基于机器学习算法的老年胆心综合征最优预测模型的筛选及预测因素分析

Optimal predicting model for cholecystocardiac syndrome in elderly patients based on machine learning algorithms and predicting factors

摘要:

目的 从多种机器学习算法中筛选出老年胆心综合征早期识别的最优预测模型，并探究胆心综合征的预测因素。方法 回顾性纳入杭州市第一人民医院2021年7月至2023年12月期间以胸前区不适为首发症状的胆心综合征患者共150例，采用倾向性评分匹配（propensity score matching，PSM）方法，按1∶1比例匹配年龄、性别、吸烟史、饮酒史等变量，选取胸前区不适为首发症状的同时间段非胆心综合征患者150例作为对照组。采集两组患者的人口学特征、入院时的生命体征、实验室指标、心电图检查结果等病例资料。比较梯度提升、logistic回归、随机森林、k近邻、多层感知器和支持向量机6种模型在老年胆心综合征患者早期识别中的预测性能，根据模型的曲线下面积（area under the curve，AUC）和Brier分数筛选出综合性能最优的模型，并依据其特征重要性排序确定关键预测因素。结果 在6种机器学习模型中，随机森林模型的AUC最高（0.84）、Brier分数最小（0.165），按预设筛选标准确定为最优预测模型。随机森林模型在独立测试集中的AUC值为0.87 （95%CI：0.735~0.908），灵敏度为86.7%（39/45），特异度为80.0%（36/45），阳性预测值为81.3%（39/48），阴性预测值为85.7%（36/42）。最重要的6个预测变量是降钙素原、血清淀粉样蛋白A、心率变异性低频与高频成分比值、心率变异性相邻正常RR间期差值>50 ms的比例、γ-谷氨酰转移酶/天冬氨酸转氨酶和直接胆红素。结论 随机森林模型可用于老年胆心综合征患者的早期风险识别，关键预测因素包括降钙素原、血清淀粉样蛋白A、心率变异性低频与高频成分比值，心率变异性相邻正常RR间期差值>50 ms的比例、γ-谷氨酰转移酶/天冬氨酸转氨酶和直接胆红素。

Abstract:

Objective To identify the optimal machine learning model for predicting cholecystocardiac syndrome (CCS) in elderly patients and determine key predictive factors. Methods A total of 150 elderly patients diagnosed as having CCS who presented with chest discomfort as the initial symptom were retrospectively included at Hangzhou First People''s Hospital from July 2021 to December 2023, with 150 propensity score-matched controls (1∶1 matching for age, gender, smoking, and drinking history) with chest discomfort but without CCS during the same period. Demographic characteristics, vital signs at admission, laboratory test results, and electrocardiogram (ECG) findings of all patients were collected. The predictive performance of six machine learning models—gradient boosting, logistic regression, random forest (RF), k-nearest neighbor (KNN), multilayer perceptron (MLP), and support vector machine (SVM) for early CCS recognition was compared. The model with the best overall performance was selected according to area under the curve (AUC), and its key predictive factors were identified based on feature importance ranking. Results Among the six machine learning models, the RF model demonstrated the highest AUC (0.84) and lowest Brier score (0.165) and was therefore identified as the optimal model. In the independent test set, the RF model achieved an AUC of 0.87 (95%CI: 0.735‑0.908), with a sensitivity of 86.7% (39/45), specificity of 80.0% (36/45), positive predictive value of 81.3% (39/48) and negative predictive value of 85.7% (36/42). The six most important predictive variables were procalcitonin (PCT), serum amyloid A (SAA), heart rate variability low-frequency to high-frequency ratio (LF/HF), percentage of successive normal RR intervals differing by >50 ms (PMN50), γ-glutamyltransferase/aspartate aminotransferase ratio (GGT/AST), and direct bilirubin (DBIL). Conclusion The RF model can predict early risk of CCS in elderly patients, with the key predictive factors including PCT, SAA, LF/HF, PMN50, GGT/AST, and DBIL.