| 摘要: |
| 目的 利用机器学习算法构建新型冠状病毒肺炎（COVID-19）患者临床结局的预测模型，并探索结局相关因子。方法 收集2020年2月5日至4月15日武汉市火神山医院及华中科技大学同济医学院附属同济医院光谷院区收治的COVID-19患者的临床指标与结局（院内死亡和院内接受气管插管治疗），利用人工神经网络（ANN）、朴素贝叶斯、logistic回归、随机森林4种机器学习算法构建患者临床结局的预测模型。结果 共纳入4 804例COVID-19患者，其中发生院内死亡100例（2.08%）、接受气管插管治疗87例（1.81%）。与院内死亡相关性最强的变量为白细胞计数、白蛋白、钙离子、血尿素氮、心肌型肌酸激酶同工酶和年龄，与院内接受气管插管治疗相关性最强的变量为白细胞计数、淋巴细胞绝对值、超敏CRP、总胆红素、钙离子和年龄，分别利用以上变量、基于4种机器学习算法构建院内死亡和院内接受气管插管治疗预测模型。4种预测模型中，相较于基于ANN、logistic回归、随机森林算法构建的模型[预测院内死亡的AUC值（95% CI）分别为0.938（0.882~0.993）、0.926（0.865~0.987）、0.867（0.780~0.954），预测院内接受气管插管治疗的AUC值（95% CI）分别为0.932（0.814~0.980）、0.935（0.817~0.981）、0.936（0.921~0.972）]，基于朴素贝叶斯算法构建的模型在预测COVID-19患者院内死亡（AUC=0.952，95% CI 0.925~0.979）和接受气管插管治疗（AUC=0.948，95% CI 0.896~0.965）方面性能均最佳。结论 4种机器学习算法在预测COVID-19患者临床结局方面性能良好，其中以基于朴素贝叶斯算法构建的预测模型最佳。白细胞计数、白蛋白、钙离子、血尿素氮、心肌型肌酸激酶同工酶和年龄可以用来预测COVID-19患者院内死亡，白细胞计数、淋巴细胞绝对值、超敏CRP、总胆红素、钙离子和年龄可以用来预测患者院内是否接受气管插管治疗。 |
| 关键词: 机器学习 算法 新型冠状病毒肺炎 医院死亡率 气管内插管 |
| DOI：10.16781/j.0258-879x.2021.10.1115 |
| 投稿时间：2021-05-07修订日期：2021-07-07 |
| 基金项目:上海市公共卫生体系建设三年行动计划学科建设项目（GWV-10.1-XK05），海军军医大学（第二军医大学）"三航"计划. |
|
| Prediction of in-hospital clinical outcomes of coronavirus disease 2019 patients based on machine learning algorithms |
| PENG Chi1,QI Ge-yao1,ZHANG Chen-xu1,GUO Yu-feng2,JIN Zhi-chao1* |
(1. Department of Health Statistics, Faculty of Health Services, Naval Medical University(Second Military Medical University), Shanghai 200433, China;
2. Medical Affair Office, Changzheng Hospital, Naval Medical University(Second Military Medical University), Shanghai 200003, China
*Corresponding author) |
| Abstract: |
| Objective To construct prediction models for the clinical outcomes of coronavirus disease 2019 (COVID-19) patients using machine learning algorithms, and explore the outcome-related factors. Methods The clinical indexes and outcomes (in-hospital mortality and receiving tracheal intubation) of COVID-19 patients who were admitted to Wuhan Huoshenshan Hospital or Guanggu Branch of Tongji Hospital of Tongji Medical College of Huazhong University of Science and Technology from Feb. 5 to Apr. 15, 2020 were collected. The prediction models for the clinical outcomes were constructed using artificial neural network (ANN), naive Bayes, logistic regression and random forest algorithms. Results A total of 4 804 COVID-19 patients were included, of whom 100 (2.08%) patients died and 87 (1.81%) patients received tracheal intubation during the hospitalization. White blood cell (WBC), albumin, calcium, blood urea nitrogen, creatine kinase-myocardial band (CK-MB) and age were the most correlated variables with in-hospital mortality. WBC, lymphocyte, hypersensitivity C reaction protein (hs-CRP), total bilirubin, calcium and age were the most correlated variables with in-hospital tracheal intubation. With the above variables and based on the 4 machine learning algorithms, the prediction models for in-hospital mortality and tracheal intubation were constructed. In the 4 prediction models, the model constructed based on naive Bayes algorithm had the best performance in predicting in-hospital mortality (area under curve[AUC]=0.952, 95% confidence interval[CI] 0.925-0.979) and tracheal intubation (AUC=0.948, 95% CI 0.896-0.965) versus the models constructed based on ANN, logistic regression and random forest algorithms (the AUC[95% CI] values for predicting in-hospital mortality were 0.938[0.882-0.993], 0.926[0.865-0.987] and 0.867[0.780-0.954], and the AUC[95% CI] values for predicting in-hospital tracheal intubation were 0.932[0.814-0.980], 0.935[0.817-0.981] and 0.936[0.921-0.972], respectively). Conclusion The 4 machine learning algorithms have good performance in predicting the clinical outcomes of COVID-19 patients. WBC, albumin, calcium, blood urea nitrogen, CK-MB and age can be used to predict the in-hospital mortality of COVID-19 patients; while WBC, lymphocyte count, hs-CRP, total bilirubin, calcium and age can be used to predict the in-hospital tracheal intubation. |
| Key words: machine learning algorithms coronavirus disease 2019 hospital mortality intertracheal intubation |
