Abstract:

Objective This study aims to investigate the value of computed tomography perfusion-based machine learning models for predicting 90-day functional outcomes in patients with acute cerebral infarction receiving conservative treatment. Methods This study performed a retrospective analysis of 192 consecutive patients with acute cerebral infarction who received conservative treatment. Functional outcomes were assessed via the modified Rankin Scale （mRS） at the 3-month follow-up， and patients were stratified into two groups： the good outcome group （mRS score ≤ 2） and the poor outcome group （mRS score > 2）. All patients were randomly divided into a training set （n = 134） and a validation set （n = 58） at a 7∶3 ratio. Univariate and multivariate logistic regression analyses were conducted to screen for independent clinical influencing factors， which were then used to construct the clinical prediction model. Radiomic features were extracted from baseline CT perfusion images for feature selection. Based on the selected radiomic features， four machine learning models-including logistic regression， support vector machine， random forest， and light gradient boosting machine-were developed and compared to identify the optimal radiomics prediction model. A combined prediction model was then established by integrating the selected independent clinical risk factors and optimized radiomic features. The predictive performance of all constructed models was systematically evaluated and compared using receiver operating characteristic （ROC） curves， calibration curves， decision curve analysis （DCA）， and the DeLong test. Results Age， body mass index （BMI）， the National Institutes of Health Stroke Scale （NIHSS） score， perfusion parameters， and the stratified Tmax thresholds including Tmax > 4 s， Tmax > 6 s， Tmax > 8 s， and Tmax > 10 s were identified as independent prognostic factors. Among the four constructed machine learning models， the random forest model achieved the optimal predictive performance， with areas under the receiver operating characteristic curve （AUC） of 0.828 in the training set and 0.826 in the validation set. The integrated prediction model， which combined clinical features and radiomic features， further improved predictive performance， yielding AUCs of 0.888 and 0.864 in the training and validation sets， respectively. Model validation confirmed that the combined model exhibited satisfactory calibration （Hosmer-Lemeshow test， P = 0.340） and favorable clinical utility. DCA demonstrated that the model provided a substantial net clinical benefit across a wide range of threshold probabilities. The DeLong test indicated that the predictive performance of the combined model was significantly superior to that of the single-modality clinical model and single-modality radiomics model in both the training and validation sets （P < 0.05）. Conclusions The integrated machine learning model constructed by incorporating clinical risk factors and radiomic features extracted from CT perfusion imaging exhibits favorable predictive performance for 90-day clinical outcomes in patients with acute cerebral infarction receiving conservative management. This model lays a solid foundation for early clinical risk stratification and personalized clinical decision-making.

Key words: acute ischemic stroke, conservative therapy, CT perfusion imaging, machine learning, prognosis