Predictive efficacy evaluation of MRI radiomics-based machine learning models for the prognosis of patients with traumatic cervical spinal cord injury
10.3760/cma.j.cn501098-20241129-00675
- VernacularTitle:基于MRI影像组学的机器学习模型对创伤性颈髓损伤患者预后的预测效能评估
- Author:
Yiqi XU
1
;
Han QIAO
1
;
Kai ZHANG
1
;
Jie ZHAO
1
Author Information
1. 上海交通大学医学院附属第九人民医院骨科，上海 200011
- Publication Type:Journal Article
- Keywords:
Spinal cord injuries;
Artificial intelligence;
Magnetic resonance imaging
- From:
Chinese Journal of Trauma
2025;41(4):345-352
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To evaluate the predictive efficacy of multiple MRI radiomics-based machine learning models for the prognosis of patients with traumatic cervical spinal cord injury.Methods:A retrospective cohort study was conducted to analyze the plain cervical MRI imaging data of 135 patients with traumatic cervical spinal cord injury admitted to the Ninth People′s Hospital Affiliated to Shanghai Jiao Tong University School of Medicine from January 2020 to January 2024, including 107 males and 28 females, aged 24-89 years [(56.0±11.9)years]. The patients were randomly divided into training group ( n=94) and test group ( n=41) at a ratio of 7∶3. According to the Japanese Orthopedic Association (JOA) score, 56 patients had no self-care ability (0-8 points), while 79 patients had self-care ability (9-17 points). The cervical spinal cord injury areas of all the patients were delineated. After standardizing the signal intensity distribution of the images using Z-Score, the radiomics features of the injury areas were extracted, on which the dimensionality reduction was performed by the Pearson correlation coefficient method. In the training group, the Lasso regression model was used to screen the radiomics features significantly related to the prognosis of no self-care ability and the features were then input into seven classifiers to construct prediction models, including the support vector machine, random forest, extremely randomized tree, light gradient boosting machine, adaptive boosting, naive Bayes, and K nearest neighbor. In the test group, the predictive performance of each model was evaluated through accuracy, sensitivity, specificity, precision, F1 score, area under the receiver operating characteristic (ROC) curve (AUC), and Hosmer-Lemeshow (H-L) goodness-of-fit test. The clinical applicability of each model was evaluated through clinical decision curve analysis (DCA). Results:A total of 14 radiomics features were selected, including 1 first-order feature and 13 texture features. In the test group, the accuracy rates of the 7 models of support vector machine, random forest, extremely randomized tree, light gradient boosting machine, adaptive boosting, naive Bayes, and K nearest neighbor were 0.85, 0.81, 0.90, 0.90, 0.71, 0.73 and 0.71, respectively; the precision rates were 1.00, 0.95, 0.92, 0.96, 0.72, 0.75 and 0.88, respectively; the sensitivity rates were 0.76, 0.72, 0.92, 0.88, 0.84, 0.84 and 0.60, respectively; the specificity rates were 1.00, 0.94, 0.88, 0.94, 0.50, 0.56 and 0.88, respectively; the F1 scores were 0.86, 0.82, 0.92, 0.92, 0.78, 0.79 and 0.71, respectively; the AUC values were 0.93, 0.92, 0.94, 0.97, 0.58, 0.67 and 0.88, respectively; the P values of the H-L goodness-of-fit test were 0.211, 0.112, 0.218, 0.089,<0.001,<0.001 and 0.105, respectively; the DCA results indicated that the support vector machine exhibited a clinical benefit rate greater than 0 within the 0-1 interval, surpassing the performance of the other models. Conclusions:Four MRI radiomics-based models, including support vector machine, random forest, extremely randomized tree, and light gradient boosting machine, have good predictive efficacy for the prognosis of patients with traumatic cervical spinal cord injury. Among them, the support vector machine model has the best clinical applicability.
