Construction and efficacy evaluation of artificial intelligence-based automatic grading model for neurological severity at acute phase of patients with traumatic cervical spinal cord injury
10.3760/cma.j.cn501098-20241128-00674
- VernacularTitle:基于人工智能的创伤性颈髓损伤患者急性期神经功能损伤严重程度自动分级模型的构建与效能评估
- Author:
Yijin WANG
1
;
Zhenzhen GUAN
1
;
Liang WANG
1
;
Xuhua LU
1
Author Information
1. 海军军医大学第二附属医院骨科，上海 200003
- Publication Type:Journal Article
- Keywords:
Spinal cord injuries;
Artificial intelligence;
Magnetic resonance imaging;
Neurological function grading
- From:
Chinese Journal of Trauma
2025;41(5):449-455
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To construct an artificial intelligence (AI)-based automatic grading model for neurological severity at acute phase of patients with traumatic cervical spinal cord injury (TCSCI) and evaluate its efficacy.Methods:A retrospective cohort study was conducted to analyze the clinical data of 315 patients with TCSCI admitted to the Second Affiliated Hospital of Naval Medical University from January 2019 to December 2023, including 243 males and 72 females, aged 30-75 years [(57.6±7.0)years]. Injured segments involved C 1-C 4 in 143 patients and C 5-C 8 in 172. According to the American Spinal Injury Association (ASIA) scale, the injuries were classified as Grade A in 15 patients, Grade B in 53, Grade C in 74, and Grade D in 173. The patients were randomly divided into training group ( n=252) and test group ( n=63) with a ratio of 8∶2. The patients′ sensory and motor functions were assessed according to the ASIA scale within 48 hours after injury. The cervical spine MRI instance segmentation model was used to extract injury severity features of TCSCI patients in sagittal T2-weighted images. The grading model consisted of a two-layer cascade network. The first layer involved gradient boosting, Gaussian naive bayes, K-nearest neighbors, decision tree, random forest and support vector machine classifier. In the training group, the 6 machine learning models were trained separately. In the second layer, the performance of the six models was optimized to obtain the corresponding optimal grading models, so as to match the models with the best grading performance for each feature. In the test group, the performance of each model was evaluated by calculating accuracy, recall, precision, average precision, and F1 score. Results:A total of 138 clinical and imaging features were included to construct an automatic grading model for neurological severity of TCSCI patients at acute phase, comprising 132 clinical neurological features (including 56 light touch sensory scores, 56 pinprick sensory scores, and 20 key muscle scores) and 6 MRI imaging features. In the test group, the accuracy, recall, precision, average precision and F1 score of the six models, including gradient boosting, Gaussian naive bayes, K-nearest neighbors, decision tree, random forest and support vector machine classifier in the first layer of the automatic grading model for neurological severity at acute phase of TCSCI patients, in the overall grading of light touch, pinprick sensory and key muscle motor function were all above 0.86. In terms of the overall light touch function grading performance, the models with the highest accuracy, recall, precision, average precision, and F1 score were K-nearest neighbors (0.90), gradient boosting (0.99), Gaussian naive bayes (0.98), random forest (0.96), and gradient boosting (0.96), respectively. In terms of the overall pinprick sensory function grading performance, the models with the highest accuracy, recall, precision, average precision, and F1 score were gradient boosting (0.98), Gaussian naive bayes (0.98), gradient boosting (0.99), decision tree (0.99), and gradient boosting (0.95), respectively. In terms of the overall key muscle motor function grading performance, the models with the highest accuracy, recall, precision, average precision, and F1 score were K-nearest neighbors (0.97), gradient boosting and support vector machine classifier (0.97), decision tree (0.95), random forest (0.95), and support vector machine classifier (0.96), respectively. In terms of sensory function, gradient boosting had the highest number of superior performances in the overall light touch and pinprick sensory function grading. In terms of motor function, the support vector machine classifier had the highest number of superior performances in the overall key muscle motor function grading.Conclusion:The automatic grading model for neurological severity at acute phase of patients with TCSCI that is constructed based on machine learning models and two-layer cascade networks can achieve the optimization of the grading performance of each feature and exhibit a strong grading ability for the sensory and motor function severity.
