Work

Preserving explainability in federated learning for clinical decision-making: SHAP explainability consistency analysis of ICU mortality prediction models

Year: 2026
Type: article
Abstract: As medical privacy regulations restrict the centralized aggregation of Electronic Health Records (EHRs), Federated Learning (FL) has emerged as a promising alternative for training predictive models across distributed clinical sites. However, clinical deployment of Machine Learning (ML) models requires not only a high predictive performance but also clinical trustworthiness, which can be supported by eXplainable Artificial Intelligence (XAI). In this paper, explainability consistency, defined as the quantitative consistency of feature importance rankings between Centralized Learning (CL) and FL models, is systematically evaluated by analyzing SHapley Additive exPlanations (SHAP) feature importances using Spearman’s Rank Correlation Coefficient (SRCC). Using three benchmark EHR datasets (MIMIC-III, MIMIC-IV, and eICU), explainability consistency in mortality prediction is evaluated across tree-based ensembles (LightGBM and XGBoost) utilizing cyclic boosting and TabPFN teachers with distilled XGBoost students. Experimental results demonstrate explainability consistency across nine architecture-matched comparisons, with SRCC values ranging from 0.6463 to 0.8374. Federated distilled students achieve Area Under the Receiver Operating Characteristic Curve (AUROC) values of 0.8294 (MIMIC-III), 0.8402 (MIMIC-IV) and 0.8515 (eICU), close to centralized students. These findings provide a quantitative framework for evaluating explainability consistency alongside the predictive performance when comparing centralized and federated mortality prediction models.
Language: English
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7225890436
FWCI:
Cites: 3
Cited by:
Related to: 10