CRII: SCH: Improving translation of deep learning in clinical ECG diagnostics

About This Grant

Electrocardiograms (ECGs) are extensively used by clinicians for the detection and monitoring of cardiac conditions and constitute an essential element of both routine and emergency cardiac care. Despite recent advancements in deep learning (DL), which have demonstrated high performance in the automatic diagnosis and interpretation of medical conditions from ECG signals, the real-world use of these models in healthcare remains limited. This is due to several key challenges, such as the lack of sufficient data for rare but clinically important heart conditions, the difficulty of understanding how these “black-box” models make decisions, and the absence of explanations that align with the diagnostic reasoning of clinical experts. This project seeks to overcome these challenges in collaboration with clinicians to build DL systems that can generate synthetic ECG signals for rare cases, enhance diagnostic reliability by identifying instances of model uncertainty, and produce comprehensive explanation reports that reflect the clinical workflow of ECG diagnosis. These explanations are intended to assist clinicians in understanding the reasoning behind deep-learning model decisions, thereby increasing trust and usability in real clinical settings. The results of this research are expected to improve the quality and transparency of DL-based decision support systems, enrich training for future healthcare professionals through tools that align with their diagnostic processes, and broaden access to accurate and interpretable diagnostic technologies. This project aims to facilitate the integration of deep learning models for Electrocardiograms(ECG)-based detection and monitoring of cardiac conditions by addressing the predominant challenges of data shortage for rare conditions, lack of alignment and interpretation with clinician diagnostic processes, and model uncertainty. This project will address four fundamental research challenges: 1) Multimodal learning: jointly learning signal-feature embedding space for controlled generation of synthetic ECG signals based on SCP statements, 2) Model uncertainty: dynamic and fluid prototypical spaces for improved reliability through abstention reducing false positives and negatives, 3) Rare ECG data: controlled ECG signal generation from SCP-ECG (Standard Communications Protocol) statements to alleviate data challenges, and 4) Clinically-aligned explanations: generation of comprehensive clinical and technical contextual, counterfactual, and contrastive explanations for classification outcomes. These four research aims will be complemented by clinical collaborators to ensure translation in real-world clinical settings and dissemination in the clinical community. The outcomes are anticipated to facilitate the integration of interpretable and reliable deep learning systems into clinical settings, paving the way for future integration of other 1D time-series signals in clinical workflow alignment. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.