Key result
PhysioDrift-Fed improves classification under gradual drift with ~3-point higher macro F1 vs baselines.
Why the study?
Consumer cardiac wearables using federated learning face model degradation due to heterogeneous physiological drift rates among users, which existing frameworks do not address fairly or detect effectively.
Does the PhysioDrift-Fed federated learning framework improve electrocardiogram classification accuracy and reduce energy consumption under physiological concept drift in consumer cardiac wearables compared to standard baselines?
Does the PhysioDrift-Fed federated learning framework improve electrocardiogram classification accuracy and reduce energy consumption under physiological concept drift in consumer cardiac wearables compared to standard baselines?
Absolute Event Rate: 0.798% vs 0.764%
The PhysioDrift-Fed framework provides an energy-efficient federated learning solution for consumer cardiac wearables that maintains AI model accuracy as user physiology evolves.
May support drift-adaptive ECG wearables; leaves open prospective clinical validation before practice change.
Consumer cardiac wearables increasingly rely on federated learning (FL) to deliver personalized AI without compromising user privacy, yet deployed models silently degrade as user physiology evolves with age, medications, and lifestyle. We identify and formalize this challenge as the Drift-Rate Heterogeneity Problem (DRHP)- users drift at fundamentally different rates, and existing federated frameworks neither detect drift through physiological signals nor distribute its consequences fairly across the user population. This paper introduces PhysioDrift-Fed, the first dedicated FL framework for driftrate heterogeneity, combining three coordinated components: a physiology-aware drift detector that fuses statistical and electrocardiogram-domain signals, a Drift-Rate Fair Aggregation (DRFA) scheme that equalizes representation across heterogeneous drift strata, and an energy-budgeted selective update protocol respecting wearable resource limits. To support reproducible study of this problem, we release DriftBench-ECG, a public driftevaluation benchmark on the PTB-XL dataset. PhysioDrift-Fed delivers higher classification accuracy under drift, lower detection delay, substantially reduced fairness variance, and lower perround energy than five state-of-the-art baselines on real wearable-class hardware.
No takes yet. Share an insight, caveat, or question.
Chaturvedi et al. (2026) studied Cardiac anomalies (PTB-XL dataset) (n=18,885). PhysioDrift-Fed framework vs. State-of-the-art federated learning baselines (e.g., FAC-Fed, FedAvg) was evaluated on Macro F1-score under gradual concept drift. PhysioDrift-Fed improved classification accuracy under gradual drift with a macro F1-score of 0.798 compared to 0.764 for the strongest baseline, while reducing energy consumption by 60%.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: