Federated Learning for Privacy-Preserving ECG Arrhythmia Classification under Non-IID Data Distributions: A Preliminary Study

Cardiovascular diseases remain the leading cause of mortality worldwide, and early detection of arrhythmias via electrocardiogram (ECG) analysis is critical for effective intervention. Centralized machine learning approaches for ECG classification require aggregating sensitive patient data, raising significant privacy concerns. Federated Learning (FL) offers a promising alternative by enabling collaborative model training without exchanging raw data. Yet, real-world clinical data across hospitals is inherently non-independent and identically distributed (non-IID) due to demographic variation, equipment differences, and local disease prevalence. This paper presents a preliminary empirical study of FL for ECG arrhythmia classification under controlled non-IID conditions. Using the MIT-BIH Arrhythmia Database, we simulate five hospital clients partitioned with Dirichlet distributions (α ∈ {0.1, 0.5, 1.0}) and compare Federated Averaging (FedAvg) and Federated Proximal (FedProx) using a compact 1D CNN with class-weighted loss. Our results reveal two findings: (i) both algorithms collapse under severe non-IID (α=0.1), achieving macro F1 ≈ 0.22 while never predicting minority classes; and (ii) under moderate heterogeneity (α=0.5), FedProx outperforms FedAvg by 4.1 percentage points (F1 = 0.617 vs. 0.575), but reverses under milder heterogeneity (α=1.0), where FedAvg leads (F1 = 0.569 vs. 0.493). This non-monotonic behavior, consistent with recent theoretical observations on the proximal term's regularization effect, is itself a contribution. All code, data partitions, and evaluation scripts are released for full reproducibility.