基于不同算法比较筛选老年卒中后衰弱预警模型

Comparison of warning models for post-stroke frailty in elderly patients based on different algorithms

中文关键词: 机器学习算法  老年  脑卒中  衰弱  预警模型

英文关键词: machine learning algorithm  elderly  stroke  frailty  warning model

中文摘要:

      目的 比较不同算法构建的老年脑卒中后衰弱预警模型的预测效能.方法 回顾性纳入 2020 年 7 月至 2025年1 月成都医学院第二附属医院收治的 417 例老年脑卒中患者,按脑卒中后是否衰弱分为衰弱组(n=196)与非衰弱组(n=221).基于多因素分析筛选出的独立危险因素,构建逻辑斯谛回归(LR)、决策树(DT)、随机森林(RF)及支持向量机(SVM)预测模型.通过曲线下面积(AUC)、F1 分数(F1)和准确率评估性能;结合决策曲线分析(DCA)量化临床净获益;沙普利加性解释(SHAP)解析特征贡献,以筛选最优预警模型.结果 多因素分析表明,高龄、低体重指数(BMI)、缺乏体育锻炼、低社会支持评定量表(SSRS)评分、低巴塞尔指数(BI)及低白蛋白(ALB)水平是老年脑卒中后衰弱的独立危险因素(P<0.05).在预测模型比较中,RF 模型表现最优,其 AUC(0.893)、F1(0.721)及综合准确率均高于 LR、DT和 SVM 模型.DCA 进一步证实,RF 模型具有更高的临床净获益.SHAP 分析显示,年龄、体育锻炼、ALB 水平、BMI、BI及 SSRS 评分是影响模型预测的最关键特征.结论 基于年龄、体育锻炼、ALB 水平、BMI、BI 及 SSRS 评分构建的 RF 算法模型预测效能与临床实用性可能最优.

英文摘要:

      Objective To compare the predictive performance of different warning models for post-stroke frailty in elderly patients constructed by different algorithms.Methods A retrospective study included 417 elderly stroke patients admitted to Second Affiliated Hospital of Chengdu Medical College from July 2020 to January 2025.Patients were divided into frailty group(n=196)and non-frailty group(n=221)based on the presence of post-stroke frailty.Based on independent risk factors identified through multivariate analysis,predictive models of Logistic regression(LR),decision trees(DT),random forests(RF)and support vector machines(SVM)were constructed.The performance of each model was evaluated based on the area under the curve(AUC),the F1 score,and the accuracy.Clinical net benefit was quantified through decision curve analysis(DCA).The Shapley additivity explanation(SHAP)was used to analyze the contribution of features.Then,the optimal early warning model was identified.Results Multivariate analysis revealed that advanced age,low body mass index(BMI),lack of physical exercise,low social support rating scale(SSRS)score,low Barthel index(BI),and low albumin(ALB)level were independent risk factors for post-stroke frailty in the elderly(P<0.05).In the comparison of predictive models,the RF model performed the best,with higher AUC(0.893),F1 score(0.721),and overall accuracy than the LR,DT and SVM models.DCA further confirmed that the RF model provided higher clinical net benefit.SHAP analysis indicated that age,physical exercise,ALB level,BMI,BI,and SSRS score were the most critical features influencing the model's prediction.Conclusion The RF algorithm model constructed based on age,physical exercise,ALB levels,BMI,BI,and SSRS scores may offer optimal predictive performance and clinical utility.