Search papers, labs, and topics across Lattice.
This paper presents a novel federated learning system for unsupervised ECG anomaly detection that adheres to stringent privacy regulations while ensuring real-time performance on edge devices. By integrating multiple autoencoder architectures and employing differentially private stochastic gradient descent alongside 8-bit integer quantization, the system achieves a robust area under the ROC curve (AUROC) of 0.782 across diverse hospital data. Notably, the findings reveal that privacy and model efficiency can be optimized independently, allowing for effective deployment without compromising detection quality or privacy guarantees.
Federated learning not only meets privacy standards but also outperforms centralized models in ECG anomaly detection, revealing a new paradigm for secure health monitoring on edge devices.
Continuous electrocardiography (ECG) monitoring could surface rhythm abnormalities before they escalate into cardiovascular events. However, a deployable system must satisfy three requirements simultaneously: legal-grade privacy (GDPR, HIPAA), real-time inference on constrained edge hardware, and detection quality under non-IID cross-hospital data. We design and evaluate an end-to-end federated system addressing all three for unsupervised 12-lead ECG anomaly detection on PTB-XL dataset, combining three autoencoder families (VanillaAE, ConvAE, VAE), Flower-based federated averaging (FedAvg) across ten simulated hospitals, client-side differentially private SGD (DP-SGD) with a R\'enyi-DP accountant, and 8-bit integer (INT8) post-training quantization with Raspberry Pi 4 benchmarking. Our main contributions are: an empirical characterization of how these mechanisms compose, practical DP-specific recommendations, and technical and security insights for a clinically sensitive setting. Federated learning matches or exceeds the centralized baseline across all architectures (ConvAE federated area under the ROC curve, AUROC, $0.782$), and an $\varepsilon$ sweep identifies $\varepsilon=4$ as the recommended clinical operating point. INT8 quantization roughly halves model size and cuts Pi 4 latency by up to $44%$ with $<0.12%$ AUROC loss. Crucially, DP and quantization penalties are empirically independent, so practitioners need not trade a strong privacy guarantee for a compact edge footprint. To our knowledge, this is the first system combining federated learning, formal $(\varepsilon,\delta)$-DP, unsupervised reconstruction-based detection, and quantized AArch64 deployment.
