广西医学

目的基于6种机器学习算法构建超重/肥胖患者的减重效果预测模型，并分析患者减重效果的影响因素。方法回顾性分析680例超重/肥胖患者的临床资料，将其随机分为训练集（475例）和验证集（205例）。采用基于随机森林算法的10折交叉验证递归特征消除算法筛选出38个特征变量。利用训练集数据，采用Logistic回归、K近邻、决策树、随机森林、支持向量机和朴素贝叶斯6种机器学习算法构建超重/肥胖患者减重效果的预测模型。利用验证集的数据对6种预测模型进行验证，并采用R语言caret包的varImp函数计算特征变量的重要性。结果 6种预测模型的曲线下面积（AUC）和准确率均＞0.8，其中随机森林预测模型的AUC最大，预测性能最好。减重时长、颈围、躯干脂肪含量百分比、右臂脂肪含量百分比、全身脂肪含量、左上臂围、右上臂围是较为重要的特征变量。结论基于6种机器学习算法构建的预测模型对超重/肥胖患者减重效果均具有一定的预测效能，其中随机森林预测模型的预测效能相对更优。减重时长、身体脂肪含量（躯干脂肪含量百分比、右臂脂肪含量百分比、全身脂肪含量）、颈围和双侧上臂围等指标与超重/肥胖者的减重效果密切相关，临床上可针对上述因素制订相应措施以提高减重效果。

ObjectiveTo establish the prediction model of weight loss effect in patients with overweight/obesity, and to analyze the influencing factors for weight loss effect of patients based on 6 categories of machine learning algorithms. MethodsThe clinical data of 680 overweight/obese patients were retrospectively analyzed, and patients were randomly divided into training set (475 cases) and validation set (205 cases). The 10-fold cross-validation recursive feature elimination algorithm based on random forest algorithm was used to screen 38 feature variables. Using the training set data, 6 categories of machine learning algorithms, involving the Logistic regression, K-nearest neighbor, decision tree, random forest, support vector machine, and naive Bayes, were used to establish the prediction model of weight loss effect in overweight/obese patients. The data from validation set was employed to validate the 6 categories of prediction models, and the varImp function of the R language caret package was also used to calculate the importance of feature variables.ResultsThe areas under the curve (AUC) and accuracy of 6 categories of prediction models were all larger than 0.8, therein the largest AUC went to random forest prediction model, and its predictive efficiency was the best. The duration of weight loss, neck circumference, percentage of trunk fat content, percentage of right arm fat content, whole body fat content, left upper arm circumference, and right upper arm circumference were relatively important feature variables. ConclusionThe prediction model based on 6 machine learning algorithms exerts predictive efficiency on weight loss effect of overweight/obese patients to a certain extent, of which the random forest prediction model has the relatively superior predictive efficiency. The duration of weight loss, body fat content (trunk fat content percentage, right arm fat content percentage, whole body fat content), neck circumference, bilateral upper arm circumferences, and other indices are closely related to weight loss effect of overweight/obese patients. Corresponding measures should be taken to improve weight loss effect in clinical practice based on the factors as mentioned above.