文章摘要
基于机器学习算法构建老年新冠肺炎患者重症风险预测模型
Construction of Severe Risk Prediction Model for Elderly Patients with COVID-19 Based on Machine Learning Algorithm
投稿时间：2025-08-04  修订日期：2026-06-12
DOI：
中文关键词: 新型冠状病毒肺炎  重症肺炎  诊断模型  横断面研究  临床预测模型  机器学习
英文关键词: Novel coronavirus pneumonia  Severe pneumonia  Diagnostic model  Cross-sectional study  Clinical prediction model  machine learning
基金项目:；吴阶平医学(320.6750.2025-01-15)
作者单位邮编
梅雯倩 安徽医科大学附属巢湖医院 238000
孟凡亮* 安徽医科大学附属巢湖医院 
摘要点击次数: 78
全文下载次数: 0
中文摘要:
      【】　背景　机器学习算法在诸多领域已获得广泛应用,然而,专门针对老年新冠肺炎患者重症情况的机器学习模型尚不多见。开发一款便捷且高效的预测模型不仅能够协助临床医生在早期准确判断患者病情的严重程度,还能优化重症监护室资源的合理分配。目的　利用机器学习算法构建老年新型冠状病毒肺炎(COVID-19)患者重症风险预测模型,并开发可实际运用的网页计算器。方法　对安徽医科大学附属巢湖医院2022年12月-2023年2月收治的≥60岁新型冠状病毒肺炎患者进行回顾性研究,从电子病历中提取人口学信息、伴随症状、入院实验室检查结果、合并症、影像学资料等。根据《新型冠状病毒感染诊疗方案(试行第十版)》诊断标准分为非重症组(轻型、中型)和重症组(重型、危重型),并按7∶3比例随机分为训练集和验证集,其中训练集用于筛选危险因素和风险模型的建立,验证集用于模型的内部验证和性能评价。采用受试者工作特征(ROC)曲线下面积AUC值、校准曲线和临床决策曲线(DCA)评价各机器学习模型对老年患者重症风险的预测能力。用Shap解释表现最佳的模型。结果　最终纳入449例老年新型冠状病毒肺炎患者,其中训练集314例(非重症组199例、重症组115例),验证集135例(非重症组85例、重症组50例)。训练集的单因素分析结果显示:重症组和非重症组的乳酸脱氢酶(LDH)、白蛋白(ALB)、 降钙素原(PCT)、血尿素氮(BUN)、白蛋白/球蛋白比值(A/G)、D-二聚体(D-D)、C反应蛋白(CRP)、白细胞(WBC)、部分凝血活酶时间(APTT)、中性粒细胞与淋巴细胞比值(NLR)、钙离子(Ca2+)等30个临床指标差异具有统计学意义(P均<0.05)。Lasso回归最终保留了12个预测因子,其中危险因素包括LDH、PCT、BUN、D-D、CRP、APTT、呼吸困难、心脏病史、细菌感染、多肺叶累及(>3个)；保护因素包括ALB、Ca2+。利用以上预测因子分别基于自适应提升(AdaBoost)、K最近邻(KNN)、梯度提升机(GBM)、极端梯度提升(XGBoost)、支持向量机(SVM)、神经网络多层感知机(NeuralNetwork(MLP))、逻辑回归(Logistic regression)这7种机器学习算法构建风险预测模型。综合AUC和校准曲线、临床决策曲线三者结果考虑,最佳模型是NeuralNetwork(MLP)模型,其在训练集的AUC值为0.895(95%CI:0.859-0.931),在验证集的AUC为 0.844(95%CI:0.778-0.909)。并基于该模型建立网页计算器。利用Shap解释各预测因子在模型中的贡献,得出最重要的特征变量是LDH,其次是ALB。
英文摘要:
      【】　Background　Machine learning algorithms have been widely used in many fields. However, there are few machine learning models specifically for the severe condition of elderly patients with COVID-19. Developing a convenient and efficient predictive model can not only assist clinical doctors in accurately assessing the severity of patients" conditions in the early stages but also optimize the rational allocation of intensive care unit resources. Objective　To construct a model for predicting the risk of severe illness in elderly patients with novel coronavirus pneumonia (COVID-19) using machine learning algorithms, and to develop a practical web calculator. Methods　A retrospective study was conducted on patients with novel coronavirus pneumonia aged ≥ 60 years who were admitted to Chaohu Hospital affiliated to Anhui Medical University from December 2022 to February 2023. Demographic information, accompanying symptoms, admission laboratory examination results, complications, imaging data, etc,were extracted from electronic medical records. According to the diagnostic criteria of Diagnosis and Treatment Plan for novel coronavirus Infection (Tenth Edition on Trial), they were divided into non severe group (light, medium) and severe group (heavy, critical), and randomly divided into training set and verification set according to the ratio of 7:3. The training set was used to screen risk factors and establish risk models, and the verification set was used for internal verification and performance evaluation of models. Evaluate the predictive ability of various machine learning models for the risk of severe illness in elderly patients using the area under the receiver operating characteristic (ROC) curve AUC value, calibration curve, and clinical decision curve (DCA). Using Shap to explain the best performing model. Results　Finally, 449 elderly patients with novel coronavirus pneumonia were included, including 314 cases in the training set (199 cases in the non-severe group, 115 cases in the severe group) and 135 cases in the validation set (85 cases in the non-severe group, 50 cases in the severe group). The results of the univariate analysis of the training set showed that there were statistically significant differences (P<0.05) in 30 clinical indicators, including lactate dehydrogenase (LDH), albumin (ALB), procalcitonin (PCT), blood urea nitrogen (BUN), albumin/globulin ratio (A/G), D-dimer (D-D), C-reactive protein (CRP), white blood cell count, neutrophil to lymphocyte ratio (NLR), and calcium ion (Ca2+), between the severe and non-severe groups. Lasso regression ultimately retained 12 predictive factors, including risk factors such as LDH, PCT, BUN, D-D, CRP, APTT, dyspnea, cardiopathy, bacterial infection, and involvement of multiple lung lobes (>3); Protective factors include ALB and Ca2+. Using the above predictive factors, risk prediction models are constructed based on seven machine learning algorithms: Adaptive Boosting (AdaBoost), K-Nearest Neighbor (KNN), Gradient Boosting Machine (GBM), Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), Neural Network (MLP), and Logistic Regression. Considering the AUC, calibration curve, and clinical decision curve, the optimal model is the Neural Network (MLP) model, with an AUC value of 0.895 (95% CI: 0.859-0.931) in the training set and 0.844 (95% CI: 0.778-0.909) in the validation set. And based on this model, establish a web calculator. Using Shap to explain the contributions of various predictive factors in the model, it is found that the most important feature variable is LDH, followed by ALB. Conclusion　Seven prediction models were established based on machine learning, among which the Neural Network (MLP) model performed the best, and a user-friendly web calculator was developed.
View Fulltext   查看/发表评论  下载PDF阅读器
关闭