Development and validation of an interpretable machine learning model integrating baseline multimodal CT perfusion and clinical data for predicting 9-month functional outcomes in acute ischemic stroke.
Zhang G, Zhang Y, et al. • Frontiers in neurology • 2026
An interpretable XGBoost-based model integrating baseline multimodal CT perfusion and clinical data achieved an AUC of 0.956 for predicting 9-month poor functional outcomes in acute ischemic stroke patients undergoing endovascular or surgical intervention, with door-to-intervention interval identified as the most influential predictor.
Key Findings
Results
The XGBoost model achieved the best predictive performance among four tested algorithms for predicting 9-month poor functional outcome in acute ischemic stroke patients.
XGBoost achieved an AUC of 0.956 (95% CI 0.917–0.995), sensitivity of 0.932, specificity of 0.897, F1-score of 0.938, and Brier score of 0.0713 (95% CI 0.0438–0.1135)
Four algorithms were compared: Logistic Regression, Random Forest, Support Vector Machine, and XGBoost
The study included 371 AIS patients who underwent endovascular or surgical intervention in a retrospective design
The outcome was defined as poor functional outcome (modified Rankin Scale score 3–6) at 9 months
Results
Door-to-intervention interval (DTI) was identified as the most influential predictor of 9-month poor functional outcome by SHAP analysis.
DTI had the highest mean SHAP value of 0.95 among all predictors
SHAP analysis was used to rank and interpret the contribution of each feature to individual predictions
DTI was followed by HU_affected (mean SHAP = 0.88), CBV < 42% (0.82), NLR (0.78), blood glucose (0.74), age (0.70), gender (0.68), ASPECT score (0.66), HU_healthy (0.50), and NWU (0.48)
The authors note that DTI highlights 'the prognostic importance of time-to-reperfusion and baseline tissue injury'
Methods
Feature selection using LASSO regression followed by SHAP-based refinement reduced the predictor set from eleven to ten variables for the final model.
Eleven predictors were initially selected by LASSO regression
After SHAP-based refinement, ten predictors remained in the final model
Predictors integrated clinical variables, laboratory markers, CT perfusion parameters (Tmax, cerebral blood flow, cerebral blood volume), and quantitative CT density measurements (Hounsfield units on affected and healthy sides, NWU, ASPECT score)
Results
CT perfusion parameter CBV < 42% was the third most influential predictor, with a mean SHAP value of 0.82.
CBV < 42% had a mean SHAP value of 0.82, ranking third among all predictors
Hounsfield units on the affected side (HU_affected) had a mean SHAP value of 0.88, ranking second
Normalized water uptake (NWU) had a mean SHAP value of 0.48, ranking tenth
These findings highlight the prognostic value of baseline tissue-level imaging parameters
Results
Calibration curves demonstrated excellent agreement between predicted and observed outcomes, and decision curve analysis confirmed superior net benefit of the XGBoost model across clinically relevant probability thresholds.
Calibration curves showed 'excellent agreement between predicted and observed outcomes'
Decision curve analysis (DCA) confirmed 'superior net benefit of the XGBoost model across clinically relevant thresholds (0.01–0.60)'
Model performance was also evaluated using AUC, sensitivity, specificity, F1-score, and Brier score
These multiple validation metrics were used to comprehensively assess model clinical utility
Methods
The study retrospectively included 371 AIS patients who underwent endovascular or surgical intervention and integrated multimodal data sources for model development.
The sample comprised 371 AIS patients in a retrospective study design
Multimodal data included clinical variables, laboratory markers (including neutrophil-to-lymphocyte ratio and blood glucose), CT perfusion parameters, and quantitative CT density measurements
The outcome of interest was poor functional outcome defined as mRS 3–6 at 9 months post-intervention
The authors note external validation has not yet been performed, stating 'After external validation, this model may serve as a practical tool'
Results
Neutrophil-to-lymphocyte ratio (NLR) and blood glucose were among the top five most influential predictors identified by SHAP analysis.
NLR had a mean SHAP value of 0.78, ranking fourth among all predictors
Blood glucose had a mean SHAP value of 0.74, ranking fifth
Age had a mean SHAP value of 0.70 (sixth) and gender had a mean SHAP value of 0.68 (seventh)
ASPECT score had a mean SHAP value of 0.66, ranking eighth
What This Means
This research developed and tested a computer-based prediction tool designed to forecast how well stroke patients will function nine months after receiving emergency treatment (either a procedure to open blocked blood vessels or surgery). The tool was built using data from 371 patients and combined information from brain CT scans, blood tests, and clinical details collected at the time of hospital arrival. Among four different types of machine learning models tested, the XGBoost model performed best, correctly identifying patients who would have poor outcomes about 93% of the time, with very few false alarms. The model's predictions also closely matched what actually happened to patients over nine months.
The research identified which factors were most important for predicting outcomes. The time from hospital arrival to the start of treatment (called door-to-intervention interval) was the single strongest predictor, reinforcing the well-established idea that faster treatment leads to better outcomes in stroke. Imaging measurements of brain tissue damage on CT scans—specifically the density of brain tissue on the affected side and the volume of brain tissue with severely reduced blood flow—were the next most important factors. Other significant predictors included markers of inflammation in the blood (neutrophil-to-lymphocyte ratio), blood sugar levels, age, sex, and a standard CT stroke scoring system called the ASPECT score.
This research suggests that combining detailed brain imaging data with basic clinical information at the time of hospital admission can provide highly accurate predictions of long-term stroke recovery. If validated in additional patient groups at other hospitals, this type of model could help doctors identify which patients are at high risk of lasting disability, enabling earlier and more targeted rehabilitation planning. The use of explainability tools (called SHAP analysis) means clinicians can see exactly why the model made a particular prediction for each patient, making it more practical and trustworthy for potential clinical use.
Zhang G, Zhang Y, Chen X, Han B, Li X, Wang Y, et al.. (2026). Development and validation of an interpretable machine learning model integrating baseline multimodal CT perfusion and clinical data for predicting 9-month functional outcomes in acute ischemic stroke.. Frontiers in neurology. https://doi.org/10.3389/fneur.2026.1877613