基于机器学习算法构建慢性阻塞性肺疾病吸入剂治疗患者不良吸入风险预警模型

Development of A Predictive Model for Adverse Inhalation Risk in COPD Inhaler Therapy Using Machine Learning Algorithms

目的构建并验证接受吸入剂治疗的慢性阻塞性肺疾病(COPD)患者不良吸入的风险预测模型,为不良吸入的个体化预防提供决策支持工具。方法采用横断面研究,收集COPD吸入剂治疗患者相关数据,形成数据集。按4:1将数据集随机分为训练集和测试集,通过采用4种不同的缺失值填补方法、3种变量特征筛选方法以及18种机器学习算法,在训练集上构建模型。在测试集中使用蒙特卡罗模拟法进行重采样,验证模型,以曲线下面积(AUC)、准确率、精准率、召回率和F1值评估模型性能。选择最优模型用于构建不良吸入预测平台。结果共纳入COPD患者308例,135例(43.8%)存在不良吸入风险。根据33个特征变量构建了216个风险预警模型,其中,集成学习算法的平均AUC最大,为0.844±0.058[95%CI=(0.843,0.845)]。216个模型在预测性能方面差异有统计学意义(P<0.01)。在集成学习算法下,吸入剂使用依从性(38.0874%)、吸入剂满意度(25.6801%)、教育水平(24.0313%)、吸入剂数量(5.4823%)、年龄(4.2045%)和过去一年急性加重频次(2.1847%)对预测模型贡献最大。该模型展现出良好的预测性能,其AUC 0.8693、准确率83.87%、精准率86.96%、召回率74.07%、F1值0.8。结论该研究构建的COPD吸入剂治疗患者不良吸入风险的预测模型预测能力良好,具有一定潜在临床应用价值。

Objective To construct and validate a risk prediction model for poor inhalation in chronic obstructive pulmonary disease(COPD)patients receiving inhaler therapy,providing a decision support tool for personalized prevention of poor inhalation.Methods A cross-sectional study was conducted to collect data related to COPD patients receiving inhaler therapy,forming a dataset.The dataset was randomly divided into a training set and a test set in a ratio of 4:1.Four different methods for missing value imputation,3 methods for variable feature selection,and 18 machine learning algorithms were employed to successfully construct 216 models on the training set.The monte carlo simulation method was used for resampling in the test set to validate the models,with the area under curve(AUC),accuracy,precision,recall,and F1 score used to evaluate model performance.The optimal model was selected to build the poor inhalation prediction platform.Results A study involving 308 patients with COPD found that 135(43.8%)were at risk of adverse inhalation.Using 33 predictor variables,216 risk prediction models were developed.Of these models,the ensemble learning algorithm yielded the highest average AUC of 0.844,with a standard deviation of 0.058[95%CI=(0.843,0.845)].The differences in predictive performance among the 216 models were statistically significant(P<0.01).Under the ensemble learning algorithm,adherence to inhaler use(38.0874%),inhaler satisfaction(25.6801%),literacy(24.0313%),number of inhalers(5.4823%),age(4.2045%)and number of acute exacerbations in the past year(2.1847%)contributed most to the predictive model.The model exhibited superior performance,with an AUC of 0.8693,an accuracy of 83.87%,a precision of 86.96%,a recall of 74.07%,and an F1 score of 0.8.Conclusion This study has developed a predictive model for poor inhalation risk in COPD inhaler therapy patients using machine learning algorithms,which exhibits strong predictive capabilities and holds potential clinical application value.