Revision Behavior and Explainability in Adaptive LLM Swarms for ICU Mortality Risk Prediction: A Two-Dataset Evaluation

PurposeTo evaluate how adaptive LLM swarm revision changes ICU mortality-risk outputs and to characterize evidence use, explanation indicators, auditability, and computational burden in final adaptive outputs relative to an independently executed fixed-voting (FV) architecture. MethodsWe retrospectively analyzed 1,607 eICU encounters (1,500 stays) and 1,607 ICU-2012 encounters. Initial (ASI) and final (ASF) adaptive outputs were compared within runs for revision engagement and risk-score drift; final ASF and separately generated FV outputs were compared for evidence use, explanation indicators, auditability, and computation. Paired differences and 95% confidence intervals used 10,000 hospital-stay-clustered bootstrap replicates. ResultsAt least one specialist revision trace occurred in 84.32% of eICU and 56.44% of ICU-2012 encounters. Mean ASF-minus-ASI risk-score changes were +0.0810 and +0.0539; 757 of 761 0.50-threshold crossings moved toward mortality, without clear AUROC or AUPRC improvement. In the independent benchmark, ASF explanations contained 1.26 and 0.55 more supporting-evidence items than FV, but counterevidence acknowledgement was 21.59 and 7.47 percentage points lower and unsupported-claim flags were 1.43 and 0.68 points higher. All final records met the reconstruction-completeness criterion, although ASF generated more warnings and required 1.97 and 1.52 times the FV runtime. ConclusionAdaptive revision materially changed swarm operating behaviour. Independently, final ASF outputs showed greater supporting-evidence use but less balanced evidence engagement, more process warnings, and greater computational burden than FV. These automated artifact-level findings do not establish superior explanation quality or isolate revision as their cause.