To compare the performance of support vector machine (SVM) -, BP neural network (BPNN) - and random forest (RF) -based models in predicting the risk of CAS in steelworkers.

4 568 steelworkers who underwent physical examination and health monitoring in Tangshan Hongci Hospital from March to June 2017 were selected for a survey using the Health Assessment Checklist developed by us for understanding their information about demographic characteristics (sex, age, BMI, education level, marital status) , personal behavior and lifestyle (smoking and drinking) , medical history (hypertension, diabetes, family history of CAS) , occupation history (current work in shifts, working under high temperature or in noisy environments) . Levels of serum cholesterol, triglyceride, homocysteine and uric acid were also collected. Variables for building SVM-, BPNN- and RF-based models for predicting the risk of CAS were determined using unconditioned multivariate Logistic regression analysis and literature review.

In predicting the risk of CAS in participants in the training set, the accuracy, sensitivity and specificity were 83.81%, 80.10%, 87.32%, respectively, for the SVM-based model, 79.27%, 66.19%, 91.62%, respectively, for the BPNN-based model, and 86.60%, 73.62%, and 98.90%, respectively, for the RF-based model. And the AUC for SVM-, BPNN- and RF-based models was 0.84, 0.79 and 0.86, respectively. The SVM-based model had the highest sensitivity, while the RF-based model had the highest accuracy and specificity (P<0.05) . In predicting the risk of CAS in participants in the test set, the accuracy, sensitivity and specificity were 85.70%, 81.63%, 90.29%, respectively, for the SVM-based model, 75.46%, 64.65%, 87.66%, respectively, for the BPNN-based model, and 73.37%, 60.00%, and 88.45%, respectively, for the RF-based model. And the AUC for SVM-, BPNN- and RF-based models was 0.86, 0.76, and 0.74, respectively. The SVM-based model had the greatest accuracy, sensitivity and AUC. The sensitivity, accuracy and AUC of the SVM-based model were significantly different from those of the BPNN- or RF-based model in predicting the CAS risk (P<0.05) .

The SVM-based model may be better than other two models in predicting the risk of CAS in steelworkers.

运用支持向量机（SVM）、BP神经网络（BPNN）与随机森林（RF）模型构建钢铁工人CAS发生风险预测模型，并比较其预测效能。

选取2017年3—6月在唐山市弘慈医院进行体检和健康监测的4 568例钢铁工人为研究对象，按照本团队编写的《健康评估检查表》进行调查，调查内容：人口学特征（性别、年龄、体质指数、文化程度、婚姻状况）、个人的行为生活习惯与方式（吸烟、饮酒）、个人病史（高血压、糖尿病、CAS家族史）、职业史（倒班、高温作业、噪声作业）。收集研究对象的实验室检查指标，如胆固醇、三酰甘油、同型半胱氨酸、尿酸。结合非条件多因素Logistic回归分析结果以及查阅相关文献，确定变量构建SVM、BPNN和RF模型并进行比较。

训练集显示SVM、BPNN和RF模型预测钢铁工人发生CAS的准确率分别为83.81%、79.27%、86.60%，灵敏度分别为80.10%、66.19%、73.62%，特异度分别为87.32%、91.62%、98.90%，受试者工作特征曲线下面积（AUC）分别为0.84、0.79、0.86。SVM模型的灵敏度最高，RF模型在准确率、特异度和AUC方面高于其余两种模型，差异有统计学意义（P<0.05）。测试集显示SVM、BPNN和RF模型预测钢铁工人发生CAS的准确率分别为85.70%、75.46%、73.37%，灵敏度分别为81.63%、64.65%、60.00%，特异度分别为90.29%、87.66%、88.45%，AUC分别为0.86、0.76、0.74。SVM模型在灵敏度、准确率和AUC方面高于其余两种模型比较，差异有统计学意义（P<0.05）。

运用SVM模型预测钢铁工人CAS发生风险的效果优于BPNN和RF模型。