Prediction of short-term mortality risk in cerebellar hemorrhage: Evaluation of four machine learning models and the significant impact of SII

Objective To construct machine learning models for predicting the short-term mortality risk in patients with cerebellar hemorrhage. Methods Data were collected from 203 patients with spontaneous cerebellar hemorrhage and subjected to LASSO regression for screening risk factors. Then, logistic regression, random forest, extreme gradient boosting (XGBoost) and TabPFN models were applied for development and validation. The performance of the models was comprehensively evaluated using ROC curves, accuracy, and other related metrics, and the SHAP plot was used to explain the importance of variables within the model in order to select the optimal model. Results The short-term mortality rate in patients with spontaneous cerebellar hemorrhage was 27.7% (56/203). The TabPFN model showed the optimal prediction performance (Training AUC: 0.959, 95%CI: 0.926-0.983; Testing AUC: 0.815, 95%CI: 0.670-0.930). SHAP analysis revealed that the key variables in the model were admission GCS score, systemic immune-inflammation index (SII), the ratio of maximum hematoma cross-section to posterior fossa area, presence of brainstem compression, and intraventricular hemorrhage. Conclusions Low GCS score, high SII, high ratio of hematoma cross-section to posterior fossa area, and intraventricular hemorrhage are independent risk factors for short-term mortality in patients with cerebellar hemorrhage. The TabPFN machine learning model can be used to predict short-term mortality in patients with spontaneous cerebellar hemorrhage, with admission GCS score and SII being the most significant predictors.