军事知识服务平台
基于人体成分及代谢指标构建2型糖尿病患者并发骨质疏松症的机器学习预测模型
Development of a machine learning prediction model for osteoporosis in type 2 diabetes patients based on body composition and metabolic indicators
摘要 目的:通过机器学习方法,基于人体成分数据与代谢指标构建和确定适用于2型糖尿病(T2DM)患者并发骨质疏松症(OP)的最优预测模型。方法:该研究为横断面研究,连续纳入2021年10月至2026年3月在苏州大学附属第一医院内分泌科住院的T2DM患者为研究对象。收集研究对象年龄、性别、尿酸、游离三碘甲状腺原氨酸(FT 3)水平、总Ⅰ型胶原氨基端延长肽(P1NP)、肌肉量、脂肪量、基础代谢率等临床数据。依据骨密度将研究对象分为T2DM合并OP组(DOP组,T值≤-2.5)和T2DM不合并OP组(NOP组,T值>-2.5)。将数据按8∶2划分为训练集与验证集,构建并比较9种机器学习模型[多因素logistic回归模型、支持向量机、K最近邻、高斯朴素贝叶斯、多层感知机、随机森林、决策树、极限梯度提升算法、轻量级梯度提升机(LGBM)]的性能。采用受试者工作特征(ROC)曲线评估模型区分度,计算曲线下面积(AUC)、灵敏度、特异度等;采用决策曲线分析(DCA)评估模型临床有效性;采用沙普利加性解释(SHAP)分析确定最优模型中纳入建模过程变量的重要性。 结果:共纳入1 213例T2DM患者。其中,DOP组412例,NOP组801例。与NOP组相比,DOP组患者年龄更大,肌肉量及尿酸水平均更低,而女性比例、P1NP及FT 3水平均更高(均 P<0.05)。9种机器学习模型中,LGBM模型整体性能最佳,在训练集中,其AUC为0.883(95%CI 0.861~0.904),灵敏度为0.829,特异度为0.798;在验证集中,其AUC为0.860(95%CI 0.812~0.907),灵敏度为0.753,特异度为0.817,且临床决策净效益良好。SHAP分析显示基础代谢率、FT 3、年龄、P1NP、脂肪量为前5位重要预测因子。 结论:不同机器学习算法构建的预测模型中,基于LGBM算法构建的模型预测T2DM患者并发OP的效能最高。
Abstract:
Objective:To construct and determine the optimal prediction model for osteoporosis (OP) in patients with type 2 diabetes mellitus (T2DM) by using machine learning methods based on body composition data and metabolic indicators.Methods:This was a cross-sectional study that consecutively enrolled T2DM patients hospitalized in the Department of Endocrinology at the First Affiliated Hospital of Soochow University from October 2021 to March 2026. Clinical data on age, sex, uric acid, free triiodothyronine (FT 3), and total procollagen type Ⅰ N-terminal propeptide (P1NP), muscle mass, fat mass, and basal metabolic rate were collected. Based on bone mineral density, the patients were divided into a T2DM with OP group (DOP group, T-score≤-2.5) and a T2DM without OP group (NOP group, T-score>-2.5). The data were randomly split into a training set and an internal validation set at an 8∶2 ratio. Nine machine learning models [multivariate logistic regression, support vector machine, K-nearest neighbors, Gaussian naive Bayes, multilayer perceptron, random forest, decision tree, extreme gradient boosting, and light gradient boosting machine (LGBM) ] were constructed and compared. The receiver operating characteristic (ROC) curve was used to evaluate model discrimination, and the area under the curve (AUC), sensitivity, and specificity were calculated. Decision curve analysis (DCA) was employed to assess clinical usefulness. SHapley Additive exPlanations (SHAP) analysis was used to determine the importance of the variables included in the optimal model. Results:A total of 1 213 T2DM patients were enrolled, including 412 in the DOP group and 801 in the NOP group. Compared with the NOP group, the DOP group had a higher proportion of women, older age, lower muscle mass and uric acid levels, and higher levels of P1NP and FT 3 ( P<0.05). Among the nine machine learning models, the LGBM model had the best overall performance. In the training set, the AUC was 0.883 (95%CI 0.861-0.904), the sensitivity was 0.829, and the specificity was 0.798. In the validation set, the AUC was 0.860 (95%CI 0.812-0.907), the sensitivity was 0.753, and the specificity was 0.817, with good clinical net benefit on decision curve analysis. SHAP analysis showed that basal metabolic rate, FT 3, age, P1NP, and fat mass were the top five most important predictive factors. Conclusion:Among the prediction models constructed using different machine learning algorithms, the model based on the LGBM algorithm had the highest performance in predicting OP in T2DM patients.