A LGBM model for predicting alimentary tract hemorrhage after intracerebral hemorrhage surgery: association with malnutrition risk and poor neurological recovery.

Background: Alimentary tract hemorrhage (ATH) after intracerebral hemorrhage (ICH) surgery is a common complication that can increase morbidity and mortality. Prevention of this complication is important for recovery of ICH patients, and early identification of high-risk patients would facilitate targeted prevention. Machine learning (ML) is a data-driven tool that can potentially be used to predict postoperative ATH in ICH surgical patients. However, there are currently no validated ML models for this purpose. Methods: A retrospective cohort study was performed with 658 ICH surgical patients from a single center. Five predictors were selected with the Boruta algorithm, and a total of 12 ML models were developed. The models were validated on a 70/30 train-test split, and further performance validation was performed with 10-fold cross-validation. The primary endpoint was postoperative ATH, and long-term functional outcome at 180 days was assessed with Modified Rankin Scale (MRS). Results: The Light Gradient-Boosting Machine (LGBM) model showed the best performance with an AUROC of 0.918 in the test set and an average AUROC of 0.949 on cross-validation. The five confirmed predictors were hemorrhage volume, Glasgow Coma Scale (GCS) score, surgery time, albumin, and glucose. In addition, ATH was significantly associated with lower odds of good functional outcome (MRS 0-2) at 180 days (log-rank Conclusion: The present study developed an accurate and easy-to-use ML model for early prediction of ATH in ICH surgical patients. Postoperative ATH was associated with worse long-term neurological recovery, further highlighting the importance of its prevention. The developed model should be externally validated and further used to guide the development of personalized prophylactic ATH strategies.