IoT-Enabled Federated TabNet Framework for Privacy-Preserving CVD Risk Prediction
Mehboob Zahedi, Zinat Habiba, Bikramjit Sarkar, Shyamalendu Kandar · Procedia Computer Science · 2026
Cardiovascular disease remains the leading cause of global mortality, necessitating accurate risk prediction systems for early intervention. This study presents an IoT-enabled federated learning framework that integrates TabNet architecture for privacy-preserving cardiovascular disease risk prediction. The proposed system combines real-time physiological data from wearable sensors (ECG, heart rate, temperature) with electronic health records through a distributed training paradigm. Five simulated healthcare clients train TabNet models locally using 253,680 patient records, sharing only model parameters via the FedAvg aggregation algorithm to preserve data privacy. The TabNet architecture provides interpretable predictions through attention-based feature selection, identifying clinically relevant risk factors. Experimental results demonstrate robust convergence across 50 federated rounds, achieving 93.34% validation accuracy and 0.934 AUC with validation loss of 0.245. Comparative analysis shows significant improvements over baseline approaches: 5.6% higher accuracy than deep neural networks and 15.2% AUC enhancement over conventional methods. The framework outperforms existing federated learning approaches while maintaining model interpretability essential for clinical decision-making. This work addresses critical healthcare AI challenges by enabling collaborative learning across institutions without compromising patient privacy, offering a scalable solution for real-world cardiovascular risk assessment.