Development and preliminary internal validation of a prediction model for brace treatment failure in adolescent idiopathic scoliosis using pretreatment and initial brace-fitting information
10.3969/j.issn.1006-9771.2026.07.012
- VernacularTitle:基于治疗前和支具初配信息的青少年特发性脊柱侧凸支具治疗失败风险预测模型构建与初步验证
- Author:
Ying MA
1
;
Jun REN
1
;
Shoujian WANG
1
;
Xin ZHOU
1
;
Tianxiang HE
1
;
Lingjun KONG
1
;
Min FANG
1
Author Information
1. Shuguang Hospital Affiliated to Shanghai University of Traditional Chinese Medicine, Shanghai 201203, China
- Publication Type:Journal Article
- Keywords:
adolescent idiopathic scoliosis;
curve progression;
brace;
machine learning;
prediction model
- From:
Chinese Journal of Rehabilitation Theory and Practice
2026;32(7):850-860
- CountryChina
- Language:Chinese
-
Abstract:
ObjectiveTo develop and preliminarily validate a prediction model for brace treatment failure risk in patients with adolescent idiopathic scoliosis (AIS), based on information obtained before brace treatment and during initial brace fitting. MethodsA total of 191 patients with AIS who initiated thoracolumbosacral orthosis treatment at Shuguang Hospital Affiliated to Shanghai University of Traditional Chinese Medicine from March, 2015 to August, 2024 and had determinable outcomes were retrospectively included. Patients were divided according to a prespecified scheme into a training set (n = 133) and a validation set (n = 58). Brace treatment failure was defined as progression of the major-curve Cobb angle by ≥ 5° from baseline during follow-up or at brace weaning, or a final major-curve Cobb angle > 45°. Logistic regression, random forest, support vector machine with radial basis function kernel, and extreme gradient boosting (XGBoost) models were developed using routinely available clinical and radiographic variables before brace treatment and during initial brace fitting. Stratified five-fold cross-validation was used for hyperparameter tuning in the training set. Model performance was evaluated in the validation set using the area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (AUPRC), Brier score, calibration curves, and decision curve analysis (DCA). SHapley Additive exPlanations (SHAP) was used to interpret the main explanatory model. ResultsIn cross-validation of the training set, both random forest and XGBoost showed good discriminative performance, with cross-validation-AUC values of 0.871 and 0.880, respectively. In the validation set, the AUC values of random forest and XGBoost were 0.888 and 0.899, the AUPRC values were 0.678 and 0.830, and the Brier scores were 0.086 and 0.076, respectively. Random forest was selected as the primary model for interpretation, with a calibration intercept of 0.091 and a calibration slope of 0.995. SHAP analysis indicated that the in-brace correction rate and the major-curve Cobb angle measured in the brace contributed most to the model predictions. ConclusionThe random forest model developed using routinely available pretreatment and initial-fitting variables shows preliminary predictive value for brace treatment failure risk in AIS, and may provide quantitative support for follow-up scheduling, brace-fit reassessment and early adherence-oriented management.
