Adaptive Bayesian Partner Selection for Federated Clinical Centers

Federated learning (FL) in healthcare is often hindered by statistical heterogeneity and concept drift, where patient populations and clinical practices evolve over time. Traditional FL approaches rely on persistent, global communication, which incurs high bandwidth costs and risks negative transfer—where collaboration with poorly aligned peers degrades local model performance. This paper asks: how can clinical centers autonomously decide when, with whom, and at what cost to collaborate to maximize predictive accuracy while minimizing bandwidth?

To address these challenges, the authors introduce Adaptive Bayesian Partner Selection (ABPS). This peer-to-peer framework allows each center to maintain a Beta-Bernoulli posterior over the potential utility of other peers, measured by their Shapley marginal contribution. Centers rank candidates using an Upper Confidence Bound (UCB) criterion and employ a propose-reject mechanism to form collaborations. Crucially, the framework includes an explicit 'rest' action, allowing centers to abstain from communication when no mutually beneficial partner is available. The authors also incorporate lightweight extensions, including head personalization, bfloat16 quantization, and a goal-aware metadata filter that allows institutions to prioritize homogeneity, diversity, or KPI alignment.

Evaluated on a binary in-hospital mortality prediction task using 230 non-IID clinical centers from the MIMIC-IV dataset, the full ABPS-X variant matches the performance of the strongest federated baseline (FedDyn) with an AUROC of 0.758, while requiring only 9% of the communication cost of FedAvg. The study demonstrates that selective, utility-aware collaboration is highly effective in environments with many small, heterogeneous centers. Furthermore, the authors provide theoretical guarantees for sublinear regret in partner selection and prove that intentional isolation is the Bayes-optimal strategy when potential peers provide negative marginal utility.