Framework for Privacy-Preserving Wearable Health Data Analysis: Proof-of-Concept Study for Atrial Fibrillation Detection

Medical wearable devices monitor health data and, coupled with analytics, cloud computing, and artificial intelligence (AI), enable early detection of disease. Privacy issues arise when personal information is sent or processed outside the device. We propose a framework that ensures privacy integrity medical while performing AI-based homomorphically encrypted analytics in cloud. The main contributions are: (i) privacy-preserving cloud-based machine learning for devices, (ii) CipherML—a library fast implementation deployment deep learning-based solutions on data, (iii) proof-of-concept study atrial fibrillation (AF) from electrocardiograms recorded In context AF detection, two approaches are considered: multi-layer perceptron (MLP) which receives as input ECG features computed device, an end-to-end convolutional neural network (1D-CNN), raw data. CNN model achieves lower mean F1-score than hand-crafted feature-based model. This illustrates benefit over networks, especially setting small training Compared to state-of-the-art results, lead, reasonable computational overhead, slightly lower, but still similar results: performance drop caused by limitations related use instead plaintext findings highlight potential proposed enhance functionality wearables through AI, providing, within amount time, results equivalent those achieved without enhancing mechanisms. While chosen homomorphic encryption scheme prioritizes utility, certain security shortcomings remain open future development.