Machine Learning Integration of Baseline CT Imaging and Clinical Parameters Predicts 90-Day Functional Outcomes in Spontaneous Intracerebral Hemorrhage

Objective This study aims to investigate the predictive value of machine learning models integrating initial computed tomography (CT) imaging features and admission clinical parameters for 90-day functional outcomes in patients with spontaneous basal ganglia hemorrhage (BGH). The objective is to establish a clinically applicable tool for early identification of high-risk populations and guide personalized intervention strategies. Methods Patients with BGH admitted to Affiliated Binhai Hospital,Kangda College of Nanjing Medical University from January 2022 to December 2024 were retrospectively collected. Their clinical and imaging data at admission were gathered. The patients were divided into a training set and a test set in a 7:3 ratio. Univariate analysis and multivariate Logistic regression were applied to screen for risk factors. Based on the screened variables, three machine learning algorithms, namely LogisticRegression (LR), RandomForest (RF), and Support Vector Machine (SVM), were used to construct a clinical-only model, an imaging-only model, and a clinical-imaging integrated model through 5-fold cross-validation, aiming to predict the 90-day prognosis of patients with BGH. The predictive efficacy of the models was evaluated using receiver operating characteristic (ROC) curves and the area under the curve (AUC). Results Based on clinical and imaging features, we constructed 9 machine learning models. Among them, the clinical-imaging RF model showed an AUC of 0.97, a sensitivity of 0.86, and a specificity of 0.94; the LR model had an AUC of 0.89, a sensitivity of 0.85, and a specificity of 0.78; and the SVM model presented an AUC of 0.88, a sensitivity of 0.88, and a specificity of 0.74. These results indicated that the RF model had superior predictive performance. In addition, compared with the clinical RF model (AUC 0.86, sensitivity 0.85, specificity 0.73) and the imaging RF model (AUC 0.79, sensitivity 0.65, specificity 0.81), the predictive performance of the clinical-imaging RF model was significantly improved. A feature map was used to clarify the importance of variables in the optimal model. The AUC of the clinical-imaging RF model reached 0.84 in the test set, suggesting that this model has stability. Conclusion The RF model established in this study based on baseline hematoma volume at admission, minor/major axis ratio of the largest axial slice of the hematoma, perihematomal edema volume, neutrophil-to-lymphocyte ratio (NLR), and Glasgow Coma Scale (GCS) score exhibits reliable predictive performance for the 90-day prognosis of patients with BGH. It holds significant clinical application value and practical guiding significance.