Federated ICU mortality models kept SHAP rankings broadly aligned with centralized training
Federated learning is moving from a privacy-preserving idea to a practical route for building clinical prediction tools across hospitals, but performance alone does not make a model usable at the bedside. In intensive care, where decisions are made quickly and uncertainty is common, clinicians also need to know whether a model is relying on the same variables it would have used if all records had been pooled in one place. This study addressed that second question directly by measuring explainability consistency between centralized learning and federated learning for ICU mortality prediction.
The comparison centered on SHAP, the common interpretability framework used to rank feature importance in machine learning models. Rather than treating explainability as a vague qualitative property, investigators quantified how closely the SHAP feature rankings matched between centralized and federated versions of the same architecture using Spearman’s rank correlation coefficient. Across nine architecture-matched comparisons, the correlations ranged from 0.6463 to 0.8374. In practical terms, the ordering of influential variables was often fairly similar across training strategies, even though the data had been distributed across sites rather than merged into a single central dataset.
One sentence captures the main message: federated training did not break the interpretability logic of these ICU models.
Three EHR datasets, multiple architectures, one consistency question
The analysis used three benchmark electronic health record datasets: MIMIC-III, MIMIC-IV, and eICU. Those datasets are well suited to this kind of evaluation because they reflect the kind of heterogeneous, multi-institutional data environment where federated learning is most attractive. The models were built with tree-based ensemble methods, including LightGBM and XGBoost, and the experiments also included cyclic boosting and TabPFN teachers with distilled XGBoost students. That design made the study more than a simple proof of concept. It tested whether feature-importance stability could survive across different model families and training configurations.
The central clinical issue is straightforward. If a federated model predicts mortality accurately but assigns importance to a different set of variables than a centralized model, trust may erode quickly. ICU clinicians do not need a black box that merely performs well on paper; they need a system whose rationale is legible and stable enough to align with existing medical reasoning, audit processes, and safety review. In this setting, explainability consistency becomes a form of quality control rather than a decorative extra.
Predictive performance remained close to centralized students
Predictive discrimination was also strong in the federated distilled student models. The reported AUROC values were 0.8294 for MIMIC-III, 0.8402 for MIMIC-IV, and 0.8515 for eICU, each described as close to the performance of centralized students. Those numbers matter because they show that privacy-preserving training did not impose a major accuracy penalty in this setup. In ICU mortality prediction, where baseline risk can shift rapidly with organ dysfunction, vasopressor use, ventilation status, and evolving laboratory abnormalities, a model must do more than fit static data. It must preserve enough signal to remain useful across distinct clinical environments.
The consistency results add another layer to that argument. A federated model can only be considered genuinely comparable to a centralized model if it not only detects risk but also identifies it in a similar way. The SHAP rankings provide a direct window into whether the learned associations are stable across the two training regimes. Correlations in the 0.65 to 0.84 range suggest moderate to strong alignment, not perfect identity, but enough similarity to support the idea that federated learning can retain a clinically intelligible feature structure.
Why SHAP alignment matters at the bedside
Interpretability in intensive care is not an abstract software feature. It influences whether a model can be reviewed, challenged, and safely integrated into workflow. If the explanation for a high-risk mortality score changes materially when the same model is trained in a federated environment, bedside confidence weakens. If the feature hierarchy stays largely consistent, the model is easier to audit and easier to compare against established clinical expectations.
That is especially relevant in ICU medicine, where predictive systems often interact with time-sensitive decisions involving escalation of care, resource allocation, and prognostic counseling. A model that relies on different features after decentralization may still be statistically valid, but it becomes harder to know whether it is capturing the same clinical phenomenon. The present analysis suggests that, at least for these architectures and datasets, federated learning can preserve a recognizable explanation profile while protecting data locality.
The study also points to a useful methodological shift. Rather than evaluating federated learning only through AUROC, sensitivity, or calibration, explainability consistency gives investigators a second axis for comparison. That is a more demanding standard, and a more honest one, because model deployment in medicine depends on both accuracy and credibility.
What the numbers do and do not settle
The findings support a quantitative framework for comparing centralized and federated mortality models, but they do not close the case on clinical deployment. The analysis was limited to benchmark datasets and selected architectures, and the results may not automatically extend to all hospital systems, all prediction tasks, or all explanation methods. SHAP itself is useful, but it is still one interpretation layer applied to a model’s internal behavior. Different feature-attribution methods could produce different consistency profiles.
Still, the study establishes a concrete way to ask whether federated learning preserves the structure of clinical reasoning encoded by a model. In a field where privacy constraints often block direct data pooling, that question is central. If federated training can deliver near-centralized AUROC while keeping SHAP feature rankings moderately to strongly aligned, the result is more than a technical curiosity. It becomes a plausible route toward deployable ICU decision support that respects both data governance and interpretive transparency.
The remaining test is broader clinical validation across more diverse institutions, more prospective workflows, and more model families, where explanation stability will have to survive contact with real-world practice rather than benchmark comparison alone.