广西HIV感染者/AIDS患者合并念珠菌感染预后及其机器学习预测模型  

Prognosis of HIV/AIDS patients with Candida infection in Guangxi and its machine learning prediction modeling study

作　　者：吴玉婷 陆贝贝 阳世雄 韦吴迪 石敏娟 孟思润 叶力 梁浩 谢志满[3] 蒋俊俊 WU Yuting;LU Beibei;YANG Shixiong;WEI Wudi;SHI Minjuan;MENG Sirun;YE Li;LIANG Hao;XIE Zhiman;JIANG Junjun(School of Public Health,Key Laboratory of AIDS Prevention and Control Research,Guangxi Medical University,Nanning,Guangxi 530021,China;Institute of Life Sciences,Collaborative Innovation Center of Regenerative Medicine and Medical Bioresource Development and Application Co-constructed by the Province and Ministry,Guangxi Medical University,Nanning,Guangxi 530021,China;Guangxi Nanning Fourth People's Hospital,Nanning,Guangxi 530021,China)

机构地区：[1]广西医科大学公共卫生学院,广西艾滋病防治研究重点实验室,广西南宁530021 [2]广西医科大学生命科学研究院,再生医学与医用生物资源开发应用省部共建协同创新中心,广西南宁530021 [3]广西南宁市第四人民医院,广西南宁530021

出　　处：《中国热带医学》2024年第10期1217-1224,共8页China Tropical Medicine

基　　金：国家重点研发计划子课题(No.2022YFC2305001);国家自然科学基金项目(No.81960602)。

摘　　要：目的预测影响人体免疫缺陷病毒感染者/艾滋病患者(human immunodeficiency virus/acquired immunodefi⁃ciency syndrome,HIV/AIDS)合并念珠菌感染预后的危险因素,为临床医生提供早期识别高危患者的预测指标。方法收集2012年1月—2019年6月在广西某传染病医院住院的HIV/AIDS合并念珠菌感染患者的临床数据,根据患者不同的预后结局分为死亡组和生存组。采用倾向性评分匹配方法按死亡∶生存=1∶3的比例随机选择病例构建模型。将数据按7∶3分为训练集和测试集,构建不同机器学习模型,综合评估模型性能选择最优模型作为最终的预测模型。最后使用SHAP值解释模型的特征,分析患者预后结局的影响因素。结果本研究共收集了3098例HIV/AIDS合并念珠菌感染患者。从2012年1月至2019年6月HIV/AIDS合并念珠菌感染患者的住院病死率呈线性平稳下降趋势(P=0.043)。使用倾向性得分匹配(propensity score matching,PSM)后得到1620例病例数据构建6种不同的机器学习模型,其中XG⁃Boost模型的性能表现最优[训练/测试集,曲线下面积(area under curve,AUC)为0.98/0.85,灵敏度为0.93/0.75,特异度为0.93/0.84]。呼吸衰竭、尿素和乳酸脱氢酶水平被认为是影响患者预后结局的三大因素。结论XGBoost模型在预测HIV/AIDS合并念珠菌感染患者的预后结局方面表现出良好的预测性能。该模型可为早期识别高危患者提供预警,协助临床医生采取个性化治疗措施,对指导临床决策具有重要意义。Objective To predict risk factors affecting the prognosis of HIV/AIDS patients with Candida infection,providing clinicians with predictors for early identifying high-risk patients.Methods Clinical data were collected from HIV/AIDS patients with Candida infection admitted to an infectious disease hospital in Guangxi from January 2012 to June 2019.Patients were divided into a death group and a survival group according to their different prognostic outcomes.Cases were randomly selected and matched using propensity score matching(PSM)at a ratio of 1∶3(death:survival)to construct the model.The data were split into a training set and a testing set at a ratio of 7∶3.Various machine learning models were built,and the optimal model was selected as the final prediction model by comprehensively evaluating the model performance.Finally,the SHAP values were used to interpret the features of the model and analyze the influencing factors of patients'prognostic outcomes.Results A total of 3098 HIV/AIDS patients with Candida infection were collected.From 2012 to June 2019,the in-hospital mortality rate of HIV/AIDS patients with Candida infection showed a linear and stable downward trend(P=0.043).After applying PSM,data from 1620 cases were used to construct six different machine learning models,among which the XGBoost model had the best performance(training/testing set,AUC=0.98/0.85,sensitivity=0.93/0.75,specificity=0.93/0.84).Respiratory failure,urea,and LDH levels were thought to be the three major factors affecting the prognostic outcomes of HIV/AIDS patients with Candida infection.Conclusions The XGBoost model showed good predictive performance in predicting prognostic outcomes of HIV/AIDS patients with Candida infection.The model can provide early warning for the identification of high-risk patients and assist clinicians to take personalized treatment measures promptly,which is of great significance for guiding clinical decision-making.

关 键 词：人体免疫缺陷病毒感染者/艾滋病患者 念珠菌感染 机器学习 预测模型 XGBoost 

分 类 号：R512.91[医药卫生—内科学]