Wearable Heart Monitors Could Learn Without Sharing Data
A new framework trains arrhythmia detection across devices while keeping patient ECG records local.
Updated on Nov. 4, 2026 in Heart Disease
Researchers have proposed FedIoMT-RPM, a split federated learning framework for classifying heart arrhythmias from wearable ECG data without pooling patient records. The framework's client-side network is 28.4 KiB per device, 6.8 times smaller than in full-model federated learning, and is aimed at hospitals and devices that must learn across sites while keeping patient data local.
Why it matters
Wearable heart monitors generate continuous electrocardiogram data, but classifiers must run on resource-constrained devices and learn across institutions without sharing identifiable patient records. A design that shrinks the on-device model and reduces data exchange addresses both constraints at once, which matters for anyone whose heart rhythm is tracked outside a clinic.
This is a benchmark study on the MIT-BIH Arrhythmia Database using a patient-independent protocol, with five clients and six seeds per method. The framework reached 93.97 ± 1.72% accuracy and a 0.731 ± 0.039 Matthews correlation coefficient, though it detected fusion beats at near-zero rates and sensitivity for supraventricular beats was 64.8%.
The details
Each device trains only a small client-side network, while the bulk of the computation is shared through a split learning design rather than sending whole models or raw signals. An adaptive weighting step adjusts how much each client contributes to the shared model based on its smoothed validation performance, with bounds to keep any one client from dominating. Because only intermediate activations travel between devices and the server, the design reduces communication compared with SplitFed. The researchers also tested whether the transmitted activations leak patient information: reconstruction attacks could recover the ECG from them, while membership inference stayed near chance.
Timeline
The framework was reported in a research publication dated November 4, 2026.
Health Landscape
Federated and split learning are emerging as ways to train medical AI across hospitals without centralizing patient data, a response to strict privacy rules around health records. This work sits within that effort, targeting the specific constraints of continuous wearable ECG monitoring on low-power devices.
If you use a wearable that tracks heart rhythm, this research points toward monitoring tools that can be trained across hospitals while keeping your data on your device. Anyone with a known arrhythmia or a wearable heart monitor may want to ask their doctor how their rhythm data is used and protected.
The takeaway
Privacy-preserving training methods can make wearable heart monitoring both more capable and more protective of patient data. If you rely on a wearable for rhythm tracking, it is worth asking your physician how any monitoring data is stored and shared.
Further reading
For background on the condition at the center of this research, explore our Heart Disease coverage.
Source note: This article includes information reported by Nature.
