On-device machine learning for secure, generalizable and real-time medical data processing: on the emergence of <i>μ</i> -trainers

Abstract This review presents a comprehensive examination of advanced artificial intelligence (AI) techniques for wearable electrocardiogram (ECG) analysis, with a focus on secure, generalizable and personalized solutions enabled by tiny machine learning (TinyML). As cardiovascular diseases continue to impose a significant global health burden, accurate and real-time ECG monitoring is essential. We detail how traditional and deep learning methods, ranging from convolutional neural networks (CNNs) to bio-inspired spiking neural networks (SNNs), are adapted for resource-constrained devices. Emphasis is placed on strategies to enhance model generalizability, such as leveraging broader datasets, lead reduction, improved loss functions and generative models, alongside approaches for individual personalization through fine-tuning and meta-learning. Furthermore, the review addresses critical security and privacy challenges inherent to on-device processing and federated learning, highlighting methods like homomorphic encryption, secure multi-party computation and differential privacy. By examining what seemed like separate fields, this article highlights developments toward robust, energy-efficient and privacy-aware wearable ECG monitoring systems, supporting improved diagnostics and personalized healthcare in edge-computing environments. Recently, μ-trainers, a novel technique addressing this intertwined area, has been proposed for on-device personalization: a compact trainer fine-tunes directly on personal data within the device, updating the full model for individualized inferences without transmitting sensitive information, thereby preserving privacy.