From Black Box to Discovery Engine: A Geometric and Topological Framework for Interpreting Graph Neural Networks in Critical Care

Effective clinical decision-making in critical care depends on interpreting complex, high-dimensional patient data. However, many advanced AI models function as uninterpretable "black boxes," limiting their trustworthiness and scientific value. This study introduces NETWISE, a geometric and topological framework designed not for mere prediction, but for building a validated patient similarity network to serve as a scientific discovery engine. We demonstrate the frameworks utility using acute cholangitis, a life-threatening emergency, as a case study. Analyzing 1,372 patients from the MIMIC-IV database, NETWISE constructs a heterogeneous graph to learn a robust patient manifold. The resulting model achieved high predictive accuracy (AUC-ROC of 82.5%) while offering multiple layers of interpretability. The frameworks potential for discovery was demonstrated through two complementary analyses: SHAP identified clinically coherent predictors, while discrete clustering autonomously generated 13 data-driven patient subgroups, presenting them as testable hypotheses for future clinical investigation. Crucially, topological analysis confirmed the networks immense structural complexity ({beta}1 = 63,187). This studys primary contribution is a transparent, validated, and network-based methodological framework capable of generating new, data-driven clinical hypotheses from complex EHR data.