Related Experiment Video
Updated: Oct 15, 2026
Cell-based Assay Protocol for the Prognostic Prediction of Idiopathic Scoliosis Using Cellular Dielectric Spectroscopy
Published on: October 16, 2013
Initial screening positivity for suspected scoliosis and machine learning-assisted risk prioritization: a
Hongyan Wang1, Xiafei Liu1, Biao Wang1
1Wuhai Hospital of Traditional Chinese and Mongolian Medicine, Wuhai, Inner Mongolia Autonomous Region, China.
Background:
Questionnaire-based prescreening is inexpensive, but its ability to identify students with abnormal scoliosis screening findings is uncertain. We assessed whether questionnaire variables improved prediction beyond a demographic and anthropometric baseline and whether simple on-site physical signs could further improve risk prioritization. The outcome was initial screening positivity, defined by a positive Adams forward-bend test together with an angle of trunk rotation (ATR) of ≥5°.
Methods:
This school-based cross-sectional study included 3,559 students aged 10-18 years from four schools in Wuhai, Inner Mongolia, China. Participants underwent anthropometry, postural examination, the Adams test, electronic ATR measurement after a visually positive Adams test, and a structured questionnaire. Questionnaire models were assessed first. As questionnaire variables did not improve the demographic and anthropometric baseline, the final models used age, sex, grade, BMI, shoulder asymmetry, and foot arch morphology. Logistic regression, random forest, and XGBoost were evaluated with five-fold stratified cross-validation repeated three times.
Results:
Of 3,559 students, 392 (11.01%) met the initial screening-positive definition. Among models without physical signs, the demographic and anthropometric baseline model performed best (ROC-AUC, 0.633; PR-AUC, 0.163), while adding all or selected questionnaire variables did not improve performance. Among physical-sign models, random forest had the highest ROC-AUC (0.787) and PR-AUC (0.400), and the lowest Brier score (0.082), although its ROC-AUC gain over logistic regression was modest (0.014; 95% CI, -0.001 to 0.029). Prioritizing the top 30% by risk identified 258 of 392 positive results (65.8%); approximately 60.4% of students required prioritization to achieve 90% sensitivity.
Conclusion:
Questionnaire information did not support individual prescreening. Simple physical signs improved internal risk ranking for initial screening positivity, but the model cannot replace the combined screening procedure or diagnosis. External validation against clinical and radiographic reference standards is needed.
