To develop a machine learning-based model for predicting the risk of gastrointestinal bleeding in patients with spontaneous intracerebral hemorrhage

Chenzhu Cai¹Jiayin Wang²Mingfa Cai¹Zhen Qi¹Xieli Guo^1*

• ¹Jinjiang Municipal Hospital, Quanzhou, China
• ²The Second Affiliated Hospital of Fujian Medical University, quanzhou, China

Background: Spontaneous intracerebral hemorrhage (sICH) is a critical illness with a poor clinical prognosis, and gastrointestinal bleeding (GIB) is a severe complication that can significantly worsen the patient's adverse outcomes. However, research on the risk factors for GIB in sICH patients is currently limited. Therefore, this study aims to construct and validate a predictive model for GIB risk in sICH patients using machine learning methods, providing decision support for the early identification of high-risk patients in clinical settings. Methods: The present study retrospectively analysed the clinical data of 738 patients with sICH from two centres. In the feature selection process, the Boruta algorithm was initially employed for preliminary screening, and subsequently, the Information-Gain method was utilised to identify significant predictors. Following this, Spearman correlation analysis was implemented to eliminate collinearity between variables. During the model construction stage, the machine learning algorithm was optimised based on the internal test set, and the model performance was finally verified by the internal test set and the external validation set. In order to enhance the interpretability of the model, the SHapley Additive exPlanations (SHAP) method was used to visualise the prediction results. Results: The Glasgow Coma Scale (GCS) score, intraventricular extension of hemorrhage (ICH with IVH), surgeries, albumin, and distance to the midline were identified as significant predictors of GIB in patients with sICH. The patients were randomly divided into training and validation cohorts in an 8:2 ratio for model development and validation. An Extra Trees Classifier algorithm was used to construct the predictive model. Internal validation showed that the area under the receiver operating characteristic (ROC) curve (AUC) was 0.803 (95% CI: 0.659– 0.947), while the AUC for external validation data was 0.757 (95% CI: 0.675–0.839). The calibration curves for both internal and external validation were close to the ideal diagonal line, and decision curve analysis (DCA) demonstrated that the model provided a substantial net benefit. Conclusion: Our prediction model for GIB in sICH patients has reliable predictive power and provides a reliable tool for clinicians to identify early the high-risk group for GIB in sICH patients.