Privacy-preserving adaptive arrhythmia monitoring with wearable ECG, smartphone PPG, and federated on-device learning.

INTRODUCTION: Wearable electrocardiography (ECG) and smartphone technology create new opportunities for arrhythmia monitoring outside clinical environments. However, deployable mobile cardiac systems must address more than just classification accuracy. They must reliably acquire physiological signals under real-world conditions, protect the privacy of sensitive cardiac data, adapt to user- and sensor-specific variability, and minimize false alarms in daily use. We present Cardio-FL, a smartphone-based digital health prototype for privacy-preserving, adaptive arrhythmia monitoring. METHODS: The system integrates wearable ECG acquisition with activity-aware on-device classification, camera-based photoplethysmography (PPG) pulse-regularity assessment, and federated on-device learning. ECG signals are streamed from a Movesense wearable sensor to the developed Android application via Bluetooth Low Energy, preprocessed locally, and converted into compact representations composed of derivative ECG samples and normalized RR-interval features. The deployed neural network follows an encoder-classifier-decoder architecture with only 5,575 parameters. The classifier distinguishes between normal beats, supraventricular ectopic beats (SVEB), and ventricular ectopic beats (VEB). At the same time, the encoder and decoder are further adapted on-device using unlabeled ECG segments and a reconstruction-based federated learning strategy. To improve signal representations without increasing runtime complexity, the model is pretrained prior to deployment. The encoder is first pretrained using masked reconstruction on the PTB-XL dataset and subsequently fine-tuned together with the classifier on the MIT-BIH beat-classification task before deployment. The developed Android application incorporates continuous ECG acquisition and preprocessing, on-device inference, and Flower-based federated encoder-decoder adaptation. Both inference and training are gated by user activity, while PPG-based pulse-regularity assessment is also performed after ECG alerts as a post-alert support workflow rather than as a clinically validated diagnostic confirmation step. RESULTS: In offline evaluation, masked pretraining combined with validation-selected SVEB-aware calibration improves the deployed model performance from 92.80% to 94.15% accuracy and from 69.24% to 74.53% macro-F1 compared with MIT-BIH-only training. SVEB F1 improved from 35.86% to 46.11%, while VEB F1 increased from 75.44% to 80.31%. DISCUSSION: The results demonstrate the feasibility of integrating wearable ECG acquisition, compact on-device inference, activity-aware operation, smartphone-camera PPG assessment, and reconstruction-based federated adaptation within a single Android application. Cardio-FL provides a practical framework for privacy-preserving and adaptive mobile arrhythmia monitoring. We further identify the validation steps required for real-world clinical integration.