Development and validation of a machine learning-based explainable prediction model for the outcome of patients with spontaneous intracerebral hemorrhage
10.3760/cma.j.issn.1673-4165.2025.06.004
- VernacularTitle:基于机器学习的自发性脑出血患者转归可解释预测模型的构建和验证
- Author:
Hong YUE
1
;
Zhi GENG
;
Zhaoping YU
;
Chi ZHANG
;
Xuechun LIU
;
Juncang WU
;
Aimei WU
Author Information
1. 安徽医科大学附属合肥医院，合肥市第二人民医院神经内科，合肥 230011
- Keywords:
Cerebral hemorrhage;
Treatment outcome;
Machine learning;
Tomography, X-ray computed;
Risk factors
- From:
International Journal of Cerebrovascular Diseases
2025;33(6):420-428
- CountryChina
- Language:Chinese
-
Abstract:
Objectives:To evaluate the predictive value of Tabular Prior-data Fitted Network(TabPFN) for short-term outcome in patients with spontaneous intracerebral hemorrhage (sICH), and compared with the Extreme Gradient Boosting (XGboost) model and traditional logistic regression (LR) model. Methods:Patients with sICH admitted to the Department of Neurology, Hefei Second People's Hospital from January 2018 to March 2024 were included retrospectively. The demographic and baseline data were collected. At 3 months after onset, the modified Rankin Scale score was used to determine the outcome, 0-2 was defined as good outcome and >2 was defined as poor outcome. All enrolled patients were randomly divided into a training set and a testing set at a ratio of 7:3. Feature selection was performed using recursive feature elimination (RFE) method, and then the selected feature variables were included into TabPFN, XGboost, and LR models for training and testing. The area under the curve (AUC) of receiver operating characteristic (ROC) curve was used to evaluate the predictive ability of the models. Shapley additive explanations (SHAP) method was used for model interpretation.Results:A total of 547 patients with sICH were enrolled, including 367 males (67.1%), with a median age of 65 (interquartile range, 54-76) years. Two hundred twenty-six patients (41.3%) had poor outcome. Age, baseline blood pressure (systolic blood pressure, diastolic blood pressure), baseline laboratory tests (white blood cell count, red blood cell count, platelet count, neutrophil count, hemoglobin, fasting blood glucose, creatinine, uric acid, urea nitrogen, alanine aminotransferase, aspartate aminotransferase), hematoma rupture into the ventricle, island sign, baseline hematoma volume, and baseline National Institutes of Health Stroke Scale (NIHSS) score were selected as characteristic variables using RFE method. ROC curve analysis showed that the ROC AUC for TabPFN, Xgboost, and LR models predicting poor short-term outcome in the testing set were 0.918 (95% confidence interval [ CI] 0.870-0.966], 0.883 (95% CI 0.826-0.940), and 0.905 (95% CI 0.854-0.957), respectively. SHAP analysis showed that the top four important variables in the TabPFN model were baseline NIHSS score, baseline hematoma volume, baseline aspartate aminotransferase, and age. Conclusions:The TabPFN model is superior to the LR model and the XGBoost model in predicting poor outcome in patients with sICH. In the TabPFN model, baseline NIHSS score, baseline hematoma volume, aspartate aminotransferase, and age are the most important predictors of poor outcome in patients with sICH.Objectives To evaluate the predictive value of Tabular Prior-data Fitted Network(TabPFN) for short-term outcome in patients with spontaneous intracerebral hemorrhage (sICH), and compared with the Extreme Gradient Boosting (XGboost) model and traditional logistic regression (LR) model. Methods Patients with sICH admitted to the Department of Neurology, Hefei Second People's Hospital from January 2018 to March 2024 were included retrospectively. The demographic and baseline data were collected. At 3 months after onset, the modified Rankin Scale score was used to determine the outcome, 0-2 was defined as good outcome and >2 was defined as poor outcome. All enrolled patients were randomly divided into a training set and a testing set at a ratio of 7:3. Feature selection was performed using recursive feature elimination (RFE) method, and then the selected feature variables were included into TabPFN, XGboost, and LR models for training and testing. The area under the curve (AUC) of receiver operating characteristic (ROC) curve was used to evaluate the predictive ability of the models. Shapley additive explanations (SHAP) method was used for model interpretation. Results A total of 547 patients with sICH were enrolled, including 367 males (67.1%), with a median age of 65 (interquartile range, 54-76) years. Two hundred twenty-six patients (41.3%) had poor outcome. Age, baseline blood pressure (systolic blood pressure, diastolic blood pressure), baseline laboratory tests (white blood cell count, red blood cell count, platelet count, neutrophil count, hemoglobin, fasting blood glucose, creatinine, uric acid, urea nitrogen, alanine aminotransferase, aspartate aminotransferase), hematoma rupture into the ventricle, island sign, baseline hematoma volume, and baseline National Institutes of Health Stroke Scale (NIHSS) score were selected as characteristic variables using RFE method. ROC curve analysis showed that the ROC AUC for TabPFN, Xgboost, and LR models predicting poor short-term outcome in the testing set were 0.918 (95% confidence interval [ CI] 0.870-0.966], 0.883 (95% CI 0.826-0.940), and 0.905 (95% CI 0.854-0.957), respectively. SHAP analysis showed that the top four important variables in the TabPFN model were baseline NIHSS score, baseline hematoma volume, baseline aspartate aminotransferase, and age. Conclusions The TabPFN model is superior to the LR model and the XGBoost model in predicting poor outcome in patients with sICH. In the TabPFN model, baseline NIHSS score, baseline hematoma volume, aspartate aminotransferase, and age are the most important predictors of poor outcome in patients with sICH.
