Federated TinyML for Privacy-Preserving Health Data Analytics on Edge Devices

Abstract

Healthcare data is inherently sensitive, and deploying machine learning solutions often raises concerns about patient privacy and compliance with data protection regulations. This paper proposes a novel framework combining TinyML and Federated Learning (FL) for privacy-preserving health data analytics at the edge. By training lightweight models locally on wearable and IoT devices, and aggregating insights through federated protocols, our approach minimizes the need for raw data sharing while ensuring robust performance. We implement federated TinyML pipelines for detecting cardiac irregularities and stress levels using physiological data, with an emphasis on minimizing communication overhead and energy consumption. Experimental evaluation demonstrates accurate detection of cardiac irregularities using high-quality physiological signals, such as photoplethysmography (PPG) and electrocardiogram (ECG). The pipeline achieves near-cloud-level accuracy while significantly reducing communication overhead, energy consumption, and inference latency on edge devices. Results show that the system effectively supports personalized healthcare interventions. By combining federated learning, TinyML, and healthcare-specific evaluation metrics, the proposed framework provides a low-power, scalable, and privacy-preserving solution for real-world deployment in digital health platforms and personalized medicine applications. The proposed system provides a scalable solution for privacy- preserving, low-power healthcare AI, offering strong potential for real-world deployment in personalized medicine and digital health platforms.