Privacy-Preserving Federated Learning for Wearable Arrhythmia Detection: A Comprehensive Trade-off Analysis

Background Wearable devices make continuous electrocardiogram (ECG) monitoring routine, but training diagnostic models on pooled patient data sits uneasily with how sensitive that data is. Federated learning avoids pooling, yet on battery-powered wearables it forces a compromise among diagnostic accuracy, formal privacy guarantees and communication cost. This study measures that compromise for a lightweight arrhythmia classifier with 4709 parameters, trained by federated averaging across ten simulated clients on the Massachusetts Institute of Technology–Beth Israel Hospital (MIT-BIH) Arrhythmia Database. Three factors were varied: communication volume relative to centralized training, the severity of non-independent and identically distributed (non-IID) label skew across clients, and the differential privacy (DP) budget ε. Results The centralized baseline reached 98.3% accuracy. Federated training without privacy noise reached 95.0% ± 1.1% (mean ± standard deviation over five seeds) while sending 44.9 times fewer uplink bytes than a single transfer of the raw training data at 64-bit precision. Accuracy fell from 95.8% to 89.9% as the Dirichlet concentration α dropped from 1.0 to 0.1, a loss of 5.9 percentage points under extreme label skew. Differential privacy behaved very differently: mean accuracy ranged from 58.4% to 79.2% across ε values from 0.1 to 10, with standard deviations of up to 27.5 percentage points. Accuracy peaked at ε = 2 and did not rise at ε = 5 or ε = 10, and no setting came close to the 95.0% obtained without noise. Conclusions For a model of this size, communication cost and client heterogeneity are manageable, whereas the privacy mechanism is the limiting factor. The non-monotonic, high-variance privacy results point to the per-round, unclipped noise injection used here and not to a fixed ceiling on private federated arrhythmia detection. A clipped, budget-accounted mechanism should be tested before any clinical deployment.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-18
DOI
https://doi.org/10.21203/rs.3.rs-11238805/v1
Primary Topic
Privacy-Preserving Technologies in Data
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Privacy-Preserving Federated Learning for Wearable Arrhythmia Detection: A Comprehensive Trade-off Analysis

Dhruv Shah
3 links
Research Square
Privacy-Preserving Technologies in Data
preprint

Privacy-Preserving Federated Learning for Wearable Arrhythmia Detection: A Comprehensive Trade-off Analysis

Dhruv Shah
preprint en

Abstract

Background Wearable devices make continuous electrocardiogram (ECG) monitoring routine, but training diagnostic models on pooled patient data sits uneasily with how sensitive that data is. Federated learning avoids pooling, yet on battery-powered wearables it forces a compromise among diagnostic accuracy, formal privacy guarantees and communication cost. This study measures that compromise for a lightweight arrhythmia classifier with 4709 parameters, trained by federated averaging across ten simulated clients on the Massachusetts Institute of Technology–Beth Israel Hospital (MIT-BIH) Arrhythmia Database. Three factors were varied: communication volume relative to centralized training, the severity of non-independent and identically distributed (non-IID) label skew across clients, and the differential privacy (DP) budget ε. Results The centralized baseline reached 98.3% accuracy. Federated training without privacy noise reached 95.0% ± 1.1% (mean ± standard deviation over five seeds) while sending 44.9 times fewer uplink bytes than a single transfer of the raw training data at 64-bit precision. Accuracy fell from 95.8% to 89.9% as the Dirichlet concentration α dropped from 1.0 to 0.1, a loss of 5.9 percentage points under extreme label skew. Differential privacy behaved very differently: mean accuracy ranged from 58.4% to 79.2% across ε values from 0.1 to 10, with standard deviations of up to 27.5 percentage points. Accuracy peaked at ε = 2 and did not rise at ε = 5 or ε = 10, and no setting came close to the 95.0% obtained without noise. Conclusions For a model of this size, communication cost and client heterogeneity are manageable, whereas the privacy mechanism is the limiting factor. The non-monotonic, high-variance privacy results point to the per-round, unclipped noise injection used here and not to a fixed ceiling on private federated arrhythmia detection. A clipped, budget-accounted mechanism should be tested before any clinical deployment.

Research Square
Chandigarh University (IN)
Good health and well-being
Privacy-Preserving Technologies in Data
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon