Warning: This is an author submission that is not peer-reviewed or edited. Preprints - unless they show as "accepted" - should not be relied on to guide clinical practice or health-related behavior and should not be reported in news media as established information.
Explainable graph neural networks for early prediction of in-hospital mortality: a case study
ABSTRACT
Background:
Early prediction of in-hospital mortality is crucial for optimizing patient management and resource allocation. Traditional clinical risk scores remain widely used but are limited by their reliance on predefined variables and restricted applicability across clinical settings. Machine learning approaches have improved predictive performance by leveraging large-scale electronic health record data, while Graph Neural Networks (GNNs) offer additional advantages by modeling the complex relationships among patients, diagnoses, medications, and procedures. Despite these advances, the limited transparency of machine learning models remains a major obstacle to their adoption in clinical practice. In particular, local explainability, which aims to identify the factors driving individual predictions, is essential for supporting high-stakes medical decisions. Although numerous explainable artificial intelligence (XAI) methods have been proposed, robust and standardized approaches for evaluating the faithfulness, robustness, and clinical relevance of explanations remain scarce, hindering reliable comparison and deployment of explainable clinical AI systems.
Objective:
This study aims at extending a previously developed graph-based model for early in-hospital mortality prediction into a fully explainable framework. Specifically, we aimed to provide patient-level explanations for model predictions and evaluate several explanations methods through a quantitative framework.
Methods:
Building on our previous work, in which a heterogeneous graph was derived from a real-world clinical data warehouse and a GNN was trained to predict in-hospital mortality using only data available during the first 2 days of hospitalization, we developed an explainability framework completing the prediction pipeline. We introduced MCTS (Monte Carlo Tree Search)-Explainer and GNNExplainer-c, which extends the GNNexplainer method, two local explanation methods designed to identify the patient-specific clinical factors and graph structures driving predictions in our use case. Explanations were quantitatively evaluated using fidelity-based metrics, compared to state-of-the-art graph explainability methods, and further submitted to a clinical expert to evaluate end-user acceptability.
Results:
MCTS-Explainer achieved the highest fidelity-based score (characterization score = 0.53, 95% CI 0.52-0.54) while allowing more concise explanatory subgraphs and a shorter computation time. Qualitative review by an expert clinician confirmed acceptability of the results due to their concise and visual nature.
Conclusions:
By extending a validated graph-based mortality prediction model with explanation generation and evaluation components, this study provides a comprehensive end-to-end framework for explainable clinical AI. MCTS-Explainer outperformed benchmark methods on quantitative evaluation metrics and was submitted to a medical expert for first qualitative assessments. These results demonstrate that graph-based models can support not only accurate prediction but also the feasibility of transparent and assessable decision support, promoting the trustworthy adoption of AI in healthcare.
Citation
Request queued. Please wait while the file is being generated. It may take some time.
Copyright
© The authors. All rights reserved. This is a privileged document currently under peer-review/community review (or an accepted/rejected manuscript). Authors have provided JMIR Publications with an exclusive license to publish this preprint on it's website for review and ahead-of-print citation purposes only. While the final peer-reviewed paper may be licensed under a cc-by license on publication, at this stage authors and publisher expressively prohibit redistribution of this draft paper other than for review purposes.