FedCareX: Trust-Aware Federated Transformer for Wearable IoT Seizure Prediction

Abstract

Behind-the-ear electroencephalography (EEG) can now be recorded continuously outside hospital, but forecasting the preictal state is hard in a distributed deployment: raw traces cannot leave the clinical site, the wearable montage is chosen per patient, and part of the labelling comes from automated annotators rather than clinicians. Federated learning removes the need to move recordings, yet weighting each site by its sample count lets a poorly annotated site dominate the shared model. FedCareX answers this with a montage-agnostic tokenizer, a lightweight preictal Transformer encoder, an adaptive reliability aggregation rule based on annotation quality, gradient consistency and probe-set agreement, and a sparsified error-feedback codec that reduces uplink traffic. On a five-centre wearable corpus FedCareX attains 83.1±0.6% sensitivity with 0.38 false predictions per hour, a 2.8 point increase in sensitivity and a 13.6% relative reduction in false prediction rate compared with the best 2026 baseline, and on a scalp benchmark with 23 patient clients it reaches 92.8±0.4% sensitivity with 0.22 false predictions per hour. The scalp margin is significant under a paired Wilcoxon signed-rank test over 23 independent client pairs; the wearable margin rests on five centres and is reported as a hierarchical bootstrap interval of [1.1,4.4] sensitivity points rather than as a pooled significance test. Uplink volume drops by 22.1× to 0.22 MB per client per round, and edge inference spends 28.9 mJ per window against the measured 55 mJ per window budget on the gateway platform.