Interpretable multimodal machine learning model for predicting health risks of patients with heart failure

Area of Science:

• Cardiology
• Medical Informatics
• Machine Learning

Background:

• Heart failure (HF) poses a significant global health burden, necessitating improved risk stratification tools.
• Accurate prediction of mortality and readmission is crucial for effective HF patient management.

Purpose of the Study:

• To develop and evaluate an interpretable multimodal machine learning framework for predicting 30-day mortality and hospital readmission in HF patients.
• To assess the contribution of different clinical data modalities (demographics, medications, lab tests, ECGs) to HF risk prediction.

Main Methods:

• Utilized clinical data from 2868 HF patients across 43 hospitals in Hong Kong.
• Trained and evaluated ten machine learning models, integrating four data modalities.
• Employed SHapley Additive exPlanations (SHAP) for model interpretability and feature importance analysis.

Main Results:

• The best model achieved an AUC of 0.881 for mortality and 0.709 for readmission.
• Laboratory tests and ECG features demonstrated the highest predictive power, with their combination yielding near-optimal results (AUC: 0.872).
• Key predictors included serum albumin, high-sensitivity troponin I, lactate dehydrogenase, and QT interval dispersion.

Conclusions:

• Interpretable multimodal machine learning enhances HF risk prediction.
• Laboratory tests and ECGs alone may suffice for risk prediction in resource-constrained environments.
• Findings support personalized HF management and scalable deployment of predictive models.