Federated Multimodal Knowledge Distillation for Heart Failure Risk Prediction Using Clinical Data

Abstract

Heart failure risk is a major global health challenge and early prediction is essential for improving patient outcomes. However, existing clinical prediction methods either rely on centralized data sharing, raising patient privacy concerns or use unimodal models that overlook complementary imaging information. In this paper, we propose Federated Multimodal Knowledge Distillation (FMKD), a framework that combines federated learning, multimodal fusion, and knowledge distillation for heart failure risk prediction. A multimodal teacher model integrates structured clinical features with chest X-ray images using an MLP and EfficientNet-B0 through late fusion, and transfers its knowledge to a lightweight student model. Training is performed across distributed hospital nodes using the FedAvg algorithm, where raw patient records remain local to each participating client. The current framework focuses on decentralized collaborative learning and does not incorporate additional privacy-enhancing mechanisms, such as secure aggregation or differential privacy. Experiments on synthetically paired Heart Failure Prediction and Chest X-Ray Pneumonia datasets show that FMKD achieves improved Accuracy and F1-score while maintaining competitive AUC, together with approximately 188× model compression and 4× faster inference, demonstrating an effective balance between predictive performance and computational efficiency.