Values of machine learning-based CT radiomics models in predicting recurrence of chronic subdural hematoma after endoscopic treatment
10.3760/cma.j.cn115354-20250811-00472
- VernacularTitle:基于机器学习的CT影像组学模型对慢性硬膜下血肿内镜治疗后复发的预测价值研究
- Author:
Qilong WANG
1
;
Yi WU
1
;
Zhongyong WANG
1
;
Jun DONG
1
;
Qing LAN
1
Author Information
1. 苏州大学附属第二医院神经外科，苏州 215000
- Publication Type:Journal Article
- Keywords:
Chronic subdural hematoma;
CT radiomics;
Machine learning;
Recurrence;
Prediction model;
Endoscopic treatment
- From:
Chinese Journal of Neuromedicine
2025;24(11):1115-1124
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To develop and validate CT radiomics models based on machine learning for predicting recurrence of chronic subdural hematoma (cSDH) after endoscopic treatment.Methods:A retrospective study was performed; 252 patients with cSDH who underwent endoscopic treatment in Department of Neurosurgery, the Second Affiliated Hospital of Soochow University from October 2016 to October 2024 were selected. The clinical and imaging data of these patients were collected, and these patients were divided into a training set ( n=176) and a validation set ( n=76) at a ratio of 7:3. Patients in both sets were further sub-divided into a recurrence group and a non-recurrence group based on whether they had recurrence within 3 months of discharge. (1) Radiomics features of cSDH on initial non-enhanced CT images were extracted using 3D-Slicer software. Optimal features were selected through univariate analysis and least absolute shrinkage and selection operator (LASSO) regression analysis; based on these optimal features, 3 machine learning algorithms (Logistic, support vector machine [SVM], and K-nearest neighbor [KNN]) were used to construct CT radiomics models. Differences in predictive performance of different radiomics models were compared by analyzing indicators such as sensitivity, specificity, and area under receiver operating characteristic (ROC) curve (AUC), and the best model was selected. (2) Based on the initial non-enhanced CT images, cSDH was classified into homogeneous type, laminar type, septated type, and trabecular type according to Nakaguchi classification system; combined these cSDH typing with clinical features (clinical Markwalder's grade and bilateral hematoma), univariate analysis and multivariate Logistic regression analysis were used to screen the independent risk factors for cSDH recurrence. Based on these factors, the 3 machine learning algorithms (Logistic, SVM, KNN) were used to construct hematoma typing-clinical feature models; differences in predictive performance of different hematoma typing-clinical feature models were compared by analyzing indicators such as sensitivity, specificity, and AUC, and the best model was selected. (3) DeLong's test was used to compare the ROC curve differences between the CT radiomics model and hematoma typing-clinical feature model. Decision curve analysis was used to compare the effective scope of the CT radiomics model and hematoma typing-clinical feature model. Results:(1) Seven optimal CT radiomics features based on wavelet transform were obtained after univariate analysis and LASSO regression: one gray-level dependence matrix feature, one first-order energy feature, two gray-level co-occurrence matrix features, two gray level size zone matrix features, and one gray-level run-length matrix feature. The KNN model constructed based on these 7 optimal features had the best performance in predicting cSDH recurrence, with an AUC of 0.845, a sensitivity of 0.833, a specificity of 0.857, a recall rate of 0.833, and an F1 score of 0.476 in patients from the validation set. (2) Three independent risk factors for cSDH recurrence were screened out through univariate analysis and multivariate Logistic regression analysis: hematoma Nakaguchi classification, Markwalder's grade, and bilateral hematoma. Logistic model constructed based on these 3 factors had the best performance in predicting cSDH recurrence, with an AUC of 0.675, a sensitivity of 0.609, a specificity of 0.654, a recall rate of 0.609, and an F1 score of 0.311 in patients from the validation set. (3) DeLong's test showed that the AUC of the CT radiomics model was significantly greater than that of the hematoma typing-clinical feature model in patients from the training set and validation set ( P=0.027 and P=0.035). Decision curve analysis showed that in the CT radiomics model, the net benefit of the model was >0 when the risk threshold was 0.05-0.95; in the hematoma typing-clinical feature model, the net benefit of the model was >0 when the risk threshold was 0.05-0.55. Conclusion:The KNN model based on 7 CT radiomics features in this study can effectively predict the cSDH recurrence in patients after endoscopic treatment, and its performance is obviously better than that of hematoma typing-clinical feature model constructed in this study.
