Development and explanation of electrocardiogram-based deep learning for predicting short-term mortality in heart failure patients

Yan Li, Lixia Cheng, Hongmin Zhou, Hongwei Yu, Zeyuan Liu, Qiuju ZhangJournal of Global Health 2026 · DOI

TL;DR

This paper introduces HF-ECGNet, a novel deep learning model that uses electrocardiogram (ECG) data to predict short-term mortality in heart failure patients. By combining EfficientNet and Transformer architectures, HF-ECGNet achieves higher accuracy than traditional biomarkers like NT-proBNP and SOFA scores, offering a more accessible and cost-effective risk stratification tool.

What Problem Does This Paper Address?

Heart failure is a growing global health concern with high mortality rates, particularly in hospitalised patients. Current risk assessment tools for predicting short-term mortality in heart failure patients are often limited by accuracy, complexity, and cost. This necessitates the development of more effective, accessible, and interpretable predictive models. Electrocardiograms (ECGs) are a readily available and non-invasive diagnostic tool that can capture crucial information about cardiac health. However, the complex, high-dimensional nature of ECG data has historically posed challenges for traditional analytical methods. This paper aims to leverage deep learning to extract meaningful predictive signals from ECGs for improved short-term mortality prediction in heart failure.

Methods

The researchers developed a deep learning model named HF-ECGNet. This model is built upon a multi-channel one-dimensional convolutional EfficientNet architecture, known for its efficiency in ECG feature extraction. To capture complex temporal relationships within the sequential ECG data, a Transformer architecture was integrated. This hybrid approach allows for the simultaneous modeling of both localized abnormalities and global dynamic trends.

To enhance predictive power, a composite model was also created. This model integrates the output from HF-ECGNet (specifically, the average ECG prediction over the first three days) with nine clinical indicators, including demographics and comorbid conditions, using a multilayer perceptron.

Interpretability was addressed using Gradient-weighted Class Activation Mapping (Grad-CAM) to identify critical ECG segments contributing to predictions and Shapley Additive Explanations (SHAP) to understand feature importance in the composite model.

Datasets

The study utilized the Medical Information Mart for Intensive Care IV (MIMIC-IV) database, version 2.2, which contains comprehensive clinical data from hospital admissions. Additionally, the MIMIC-IV-ECG module, version 1.0, was used, providing ECG data specifically. The study focused on adult patients diagnosed with heart failure.

Key Insights and Ideas

  • ECG as a powerful predictor: The study demonstrates that ECG data, when analyzed with advanced deep learning techniques, can be a highly effective predictor of short-term mortality in heart failure patients, often outperforming traditional biomarkers.
  • Hybrid architecture's synergy: The combination of EfficientNet (for local feature extraction) and Transformer (for temporal dependencies) proved crucial for HF-ECGNet's ability to capture complex patterns in ECG signals.
  • Temporal dynamics matter: Predictive accuracy improves as the ECG recording time approaches the outcome event, suggesting that the model captures evolving patient conditions.
  • Composite model's advantage: Integrating clinical features with ECG predictions significantly boosts predictive performance, highlighting the value of a multi-modal approach.
  • Interpretability is achievable: Techniques like Grad-CAM and SHAP provide insights into why the model makes certain predictions, bridging the gap between complex AI and clinical understanding.

Key Results

  • HF-ECGNet achieved an Area Under the Curve (AUC) of 0.664 for the first ECG during initial admission, improving to 0.721 for the last ECG.
  • Incorporating three-day ECG data further enhanced performance, with AUCs of 0.691 (first admission) and 0.698 (last admission).
  • The composite model, integrating three-day average ECG output with clinical features, achieved the highest AUC of 0.725 for the first admission and 0.720 for the last admission in the test set.
  • HF-ECGNet outperformed traditional metrics like NT-proBNP (AUC of 0.598) and SOFA scores (AUC of 0.650) in the test set.
  • SHAP analysis identified the average risk score from the initial three days of ECGs as the most crucial feature, with age also being a significant predictor.

Limitations and Future Work

The authors acknowledge several limitations:

  • Lack of external validation: The model was developed and validated using data from a single center, limiting its generalizability to diverse populations and healthcare systems.
  • Data set scope: The dataset may not fully represent the broader spectrum of heart failure patients, potentially affecting applicability across different patient groups.
  • Single ECG analysis: Relying on a single ECG might overlook critical warning signs, although the study did explore data from multiple time points.
  • Benchmark limitations: NT-proBNP and SOFA scores were used as comparisons, but they may not be ideal benchmarks due to data availability.
  • Performance threshold: The current model performance, while promising, is still below the threshold for definitive clinical decision-making.

Future work should focus on:

  • Multi-center external validation: Crucial for assessing robustness and generalizability.
  • Integration of longitudinal ECG monitoring and multimodal clinical data: To improve predictive power and capture more comprehensive patient information.
  • Further optimization: To reach a level suitable for widespread clinical adoption.
  • Addressing low positive predictive value: Especially for rare outcomes, to minimize alert fatigue and ensure clinical utility.

Summary updated

Abstract

HF-ECGNet demonstrates potential as a powerful tool for predicting short-term mortality in heart failure patients. Its innovative architecture and integration of clinical data enable more accurate and interpretable risk stratification. Future multi-centre validation is the critical step to fully ascertain its clinical utility and generalisability.
BibTeX
@article{li2026development,
  author = {Yan Li and Lixia Cheng and Hongmin Zhou and Hongwei Yu and Zeyuan Liu and Qiuju Zhang},
  title = {{Development and explanation of electrocardiogram-based deep learning for predicting short-term mortality in heart failure patients}},
  journal = {Journal of Global Health},
  year = {2026},
  doi = {10.7189/jogh.16.04048},
}
Figure