基于不同机器学习算法的急诊老年患者30 d内非计划再入院的风险预测模型构建和验证

Construction and validation of risk prediction models for unplanned readmissions within 30 days in elderly emergency patients based on different machine learning algorithms

目的基于不同机器学习算法构建急诊老年患者30 d内非计划再入院的风险预测模型,以帮助临床医护人员早期识别高风险患者,制订预防性干预措施。方法回顾性选取2022年5月至2023年12月中国人民解放军海军军医大学第一附属医院急诊科收治的1207例老年患者作为研究对象,按照约7∶3比例分为训练集(n=842)和测试集(n=365)。采用最小绝对收缩和选择算子(LASSO)回归分析30 d内老年患者非计划再入院的影响因素,基于机器学习算法分别构建极端梯度提升(XGBoost)、轻量梯度提升机(LightGBM)、自适应增强(AdaBoost)、Logistic回归、K最近邻(KNN)、高斯朴素贝叶斯分类(GNB)6种预测模型。对各模型进行综和评估与验证,并用Shapley加法解释分析关键变量的重要性。结果1207例老年患者中,训练集842例,男430例,中位年龄77岁,测试集365例,男176例,中位年龄78岁;采用LASSO回归筛选出8个变量特征,以此构建的XGBoost、LightGBM、AdaBoost、Logistic回归、KNN、GNB 6个预测模型中,GNB模型表现最优,其中测试集的AUC为0.818,灵敏度为0.890,特异度为0.660,且训练集和验证集具有较强的拟合能力和较高的稳定性;影响急诊老年患者30 d内非计划再入院8个特征重要性排序分别为年龄、慢性阻塞性肺疾病、住院时间、察尔森合并症指数≥3、低蛋白血症、异常生命体征≥2个、卒中、贫血。结论基于机器学习算法构建的急诊老年患者30 d内非计划再入院的GNB模型具有良好的预测效果,有助于医护人员在出院前尽早识别高危患者,制订针对性的预防性措施,从而降低患者短期内非计划再入院率,提高患者生命质量。

Objective To construct the risk prediction model of unplanned readmission for elderly patients in emergency department within 30 days based on different machine learning algorithms,so as to help clinical staff identify high-risk patients early and formulate preventive interventions.Methods A total of 1207 elderly patients admitted to the emergency department of the First Affiliated Hospital of Naval Medical University from May 2022 to December 2023 were retrospectively selected as the study objects and were divided into the training set(n=842)and the test set(n=365)in a ratio of approximately 7∶3.Least absolute shrinkage and selection operator(LASSO)regression analysis was used to screen the factors affecting the unplanned readmission of elderly patients within 30 days.Six prediction models,including extreme Gradient Boost(XGBoost),Light Gradient Boosting Machine(LightGBM),Adaptive Boosting(AdaBoost)Logistic regression,K-nearest neighbor(KNN)and Gauss Naive Bayes classification(GNB)were constructed respectively.The models were summarized,evaluated and validated,and the importance of key variables was analyzed using Shapley Additive Interpretation(SHAP).Results Among 1207 elderly patients,there were 842 in the training set,430 males with a median age of 77 years,and 365 in the test set,176 males with a median age of 78 years.Eight variable features were selected by LASSO regression.The GNB model performed the best among the 6 prediction models constructed based on XGBoost,LightGBM,AdaBoost,Logistic regression,KNN,GNB.The AUC of the test set was 0.818,and the sensitivity was 0.890,while the specificity was 0.660,and the train set and the verification set had strong fitting ability and high stability.The eight characteristics affecting the unplanned readmitted of elderly patients in the emergency department within 30 days were ranked in importance by age,chronic obstructive pulmonary disease,length of stay,Charson comorbidity index≥3,hypoproteinemia,abnormal vital signs≥2,stroke,anemia.Conclusions The GNB model based on machine learning algorithms for unplanned readmission of elderly emergency patients within 30 days has good predictive performance,which helps medical staff to identify high-risk patients as early as possible before discharge,formulate targeted preventive measures,thereby reducing the short-term unplanned readmission rate of patients and improving their quality of life.