and machine learning for cardiovascular applications.