An Interpretable Multimodal CT-Clinical Model for Predicting 90-Day Functional Outcome After Acute Ischemic Stroke

Abstract

Abstract BACKGROUND A ccurate and individualized prediction of functional outcome after acute ischemic stroke (AIS) remains challenging. Conventional prognostic markers such as baseline National Institutes of Health Stroke Scale (NIHSS) score and age show substantial inter-patient variability and do not capture the spatial heterogeneity of tissue injury and hemodynamic disturbance. Although multimodal CT imaging is routinely acquired in acute stroke care, its integration into interpretable and generalizable outcome prediction frameworks remains limited. METHODS In this multicenter retrospective study of 720 AIS patients from five institutions, we developed an interpretable multimodal prognostic model integrating non-contrast CT, CT angiography, CT perfusion, and clinical variables to predict 90-day functional outcome (mRS). The model was trained in a derivation cohort and externally validated in four independent cohorts. Feature attribution methods, including SHapley Additive exPlanations (SHAP) and Gradient-weighted Class Activation Mapping (Grad-CAM), were used to identify key predictors and derive a spatially resolved risk representation by integrating NIHSS, Tmax, and age into a composite biomarker (C-SHAP), enabling region-specific prognostic interpretation. RESULTS Among 720 patients with AIS, our model achieved superior and stable performance across the derivation and four external validation cohorts, with AUCs ranging from 0.72 to 0.82, consistently outperforming clinical-only, imaging-only, and conventional multimodal models. SHAP analysis identified baseline NIHSS, Tmax, and age as the most influential predictors, with patient-specific contributions varying by clinical context. Spatial attribution using Grad-CAM localized outcome-relevant information to functionally meaningful brain regions. The derived Combined SHAP (C-SHAP) biomarker captured regionally coherent prognostic risk patterns and showed stronger associations with 90-day mRS than lesion burden alone. Integrating C-SHAP with ischemic lesion distribution enabled more individualized and informative outcome assessment. CONCLUSIONS Integration of multimodal CT imaging with clinical variables enables accurate, generalizable, and interpretable prediction of 90-day functional outcome after acute ischemic stroke. The proposed spatially resolved risk representation extends beyond lesion-based assessment and supports individualized prognostic evaluation in clinical practice.

Ähnliche Arbeiten