PaperlayerView exact revision on arXiv

Source-linked AI summary

Adaptive Bayesian Partner Selection for Federated Clinical Centers

Navid Seidi, Satyaki Roy, Sajal K. Das

arXiv:2609.16446v1cs.LGcs.DC

TL;DR

Clinical federated learning must contend with heterogeneous centers and data distributions that drift over time, while partner choice and communication remain open challenges. ABPS uses Bayesian utility estimates and a propose–reject mechanism that permits isolation. On MIMIC-IV with 230 non-IID centers, ABPS-X matches FedDyn at 0.09× FedAvg’s bandwidth; the experiments fix the temporal window size at two years.

  • Problem

    Clinical FL faces non-IID data across institutions and concept drift, while existing approaches leave partner choice and collaboration abstention insufficiently addressed.

    Source: §1 · p. 1, §2.3 · p. 3

  • Method

    ABPS estimates candidate peers’ Shapley utility with a Bayesian posterior, selects collaborators through UCB-ranked propose–reject decisions, and permits an explicit rest action.

    Source: §1 · p. 2, §4.3 · p. 6

  • Results

    ABPS-X matches FedDyn while using 0.09× FedAvg’s bandwidth on MIMIC-IV with 230 non-IID centers.

    Source: §7 · p. 10

  • Takeaways & Limitations

    On MIMIC-IV, the diversity goal activates the rest regime for 41% of centers per round, alongside the reported accuracy-and-bandwidth result.

    Source: §7 · p. 10

    The experiments fix the medium-scale window at two years and leave online window adaptation to future work.

    Source: §3 · p. 4

Abstract

from arXiv · show

Federated learning (FL) in healthcare faces pronounced heterogeneity and temporal concept drift across clinical centers, where evolving patient populations and care practices shift data distributions. Existing approaches rely on persistent global communication, incurring substantial bandwidth overhead while risking negative transfer from poorly aligned peers. We propose Adaptive Bayesian Partner Selection (ABPS), a peer-to-peer framework that governs who collaborates, when, and at what cost. Each center maintains a Beta-Bernoulli posterior over prospective peers' Shapley marginal utility, ranks candidates with an Upper Confidence Bound (UCB) criterion, and forms collaborations through a lightweight propose-reject mechanism, with the option to abstain from communication when no mutually beneficial partner exists. The framework admits a stochastic decision interpretation, yielding finite-sample concentration guarantees and O(kappa log T) regret in partner selection, along with conditions under which intentional isolation is optimal under negative transfer. Lightweight extensions (head personalization, bfloat16 quantized communication, and a tunable active-set size) further improve efficiency, and a goal-aware metadata filter enables institution-specific collaboration strategies. On binary in-hospital mortality prediction over the first 24 hours of an ICU stay, with 230 non-IID clinical centers drawn from MIMIC-IV, the full ABPS-X variant matches the strongest federated baseline (FedDyn, AUROC 0.758) at 0.09x the communication cost of FedAvg, with reduced variability. A diversity-driven configuration activates intentional isolation for a substantial fraction of centers. These results show that adaptive, utility-aware collaboration reduces communication without sacrificing accuracy when centers are numerous and small, offering a scalable paradigm for healthcare FL.

1 Introduction

Clinical centers differ structurally and their data distributions shift over time, complicating federated learning. ABPS selects peers by estimated utility and reports matching the strongest federated baseline with substantially lower communication cost.

1 Introduction

  • Clinical FL faces both non-IID data across institutions and concept drift as patient demographics and care protocols evolve.The introduction describes these as persistent obstacles to realizing FL in clinical settings.Source: §1 · p. 1
  • Healthcare institution types have divergent performance goals and data profiles, and forcing aggregation across them can induce negative transfer that degrades local performance.Integrated delivery networks, academic medical centers, and community hospitals represent distinct profiles.Source: §1 · p. 2
  • ABPS-X matches FedDyn while using 0.09× FedAvg’s bandwidth on MIMIC-IV v3.1 with 230 non-IID centers.Applying the same extensions to FedAvg degrades performance by 6.4 AUROC points, leaving ABPS-X ahead of FedAvg-X by 6.7 points.Source: §1 · p. 2
  • ABPS estimates each candidate peer’s Shapley marginal contribution with a Beta–Bernoulli posterior, ranks peers using an ϵ-greedy UCB policy, and uses propose–reject collaboration with an explicit rest action.The paper reports posterior concentration, O(κ log T) regret, and Bayes-optimality of isolation guarantees.Source: §1 · p. 2

2 Related Work

Prior work addresses divergence-aware selection, peer-to-peer learning, drift adaptation, and communication efficiency, but leaves open how to choose partners and abstain when collaboration is unsuitable. The paper frames ABPS as addressing these connected gaps.

2 Related Work

  • The related-work discussion organizes its open questions around divergence-aware partner selection, peer-to-peer architectures, drift adaptation, and communication efficiency.These domains are addressed in the order their corresponding questions appear in the paper.Source: §2 · p. 2

2.1 Divergence Metrics for Non-IID Collaboration

  • Divergence-based collaboration methods use distributional disparity for regularization, selection, or generalization bounds, but typically impose one network-wide objective rather than center-specific collaboration goals.ABPS’s metadata pre-filter lets each center select cosine, anti-cosine, or KPI-matching admission rules without exchanging model parameters.Source: §2.1 · p. 2

2.2 P2P Federated Learning and Partner Selection

  • Peer-to-peer FL decentralizes aggregation, but selecting an appropriate topology under non-IID heterogeneity remains open; related UCB and Shapley approaches address scheduling or contribution estimation.The cited approaches include server-mediated UCB scheduling, decentralized non-stationary bandits, and server-side Shapley selection.Source: §2.2 · p. 2
  • Existing peer-to-peer, bandit-based, and Shapley-based methods generally assume mandatory participation, while heterogeneous aggregation can reduce local performance below the no-collaboration baseline.ABPS addresses selective collaboration with a Bayesian propose–reject mechanism that allows a center to abstain when posterior UCB scores fall below its acceptance threshold.Source: §2.2 · p. 3

2.3 Concept Drift and Bayesian Adaptation

  • Prior Bayesian FL work models parameter uncertainty or filters updates by confidence, but does not treat abstention as a principled decision with formal guarantees under drift.The paper positions an explicit rest action and Bayes-optimal isolation criterion as its response to this gap.Source: §2.3 · p. 3

2.4 Personalization, Quantization, and Communication-Efficient FL

  • Personalization, quantization, and server momentum can improve efficiency or stability, but prior work studies bandwidth-saving mechanisms separately from partner choice.The paper combines head personalization, bfloat16 quantization, and tunable active-set size with partner selection on an accuracy–bandwidth frontier.Source: §2.4 · p. 3

3 Problem Formulation

The problem formulation seeks accurate local predictors across centers and rounds while limiting communication under evolving data distributions. It treats the communication budget, temporal drift, and the option to rest as explicit parts of the decision problem.

3 Problem Formulation

  • Each center’s utility is its AUROC for binary in-hospital mortality classification, and each center observes samples from an unknown, time-varying distribution.The formulation considers a sequence of local predictors for evolving center-specific data.Source: §3 · p. 3
  • Centers fit local parameters and may exchange them with an active peer set; an empty set is a legal intentional-rest action.The objective includes choosing both the active peer set and the local update rule.Source: §3 · p. 3
  • The objective is to maximize mean predictive accuracy across centers and rounds while keeping cumulative communication within a budget and temporal divergence within scale-specific tolerances.When a realized shift exceeds its tolerance, the stated model allows posterior decay, UCB re-ranking, and re-formation of the active peer set.Source: §3 · pp. 3–4
  • Communication cost counts parameter exchange over active peer pairs and rounds, with uniform model sizes reducing the expression to a cost proportional to active edges.The uniform-architecture form is C_total = 2 c_egress |W| times the sum of active peer-pair counts.Source: §3 · p. 4
  • The drift model uses short-, medium-, and long-term temporal scales, while Figure 1 illustrates within-group variation and between-group divergence across careunit-by-window centers.In the MIMIC-IV instantiation, each cell is a careunit × 2-year window center, and shade encodes in-hospital mortality rate.Source: §3 · p. 4, Figure 1 · p. 4
  • The experiments fix the medium-scale window at two years and leave online window adaptation to future work.The authors state that adaptive estimation is orthogonal to their partner-selection contribution.Source: §3 · p. 4

4 Methodology: Adaptive Bayesian Partner Selection

ABPS builds adaptive peer-to-peer collaborations by scoring peers’ marginal utility, filtering candidates to fit each center’s goals, and allowing centers to rest when proposals are rejected. Bayesian belief updates and UCB-ranked proposals govern partner selection over time.

4 Methodology: Adaptive Bayesian Partner Selection

  • ABPS combines goal-aware filtering, Beta-Bernoulli utility beliefs, UCB propose-reject selection, and a rest action when no proposal is accepted.The framework also includes head personalization, bfloat16 quantization, and tunable κ as communication-efficiency extensions.Source: §4 · p. 4

4.1 Goal-Aware Metadata Pre-Filter

  • The goal-aware pre-filter admits peers by center-specific homogeneity, diversity, or alignment scores, calibrated to retain 25% of candidate pairs.It uses non-sensitive cohort metadata; the diversity goal activates intentional isolation in about 40% of rounds on the biomedical dataset.Source: §4.1 · pp. 4–5

4.2 Bayesian Belief Updating on Marginal Utility

  • Each center models a peer’s clipped Shapley marginal utility with a Beta-Bernoulli belief, whose bounded support and variance support concentration and UCB ranking.Shapley utility measures the peer’s performance gain across historical collaboration subsets; clipping uses thresholds [−0.1, +0.1].Source: §4.2 · p. 5

4.3 Collaborative Topology Formulation

  • Centers propose to peers in descending UCB order, while recipients accept only when their own UCB exceeds τacc; rejected proposals lead to rest if no peer accepts.Exploration can query a uniformly random unproven neighbor, and resting is per-round: updated beliefs and exploration can renew collaboration later.Source: §4.3 · pp. 5–6

5 Theoretical Guarantees and Complexity Analysis

The analysis gives posterior concentration and partner-selection regret guarantees under stated assumptions, establishes when resting is optimal under negative transfer, and bounds per-round computation and communication costs.

5.1 Convergence of the Beta Posterior

  • Under A1–A3, posterior error is bounded by Hoeffding noise, prior decay, and within-window drift, with convergence as K grows and drift vanishes.The assumptions include bounded utility, conditional independence of observations, and quasi-stationary drift within a window.Source: §5.1 · p. 6, §5 · p. 6

5.2 Regret of the UCB Partner Selection Bandit

  • In stationary windows, ABPS achieves O(κ log T) partner-selection regret and matches an oracle proposing to the κ best peers.The filtered-arm refinement improves constants when the goal-aware pre-filter is active, under A4.Source: §5.2 · p. 6

5.3 Optimality of Intentional Isolation

  • When every peer’s expected marginal utility is below the negative-transfer threshold, resting is optimal and saves 2|P| · c_egress · |W| bandwidth per round.For the stated clipping thresholds, cneg = 0.5, equal to τacc; posterior convergence lets ABPS approach this oracle by resting when every UCB is below τacc.Source: §5.3 · pp. 6–7

5.4 Computational Complexity

  • Per-round computation is O(nd + ρn log(ρn) + 2κ), while communication is zero for a center-round with no active peers.At default ρ = 0.25 and κ = 1, computation reduces to O(nd), independent of model parameter count |W|.Source: §5.4 · p. 7

6 Experimental Evaluation

The evaluation tests ABPS on careunit-by-year MIMIC-IV centers across predictive accuracy, communication, intentional isolation, and extension ablations. The full ABPS-X result matches FedDyn’s AUROC with substantially lower bandwidth, while a diversity goal activates isolation for some centers.

6.1 Dataset and Federation Setup

  • The primary task is binary in-hospital mortality prediction from the first 24 ICU hours, partitioned into 230 careunit-by-consecutive-two-year centers.The cohort contains roughly 76,000 stays after the age filter; the primary results use the careunit-by-year partition.Source: §6.1 · pp. 7–8

6.2 Baselines

  • The comparison covers ten methods, including centralized and local-only references, centralized federated methods, decentralized P2P methods, and Bayesian FL.DeFTA is identified as the closest peer-selection competitor; WPFed uses metadata similarity filtering without Bayesian posterior updates.Source: §6.2 · p. 8

6.3 Implementation and Metrics

  • The evaluation reports mean center AUROC and cumulative transmitted bytes, with Shapley contributions computed exactly for active sets up to five and by truncated Monte Carlo otherwise.The implementation uses five seeds; transmitted bytes are counted once per undirected edge.Source: §6.3 · p. 8

6.4 Headline Result: Accuracy vs. Bandwidth

  • The extensions move the trade-off in complementary ways: personalization raises base ABPS AUROC to 0.757 ± 0.015, while bfloat16 halves bandwidth without accuracy loss.The combined personalization, quantization, and κ=1 configuration reaches 0.748 ± 0.012 AUROC at 0.25× bandwidth.Source: §6.4 · p. 9
  • ABPS-X reaches AUROC 0.758 ± 0.010, matching FedDyn’s AUROC (0.758), at 0.09× FedAvg bandwidth.Figure 3 plots the accuracy-bandwidth comparison; Table 1 reports results for 230 centers over 50 rounds and five seeds.Source: §6.4 · pp. 8–9, Table 1 · p. 8, Figure 3 · p. 9

6.5 Empirical Validation of Lemma 1 (Intentional Isolation)

  • Diversity admits dissimilar peers, activating the rest action for 41% of centers and yielding AUROC 0.694 at 0.16× FedAvg bandwidth.Homogeneity and alignment rarely isolate centers; their expected active-set size is 0.99, with AUROCs 0.742 and 0.739 at 0.25× bandwidth.Source: §6.5 · p. 9

6.4 Headline Result: Accuracy vs. Bandwidth

  • At just 0.25× bandwidth, Bayesian Shapley-UCB selection gives an additional 2× reduction relative to FedAvg+P+Q, which reaches 0.758 ± 0.017 at 0.50× but cannot operate at κ=1.The star topology has no mechanism for selecting a single best peer.Source: §6.5 · p. 9, Figure 3 · p. 9

7 Conclusion and Future Work

The paper presents ABPS as a Bayesian P2P framework for utility-based partner selection, including an explicit isolation action and communication-efficiency extensions. On 230 MIMIC-IV centers, ABPS-X matches FedDyn at 0.09× FedAvg bandwidth, while the authors limit the matched-accuracy claim to federations of many small centers.

7 Conclusion and Future Work

  • On 230 MIMIC-IV careunit-by-year centers, ABPS-X matches FedDyn at 0.09× FedAvg bandwidth, while diversity activates the rest regime for 41% of centers per round.Applying the same extensions to FedAvg drops 6.4 AUROC points, which the authors attribute to the Bayesian selection.Source: §7 · p. 10
  • The matched-accuracy result is scoped to many small centers; on a 40-center partition, ABPS-X peaks below FedAvg and FedDyn AUROC, though it reaches intermediate targets with four to six times fewer bytes.The evaluation also covers a single dataset and binary clinical task, and provides no formal (ε, δ)-DP, survival, or Byzantine guarantees.Source: §7 · p. 10
  • The authors identify lower bandwidth at matched accuracy as a potential way to lower entry barriers for resource-constrained sites serving under-represented populations.They also note metadata re-identification and goal misuse risks, recommending a DP wrapper and governance over goal declarations.Source: §7 · p. 10

A Empirical Evidence of Concept Drift on MIMIC-IV

The appendix compares routinely charted ICU vital signs across two MIMIC-IV eras and reports statistically significant distributional shifts. It also states that Algorithm 1 operationalizes ABPS’s Bayesian-core components.

A Empirical Evidence of Concept Drift on MIMIC-IV

  • Heart-rate means were 84.9 versus 83.9 beats per minute, and respiratory-rate means were 19.3 versus 19.4 breaths per minute across 2008–2010 and 2020–2022.Welch tests report p = 5 × 10^-8 for heart rate and p = 0.038 for respiratory rate; the later respiratory-rate spread was narrower.Source: A Empirical Evidence of Concept Drift… · p. 14, Figure 4 · p. 14

B Algorithm Pseudocode

  • Algorithm 1 operationalizes metadata pre-filtering, ε-greedy UCB ranking, propose-reject topology with an explicit rest action, and Beta-Bernoulli posterior updating.The appendix identifies the algorithm as a reference in the regret proof.Source: B Algorithm Pseudocode · p. 14

C Full Proofs

The appendix states the assumptions used for the proofs of Theorems 1 and 2 and Lemma 1.

C Full Proofs

  • The proofs take bounded utility, conditional independence of per-round observations given the true utility, and quasi-stationary drift as assumptions.The passage specifies bounded utility in [0, 1].Source: C Full Proofs · p. 14

C.1 Proof of Theorem 1 (Beta-belief Concentration)

The proof decomposes the posterior-mean error into prior pull, sampling fluctuation, and drift bias. The prior contribution decays as observations accumulate, while the sampling term concentrates under bounded, conditionally independent observations.

C.1 Proof of Theorem 1 (Beta-belief Concentration)

  • Starting from prior Beta(α0, β0), after K updates the posterior is Beta(α0 + SK, β0 + K − SK), with posterior mean (α0 + SK)/(α0 + β0 + K).Source: §C.1 · p. 14
  • As K grows, the prior-pull and sampling terms vanish, yielding convergence in probability when the drift term is controlled.The combined bound includes an additive drift-bias term ϵm.Source: §C.1 · pp. 15–16
  • The deterministic prior-pull error is bounded by (α0 + β0)/(α0 + β0 + K), so it decays at Θ(1/K).This term does not depend on randomness in the observations.Source: §C.1 · p. 15
  • Hoeffding’s inequality bounds the sampling error because observations lie in [0, 1] and are conditionally independent given the true mean.The proof combines this sampling bound with the prior-pull and drift-bias terms to obtain the stated finite-sample concentration bound.Source: §C.1 · pp. 15–16

C.2 Proof of Theorem 2 (Sublinear Partner-Selection Regret)

The proof treats partner selection in a stationary window as a multi-armed bandit with κ accepted peers per round. A time-varying exploration schedule gives O(κ log T) regret, while goal-aware filtering can reduce constants subject to an optimal-peer coverage condition.

C.2 Proof of Theorem 2 (Sublinear Partner-Selection Regret)

  • In a stationary window, each center selects κ arms from its post-filter candidate set, and the proof bounds regret by analyzing greedy and exploratory phases separately.The greedy phase uses UCB1 analysis across κ acceptance slots; the exploratory phase swaps in a uniformly random peer with probability ϵ.Source: §C.2 · p. 16
  • O(κ log T) regret follows with ϵt = min(1, c/t), and average per-round regret tends to zero.The proof decomposes regret into greedy and exploratory phases; the time-varying schedule makes cumulative exploration O(log T).Source: §C.2 · p. 16
  • The goal-aware filter reduces explore-greedy regret by restricting the sums over suboptimal peers to a strict subset of candidates.The regret bound’s candidate-set size becomes approximately ρn, preserving the O(κ log T) rate while improving constants proportionally to ρ.Source: §C.2 · p. 17
  • Filtering can incur additive regret Δj⋆·T if the oracle-best peer fails the similarity threshold, but the stated coverage bound makes this term negligible under its assumptions.With an L-Lipschitz-informative descriptor, the miss probability is at most exp(−cρ|Ci|); for ρ = 0.25 and n = 230, the passage gives recovery probability above 1 − 10^-25.Source: §C.2 · p. 17

C.3 Proof of Lemma 1 (Optimality of Intentional Isolation)

The isolation proof connects one-step utility to Shapley marginal contributions and relates the oracle isolation condition to the algorithm’s noisy posterior rule. It also accounts for the bandwidth saved by resting and for bias from approximate Shapley estimates.

C.3 Proof of Lemma 1 (Optimality of Intentional Isolation)

  • Resting uses zero bandwidth, whereas collaborating with |P| peers costs 2|P| c_egress|W| per round.The stated saving is the collaboration cost for that round.Source: §C.3 · pp. 17–18
  • When every active peer has negative expected Shapley contribution, the proof identifies intentional isolation as optimal under the stated utility relationship.The passage states that held-out AUROC utility used for Shapley values and expected post-aggregation utility coincide under the protocol.Source: §C.3 · p. 17
  • With truncated Monte-Carlo Shapley estimates, strict isolation dominance survives when the true negative-gain margin exceeds η|P|.The estimator’s additive bias is bounded by η per peer.Source: §C.3 · pp. 17–18
  • As K grows and drift diminishes, the posterior-based resting rule converges to the oracle rule under the stated condition.The bound’s sampling, prior-decay, drift, and exploration terms vanish in the limit described.Source: §C.3 · p. 18
Loading 2609.16446v1…