基于机器学习算法构建青少年特发性脊柱侧凸进展风险预测模型

Construction of a predictive model for the progression of adolescent idiopathic scoliosis based on machine learning algorithms

中文关键词:  青少年特发性脊柱侧凸  侧凸进展  机器学习  预测模型

英文关键词:Adolescent idiopathic scoliosis  Scoliosis progress  Machine learning  Prediction model

中文摘要:

  【摘要】 目的：通过机器学习算法筛选青少年特发性脊柱侧凸（adolescent idiopathic scoliosis，AIS）进展的核心预测变量，并构建风险预测模型。方法：选取山西医科大学第二医院2018年1月～2023年6月首次确诊的361例AIS患者追踪随访。收集患者基本资料和影像学资料，根据侧凸进展（主弯Cobb角增长≥6°）与否将患者分为进展组与非进展组，并按8∶2随机分为训练集与测试集。采用LASSO回归和随机生存森林（random survival forest，RSF）筛选侧凸进展的预测变量，并分别构建RSF模型、生存支持向量机（survival support vector machine，SSVM）模型和极限梯度提升（extreme gradient boosting，XGBoost）模型，采用C-index和综合Brier分数比较选择最优模型。使用测试集数据，采用受试者工作特征（receiver operating characteristic，ROC）曲线和曲线下面积（area under the curve，AUC）、校正C-index、校准曲线、综合Brier分数和决策曲线分析（decision curve analysis，DCA）分别对最优模型的区分度、准确度和临床应用价值进行评价。结果：361例AIS患者的侧凸进展率为41.27%，LASSO回归分析和RSF算法筛选的共同预测因子分别是：初诊Cobb角、Risser征、顶椎旋转度、脊柱增长速率、是否支具治疗、T1椎体倾斜角和顶椎偏距。RSF模型、SSVM模型和XGBoost模型的C-index为0.837、0.790和0.743，综合Brier分数为0.084、0.161和0.133。综合模型的区分度和准确度，RSF模型表现更优，其6个月、12个月、18个月和24个月的时间依赖性AUC值分别为：0.903（95%CI：0.829~0.977）、0.870（95%CI：0.756~0.985）、0.858（95%CI：0.742~0.973）和0.862（95%CI：0.728~0.997），校正C-index为0.842（95%CI：0.749~0.917），模型区分度好。校准曲线显示实际观察结果与预测结果基本一致，综合Brier分数0.084，预测准确度高。DCA显示6个月、12个月、18个月和24个月的风险阈值概率分别在5%~20%、10%~80%、10%~70%和25%~85%时，使用本模型可使患者的净受益率增加。结论：基于初诊Cobb角、Risser征、顶椎旋转度、脊柱增长速率、是否支具治疗、T1椎体倾斜角和顶椎偏距构建的RSF模型可以较为准确地预测AIS患者首次确诊后在未来不同时间点侧凸进展的风险概率。

英文摘要:

  【Abstract】 Objectives: To identify the core predictor variables for adolescent idiopathic scoliosis(AIS) progression using machine learning algorithms and develop a risk prediction model. Methods: A cohort of 361 AIS patients initially diagnosed at the Second Hospital of Shanxi Medical University between January 2018 and June 2023 was enrolled and followed. Demographic data and imaging data of the patients were collected. The patients were stratified into progression(Cobb angle increased ≥6° in the primary curve) and non-progression groups, and were randomly divided into training and test sets at an 8∶2 ratio. Predictor variables were selected using LASSO regression and random survival forest(RSF) algorithms. Three prediction models were developed: a RSF model, a survival support vector machine(SSVM) model, and an extreme gradient boosting(XGBoost) model. An optimal model was selected according to the concordance index(C-index) and integrated Brier score, and the discrimination, accuracy and clinical application value of the optimal model were assessed using receiver operating characteristic(ROC) and area under the curve(AUC), bootstrap-corrected C-index, calibration curves, integrated Brier score, and decision curve analysis(DCA) based on the training set. Results: The scoliosis progression rate of the 361 AIS patients was 41.27%. The predictors consistently identified with LASSO regression and RSF algorithm were initial Cobb angle, Risser sign, apical vertebral rotation, spinal growth rate, brace treatment or not, T1 tilt angle, and apical vertebra translation. The RSF model demonstrated superior performance with a C-index of 0.837(vs. 0.790 for SSVM and 0.743 for XGBoost) and the lowest integrated Brier score(0.084 vs. 0.161 and 0.133). Time-dependent AUC values for the RSF model were 0.903(95%CI: 0.829-0.977) at 6 months, 0.870(95%CI: 0.756-0.985) at 12 months, 0.858(95%CI: 0.742-0.973) at 18 months, and 0.862(95%CI: 0.728-0.997) at 24 months. Bootstrap-corrected C-index remained robust at 0.842(95%CI: 0.749-0.917). Calibration curves showed close alignment between predicted and observed outcomes, with an integrated Brier score of 0.084. DCA revealed clinical utility across threshold probabilities of 5%-20%(6-month), 10%-80%(12-month), 10%-70%(18-month), and 25%-85%(24-month). Conclusions: The RSF model incorporating initial Cobb angle, Risser sign, apical vertebra rotation, spinal growth rate, brace treatment or not, T1 tilt angle, and apical vertebra translation can effectively predict risk probability of scoliosis progression in future for AIS patients after initial diagnosis.