Research on predicting risk factors for re-bleeding in the acute phase of intracerebral hemorrhage using machine...

To investigate risk factors for rebleeding in patients during the acute phase of intracerebral hemorrhage, compare the predictive performance of various machine learning models, and develop an optimal predictive model based on SHAP interpretation.

A retrospective analysis of clinical data from 368 patients with intracerebral hemorrhage was conducted. First, the chi-square test and the Mann-Whitney U test were used to compare the rebleeding group with the non-rebleeding group; Lasso regression was then adopted to screen predictive factors. Based on the six core variables identified, the 10-fold cross-validation performance of nine models, XGBoost, logistic regression, LightGBM, random forest, AdaBoost, decision tree, GBDT, Gaussian Naive Bayes, and k-nearest neighbors, was compared. The most effective GBDT algorithm was used as the foundation, and four additional potential variables were incorporated based on clinical relevance to construct a final GBDT model comprising 10 variables. Bootstrap resampling (1,000 times) was used to calculate the confidence interval for the model's AUC. Calibration curves and decision curves were employed to evaluate the model's calibration and clinical utility, while SHAP values were utilized for global and local interpretation. SHAP dependency plots were generated for each feature, and a predictive website was ultimately developed based on this model.

There were statistically significant differences between the two groups in the history of anticoagulant or antiplatelet therapy, hematoma morphology, intraventricular hemorrhage, body temperature, white blood cell count, time from onset to CT scan, hematoma volume, Hematoma Heterogeneity Index (HII), GCS score, lactate, and partial pressure of carbon dioxide (p < 0.05). Lasso regression identified six variables: Shape, history of anticoagulant use, time from onset to CT, HII, GCS, and age. In the six-variable model, GBDT achieved the highest AUC (0.967, 95% CI 0.932-1.000) on the validation set. After incorporating DD, intraventricular hemorrhage, history of hypertension, and hematoma volume (V), the final 10-variable GBDT model achieved an AUC of 0.85 on the test set; the mean AUC from bootstrap internal validation was 0.964 (95% CI 0.939-0.982). The calibration curve demonstrated good agreement between the model's predicted probabilities and actual incidence rates, while the decision curve indicated a net benefit across a wide range of thresholds. The top three variables in the SHAP importance ranking were HII, GCS, and age. The dependency plot showed that a higher HII value, lower GCS score, more irregular hematoma shape, shorter time from onset to CT scan, and larger hematoma volume were all associated with a higher risk of rebleeding.

In this single-center retrospective study, a 10-variable GBDT model accurately predicted early rebleeding risk after acute intracerebral hemorrhage, showing stable Bootstrap performance and improved interpretability via SHAP, which can aid early detection of high-risk patients.