基于机器学习的住院老年慢性心力衰竭患者认知衰弱风险预测模型构建与验证 [中文引用][英文引用]

目的 探讨住院老年慢性心力衰竭（chronic heart failure，CHF）患者认知衰弱的影响因素，通过不同机器学习算法构建 8 种预测模型，筛选出最佳模型，为临床干预提供依据。方法 采用便利抽样法，选择 2023 年 9 月— 2024 年 6 月本市 2 所三级甲等综合医院 650 例住院老年 CHF 患者进行横断面研究，共 607 例完成研究。将是否发生认知衰弱作为结局指标，分为认知衰弱组和非认知衰弱组。通过单因素分析和逐步 Logistic 回归分析筛选最终纳入模型的变量，按照 7：3 的比例将总样本随机划分为训练集（n=424）和测试集（n=183）。在训练集上，基于神经网络（neural networks，NN）、线性判别分析（lineardiscriminant analysis，LDA）、K 最邻近（k-nearest neighbor，KNN）、支持向量机（support vector machine，SVM）、朴素贝叶斯（naiveBayes，NB）、Logistic 回归、决策树（decision tree，DT）与随机森林（random forest，RF）算法分别构建住院老年 CHF 认知衰弱预测模型，并利用测试集数据对模型的预测性能进行比较。结果 住院老年 CHF 患者认知衰弱患病率为 48.3%。Logistic 回归结果显示：年龄、婚姻状况、文化程度、体质量指数、多病共存、营养状况、用药情况、每周锻炼频率、居住状况是住院老年 CHF 患者的影响因素（均 P<0.05）。8 种算法构建的预测模型总体分类精度（accuracy）范围 0.803~0.847，F 值（F1-score）范围 0.778~0.833，精确度（precision）范围 0.848~0.897，召回率（recall）范围 0.700~0.778；受试者工作特征（receiver operating characteristic curve，ROC）曲线下面积范围 0.820~0.901。结论 在构建的多个预测模型中，综合各模型评估指标，LDA 模型的表现最好，综合预测性能最佳，NN 模型综合预测性能最差。

Objective To explore the factors influencing cognitive frailty in elderly inpatients with chronic heart failure（CHF） during hospitalisation, 8 prediction models were developed with various machine learning algorithms to identify the best model as a guidance for medical staff on clinical interventions. Methods Convenience sampling method was used to select 650 elderly CHF inpatients who stayed in our hospital between September 2023 and June 2024 as the study objects in the cross-sectional investigation. A total of 607 patients had completed the study. The patients were divided into a cognitive frailty group and a non-cognitive frailty group according to the presence or absence of cognitive frailty. Variables were initially screened using univariate analysis and stepwise Logistic regression. The total sample was then randomly divided into a training set（n=424）and a testing set（n=183）of a 7：3 ratio. Eight predictive models were created using the algorithms of neural network（NN）, k-nearest neighbour（KNN）, linear discriminant analysis （LDA）, support vector machine（SVM）, naive Bayes（NB）, logistic regression, decision tree（DT）and random forest（RF）on the training set. The predictive performance of the models was compared using the data of the testing set. Results The prevalence of cognitive frailty in elderly CHF inpatients was 48.3%. Results of Logistic regression showed that age, marital status, education, body mass index, multi-morbidity, nutritional status, medication, frequency of weekly exercise and the living conditions were the key factors（P<0.05）. The overall accuracy in classification of the eight predictive models ranged from 0.803 to 0.847, with F1-values of 0.778 to 0.833, precision of 0.848 to 0.897, and recall rate of 0.700 to 0.778. The area under the receiver operating characteristic curve was 0.820 to 0.901. Conclusion Of the eight predictive models, the prediction model created with LDA shows the best performance and prediction in terms of comprehensive prediction metrics, while the prediction model created with NN shows the worst performance in comprehensive prediction.