Modeling Day-Long ECG Signals to Predict Heart Failure Risk with Explainable AI

AI can identify high-risk patients from routine medical data: The DeepHHF model achieved an area under the receiver operating characteristic curve of 0.80 by analyzing 24-hour ECG recordings, identifying individuals with twice the risk of hospitalization or death from heart failure—demonstrating that AI can extract meaningful health insights from existing, non-invasive tests.

Longer data collection improves prediction accuracy: The model trained on full 24-hour ECG recordings significantly outperformed models using only 30-second segments and traditional clinical scores, suggesting that AI systems benefit from comprehensive data that captures irregular heart events that may occur sporadically throughout the day.

Explainability revealed what the AI actually learned: Analysis showed DeepHHF focused on medically relevant features—arrhythmias and heart abnormalities—rather than spurious patterns, providing transparency into the model's decision-making process and building confidence that it identified clinically meaningful signals rather than statistical artifacts.

Accessibility and cost-effectiveness matter for real-world impact: Because the approach uses single-lead Holter ECG (a standard, inexpensive, non-invasive test already widely available), it could potentially reach large populations for heart failure screening, rather than being limited to patients who can access expensive or specialized diagnostic tools.

Large historical datasets enable responsible validation: The study used 20 years of data from nearly 48,000 patients, providing a robust foundation for testing the model's reliability before clinical deployment—a responsible approach that reduces the risk of deploying AI systems on real patients without adequate evidence.