Explainable Federated Learning for Multi-Class Heart Disease Diagnosis via ECG Fiducial Features.

Area of Science:

• Artificial Intelligence in Medicine
• Cardiology
• Machine Learning for Healthcare

Background:

• Cardiovascular disease (CVD) is a major global health concern, necessitating accurate and timely diagnosis.
• Electrocardiograms (ECGs) are crucial for assessing cardiac function but are challenging to interpret across diverse datasets.
• Patient privacy is a significant barrier to collaborative analysis of sensitive health data.

Purpose of the Study:

• To develop an explainable federated learning (FL) framework for multi-class heart disease classification using ECGs.
• To enable collaborative analysis of ECG data across multiple institutions while preserving patient privacy.
• To enhance the interpretability of AI models in cardiac diagnosis.

Main Methods:

• Implementation of a federated learning framework integrated with long short-term memory (FL-LSTM) networks.
• Training and evaluation on three heterogeneous ECG datasets for classifying arrhythmia, ischemia, and healthy states.
• Application of SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) for model interpretability.

Main Results:

• Achieved 92% accuracy, 99% AUC, and 91% F1 score in heart disease classification.
• Demonstrated superior performance compared to existing federated approaches.
• Identified key ECG biomarkers (e.g., P-wave, R-wave, QRS, RR, QT intervals) contributing to classification.

Conclusions:

• The FL-LSTM framework offers a secure and effective method for collaborative cardiac diagnosis in distributed settings.
• Explainable AI techniques provide transparency, facilitating clinical decision-making.
• The integration of temporal modeling, federated learning, and interpretable AI advances diagnostic capabilities for cardiovascular diseases.