Development of a spatial radiomics nomogram for predicting unfavorable outcomes after acute ischemic stroke: A...

To develop and externally validate a spatial radiomics nomogram integrating infarct location, lesion radiomics, and clinical predictors for individualized prediction of 90-day unfavorable outcomes in patients with acute ischemic stroke (AIS).

This multicenter retrospective study included 3,502 patients in the training cohort and 421 in the external validation cohort. Unfavorable outcome was defined as a 90-day modified Rankin Scale score of 3-6. Radiomics features were extracted using PyRadiomics, and location features were quantified by mapping infarct masks to a 425-parcellation brain atlas. Six machine learning algorithms were compared for radiomics-based and location-based models. SHapley Additive exPlanations (SHAP) analysis was used to interpret the optimal location-based model. A nomogram integrating Radscore, Locscore, and independent clinical predictors was developed.

The optimal location-based model outperformed the optimal radiomics-based model, with area under the receiver operating characteristic curve (AUC) values of 0.793 and 0.767, respectively (p < 0.01). SHAP analysis identified the superior corona radiata, superior longitudinal fasciculus, internal capsule, and precentral gyrus as the most influential regions. The final nomogram incorporated Radscore, Locscore, age, sex, admission National Institutes of Health Stroke Scale score, Trial of ORG 10172 in Acute Stroke Treatment classification, blood glucose level, and infarct volume on DWI. It achieved AUCs of 0.894 and 0.874 in the training and external validation cohorts, with good calibration and clinical utility.

Infarct location provides incremental prognostic information beyond radiomics. The proposed spatial radiomics nomogram may facilitate early individualized risk stratification in AIS.