Interpretable Semi-federated Learning for Multimodal Cardiac Imaging and Risk Stratification: A Privacy-Preserving...

Area of Science:

• Cardiovascular disease prediction
• Machine learning in healthcare
• Federated learning for medical data

Background:

• Increasingly diverse cardiac patient data from hospitals and wearables requires advanced predictive modeling.
• Existing models often lack personalization, interpretability, and robust privacy safeguards.
• Need for real-time, adaptive risk stratification systems tailored to individual patient data.

Purpose of the Study:

• Introduce PerFed-Cardio, a lightweight, interpretable semi-federated learning (Semi-FL) system for real-time cardiovascular risk stratification.
• Utilize multimodal data (cardiac imaging, physiological signals, EHR) for enhanced predictive accuracy.
• Develop a privacy-conscious system adaptable to heterogeneous data sources.

Main Methods:

• Implemented a personalized Semi-FL approach with high-capacity nodes (hospitals) and edge devices (wearables).
• Employed lightweight convolutional neural networks with attention for image analysis and attention-based fusion for multimodal data integration.
• Integrated Local Interpretable Model-agnostic Explanations (LIME) and Grad-CAM for model transparency.

Main Results:

• Achieved an AUC-ROC of 0.972 with a low inference latency of 130 ms on real multimodal datasets.
• Reduced communication load by 28% through customized model calibration and targeted training.
• Maintained an F1-score exceeding 92% even in noisy data conditions.

Conclusions:

• PerFed-Cardio demonstrates a scalable, privacy-conscious, and interpretable solution for cardiac risk assessment.
• The adaptive nature of the system effectively handles heterogeneous multimodal patient data.
• Highlights the potential of personalized Semi-FL in real-time clinical decision support systems.