机器学习辅助处方合理性预测模型在围手术期合理用药管理中的应用

Application of a machine learning-assisted prescription rationality prediction model in perioperative rational drug use management

目的探讨围手术期合理用药的影响因素,并基于机器学习建立处方合理性预测模型,以辅助药师审核处方。方法回顾性分析2021年3月至2023年3月山西省某三甲医院和某中心医院神经外科围手术期患者的处方数据。通过单因素分析和多因素Logistic回归模型分析筛选影响合理用药的因素,并结合Lasso回归和多重共线性分析筛选出重要变量,将数据按照7∶3的比例分成训练集和测试集,构建基于决策树(DT)、多层感知器(MLP)、极限梯度提升(XGBoost)、支持向量机(SVM)和随机森林(RF)5种机器学习算法的处方合理性预测模型。结果共纳入1500条处方,其中合理处方668条,不合理处方832条。在训练集和测试集中,DT、XGBoost和RF模型的受试者工作特征曲线下面积值均超过0.9,其中DT模型的敏感性最高(0.81),RF模型的特异性最高(0.90)。在RF模型中,合并症数、术前等待天数、住院总费用、开方医师职称和不良反应发生情况对处方合理性呈负向影响,同时开具药品数、年龄和给药途径则对处方合理性呈正向影响。结论基于机器学习的处方合理性预测模型具有良好的预测性能,能有效辅助药师进行处方审核,有助于降低不合理用药的发生率。

Objective To explore the influencing factors of rational perioperative drug use,and to establish a rationality prediction model based on machine learning to assist pharmacists in prescription review.Methods A retrospective analysis was conducted on the perioperative prescription data of neurosurgery patients from a tertiary hospital and a central hospital in Shanxi Province between March 2021 and March 2023.Univariate analysis and multivariate Logistic regression were initially used to identify factors influencing rational drug use,followed by Lasso regression and multicollinearity analysis to select important variables.The data was split into a training set and test set at a ratio of 7∶3,and decision tree(DT),multi-layer perceptron(MLP),extreme gradient boosting(XGBoost),support vector machine(SVM),and random forest(RF)learning models were constructed.Results A total of 1500 prescriptions were included,of which 668 were classified as rational and 832 as irrational.In both the training and test sets,the AUC values of the DT,XGBoost,and RF models exceeded 0.9.The DT model showed the highest sensitivity(0.81),while the RF model demonstrated the highest specificity(0.90).In the RF model,the number of comorbidities,preoperative waiting time,total hospitalization cost,prescribing physician's title,and adverse reaction occurrence negatively impacted prescription rationality,whereas the number of drugs,age,and administration route positively influenced rationality.Conclusion The machine learning-based rational drug use prediction model demonstrates strong predictive performance,effectively assisting pharmacists in prescription review and helping to reduce the incidence of irrational drug use.