Transferring Clinical Knowledge into ECGs Representation

arXiv ID: 2512.07021v1

Published: 2025-12-07

Authors: Jose Geraldo Fernandes, Luiz Facury de Souza, Pedro Robles Dutenhefner, Gisele L. Pappa, Wagner Meira

Categories: cs.LG, cs.AI

Relevance Score: 1.00 / 1.00

View on arXiv Download PDF

Summary

This paper introduces a novel three-stage deep learning paradigm to overcome the "black box" nature of ECG classification models by transferring knowledge from multimodal clinical data (lab exams, vitals) into a unimodal ECG encoder. The approach creates an ECG representation enriched with contextual clinical information, leading to improved diagnostic accuracy and providing physiologically grounded explanations through the prediction of associated laboratory abnormalities directly from the ECG.

Medical Relevance

This work is highly relevant to medical practice as it directly tackles the major barrier of AI adoption in healthcare: trust and interpretability. By providing accurate ECG diagnoses alongside physiologically grounded explanations, it enables safer, more effective, and ethically sound integration of AI into clinical decision-making processes.

AI Health Application

Developing advanced, interpretable, and trustworthy deep learning models for automated ECG analysis and multi-label diagnosis classification. The AI application aims to enhance diagnostic accuracy, predict associated laboratory abnormalities from ECGs, and provide physiologically grounded explanations to facilitate the safe and effective integration of artificial intelligence into clinical workflows and medical decision-making.

Key Points

Addresses the critical issue of deep learning models' 'black box' nature, which hinders clinical adoption due to lack of trust and interpretability in ECG classification.
Proposes a novel three-stage training paradigm to transfer clinical knowledge from multimodal data (lab exams, vitals, biometrics) into a unimodal ECG encoder.
Utilizes a self-supervised, joint-embedding pre-training stage to create an ECG representation enriched with contextual clinical information.
A key practical advantage is that the model only requires the ECG signal at inference time, despite leveraging multimodal data during training.
Enhances interpretability by training the model to predict associated laboratory abnormalities directly from the ECG embedding, transforming abstract predictions into 'physiologically grounded explanations'.
Evaluated on the MIMIC-IV-ECG dataset, the model significantly outperforms a standard signal-only baseline in multi-label diagnosis classification.
Successfully bridges a substantial portion of the performance gap to fully multimodal models that require all data at inference, demonstrating improved accuracy with unimodal input.

Methodology

The methodology involves a novel three-stage training paradigm. First, a self-supervised, joint-embedding pre-training stage is employed to create an ECG representation enriched with contextual clinical information derived from multimodal data (laboratory exams, vitals, biometrics). Second, the pre-trained ECG encoder is fine-tuned for multi-label diagnosis classification. Third, to improve interpretability and provide explanations, the model is simultaneously trained to predict associated laboratory abnormalities directly from the ECG embedding. A crucial aspect is that only the ECG signal is required during inference, making it practical for clinical deployment.

Key Findings

The proposed model achieves superior performance over a standard signal-only baseline in multi-label diagnosis classification on the MIMIC-IV-ECG dataset. Critically, it successfully bridges a significant portion of the performance gap compared to fully multimodal models (which require all data at inference), demonstrating that rich clinical knowledge can be effectively embedded into a unimodal ECG representation for highly accurate and interpretable predictions.

Clinical Impact

This research provides a practical and effective method for developing more accurate and trustworthy AI models for ECG classification, fostering greater clinical adoption. By offering physiologically grounded explanations for predictions (e.g., associated lab abnormalities), it enhances clinician trust and facilitates the safer integration of AI into diagnostic workflows, potentially leading to earlier, more precise, and personalized patient care without needing extensive additional data at the point of care.

Limitations

The abstract does not explicitly state limitations of the proposed model itself. However, it addresses the core limitation of existing deep learning models in healthcare: their 'black box' nature and the resulting lack of trust and interpretability, which this work aims to mitigate.

Future Directions

The paper suggests that this approach offers a 'promising path toward the safer integration of AI into clinical workflows.' This implies future work will likely focus on further expanding the application of such interpretable and accurate models across various clinical domains, refining the knowledge transfer mechanisms, and exploring broader validation in diverse clinical settings.

Medical Domains

Cardiology Clinical Decision Support Diagnostic Medicine Critical Care

Keywords

Deep Learning Electrocardiogram (ECG) Clinical Knowledge Transfer Multimodal Learning Self-supervised Learning Interpretability Medical AI MIMIC-IV-ECG

Abstract

Deep learning models have shown high accuracy in classifying electrocardiograms (ECGs), but their black box nature hinders clinical adoption due to a lack of trust and interpretability. To address this, we propose a novel three-stage training paradigm that transfers knowledge from multimodal clinical data (laboratory exams, vitals, biometrics) into a powerful, yet unimodal, ECG encoder. We employ a self-supervised, joint-embedding pre-training stage to create an ECG representation that is enriched with contextual clinical information, while only requiring the ECG signal at inference time. Furthermore, as an indirect way to explain the model's output we train it to also predict associated laboratory abnormalities directly from the ECG embedding. Evaluated on the MIMIC-IV-ECG dataset, our model outperforms a standard signal-only baseline in multi-label diagnosis classification and successfully bridges a substantial portion of the performance gap to a fully multimodal model that requires all data at inference. Our work demonstrates a practical and effective method for creating more accurate and trustworthy ECG classification models. By converting abstract predictions into physiologically grounded \emph{explanations}, our approach offers a promising path toward the safer integration of AI into clinical workflows.