CMST-FL: Cross-Missingness Stable Trust with Evidence-Guided Preprocessing for Federated Healthcare Signal Imputation

Abstract

Missing physiological samples create preprocessing uncertainty and uneven client reliability in federated learning. This study proposes Cross-Missingness Stable Trust federated learning (CMST-FL), a two-stage framework that selects an imputer using development patients and then weights local updates from their marginal validation-quality gains across MCAR, MAR, and contiguous-block masks. Twelve BIDMC records screened five imputers, and 15 disjoint records evaluated five training strategies under three mechanisms, three rates, and three seeds. Iterative imputation minimized the predeclared multi-objective score (0.35) and achieved the lowest standardized imputation RMSE (0.7712) under the tested budget. Across 27 matched evaluation conditions, CMST-FL obtained mean test R² of 0.0130, RMSE of 0.3298, and MAE of 0.2624. Compared with FedAvg, its mean R² gain was 0.00658, and its RMSE reduction was 0.00110. CMST-FL was favored in all 27 conditions for both metrics, and the paired tests remained statistically significant after Holm correction. Client-level cluster resampling, however, produced wide intervals because only five clients were available, and the worst-client gain was not significant. Trust logs showed meaningful but moderate departures from equal sample weighting. The evidence therefore supports iterative preprocessing followed by CMST-FL as a promising benchmark-specific configuration, while the low absolute R² and small error reductions preclude claims of clinical or universal superiority.