Death risk prediction model for patients with non-traumatic intracerebral hemorrhage.

Abstract

Background: This study aimed to assess the risk of death from non-traumatic intracerebral hemorrhage (ICH) using a machine learning model.

Methods: 1274 ICH patients who met the specified inclusion and exclusion criteria were analyzed retrospectively in the MIMIC IV 3.0 database. Patients were randomly divided into training, validation, and testing datasets in a ratio of 6:2:2 based on the outcome distribution. Data from the Second Hospital of Lanzhou University were used as an external validation set. This study used LASSO regression and multivariable logistic regression analysis to screen for features. We then employed XGBoost to construct a machine-learning model. The model's performance was evaluated using ROC curve analysis, calibration curve analysis, clinical decision curve analysis, sensitivity, specificity, accuracy, and F1 score. Conclusively, the SHapley Additive exPlanations (SHAP) method was employed to interpret the model's predictions.

Results: Deaths occurred in 572 out of the 1274 ICH cases included in the study, resulting in an incidence rate of 44.9%. The XGBoost model achieved a high AUC when predicting deaths in ICH patients (train: 0.814, 95%CI: 0.784 - 0.844; validation: 0.715, 95%CI: 0.653 - 0.777; test: 0.797, 95%CI: 0.743 - 0.851). The importance of SHAP variables in the model ranked from high to low was: 'GCS motor', 'Age', 'GCS eyes', 'Low density lipoprotein (LDL)', ' Albumin', ' Atrial fibrillation', and 'Gender'. The XGBoost model demonstrated good predictive performance in both the validation and external validation datasets.

Conclusions: The XGBoost machine learning model we built has demonstrated strong performance in predicting the risk of death from ICH. Furthermore, the SHAP provides the possibility of interpreting machine learning results.