Federated learning models designed for intensive care unit mortality prediction can preserve explainability consistency with centralized learning models while achieving comparable predictive performance, according to research published by Byounghoon Son and Jong-Moon Chung in Scientific Reports. In an analysis spanning nine architecture-matched model comparisons across three intensive care benchmark datasets, the researchers observed that feature importance rankings derived from federated models aligned closely with those from centralized counterparts, yielding Spearman’s rank correlation coefficient values that ranged between 0.6463 and 0.8374.
The investigation directly addresses critical regulatory and operational barriers in contemporary digital healthcare systems. Because modern medical privacy regulations place strict legal restrictions on aggregating patient electronic health records into central repositories, federated learning has emerged as an attractive distributed paradigm that enables training predictive machine learning models across decentralized hospital networks without sharing raw patient data. Nevertheless, translating such models into clinical deployment demands both strong predictive power and clinical trustworthiness, an objective increasingly reliant on explainable artificial intelligence frameworks.
To determine whether distributed training alters the underlying logic of clinical decision-making tools, the authors systematically analyzed what they defined as explainability consistency, representing the quantitative consistency of feature importance rankings between centralized learning and federated learning models. This assessment utilized SHapley Additive exPlanations to quantify individual feature contributions across the trained systems, followed by calculating the Spearman’s rank correlation coefficient to measure rank-order agreement between centralized and federated predictive structures.
The experimental framework examined mortality prediction across three established benchmark electronic health record datasets, identified as MIMIC-III, MIMIC-IV, and eICU. Methodologically, the evaluations encompassed tree-based ensemble models consisting of LightGBM and XGBoost utilizing cyclic boosting, as well as configurations employing TabPFN teacher models paired with distilled XGBoost student models to test whether knowledge distillation preserved feature rankings under federated parameters.
In addition to maintaining explainability consistency, the federated distilled student models demonstrated predictive efficacy on par with centralized training. Specifically, the federated student models recorded Area Under the Receiver Operating Characteristic Curve values of 0.8294 on the MIMIC-III dataset, 0.8402 on the MIMIC-IV dataset, and 0.8515 on the eICU dataset, closely tracking the performance metrics attained by centralized students. The authors concluded that their findings establish an objective quantitative framework for jointly evaluating explainability consistency and predictive performance whenever decentralized clinical mortality prediction models are compared against centralized alternatives.
Subject of Research: Explainability consistency of federated learning models in ICU mortality prediction
Article Title: Preserving explainability in federated learning for clinical decision-making: SHAP explainability consistency analysis of ICU mortality prediction models
Article References: Son, B., & Chung, J.-M. (2026). Preserving explainability in federated learning for clinical decision-making: SHAP explainability consistency analysis of ICU mortality prediction models. Scientific Reports. https://doi.org/10.1038/s41598-026-79688-4
Image Credits: AI Generated
DOI: 10.1038/s41598-026-79688-4
Keywords: federated learning, explainable artificial intelligence, SHAP, Spearman rank correlation coefficient, ICU mortality prediction, electronic health records, MIMIC-III, MIMIC-IV, eICU, LightGBM, XGBoost, cyclic boosting, TabPFN, knowledge distillation, explainability consistency, AUROC, centralized learning, clinical decision-making
Cite Scienmag News
Veronica Carney. (November 12, 2026). Federated Learning Preserves Feature Explainability in ICU Mortality Prediction Models. Scienmag. https://scienmag.com/federated-learning-preserves-feature-explainability-in-icu-mortality-prediction-models/
Veronica Carney. "Federated Learning Preserves Feature Explainability in ICU Mortality Prediction Models." Scienmag, 12 November 2026, https://scienmag.com/federated-learning-preserves-feature-explainability-in-icu-mortality-prediction-models/. Accessed 16 November 2026.
Veronica Carney. "Federated Learning Preserves Feature Explainability in ICU Mortality Prediction Models." Scienmag. November 12, 2026. https://scienmag.com/federated-learning-preserves-feature-explainability-in-icu-mortality-prediction-models/








