Development and validation of prediction model for AIDS-related death after antiretroviral therapy based on random survival forest algorithm
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
NETL
NSTL
万方数据
目的 建立并验证接受抗逆转录病毒治疗(antiretroviral therapy,ART)的HIV感染者或AIDS患者(persons living with HIV or AIDS,PLWHA)发生AIDS相关性死亡的预测模型.方法 基于2003-2019年中国艾滋病综合防治信息系统中符合条件的PLWHA建立回顾性队列,采用随机过采样少数类(random over-sampling minority examples,ROSE)技术和随机生存森林算法建立和验证AIDS相关性死亡的预测模型,通过时间依赖Brier评分和时间依赖ROC曲线下面积对预测模型进行评价.根据训练集预测风险值的三分位数对验证集进行风险分层,通过Kaplan-Meier曲线和log-rank检验比较三组的生存率.结果 本研究共纳入360例研究对象,随访时间中位数为36.72(10.62,60.69)个月,120例研究对象发生AIDS相关性死亡.根据候选预测因子的重要性,选择年龄、CD4+T淋巴细胞计数、Hb和HCV感染建立预测模型1;在预测模型1的基础上,剔除变量重要性较低的HCV感染,建立预测模型2.在验证集中,两个预测模型预测1～5年AIDS相关性死亡风险的Brier评分均低于0.15,预测18～42个月AIDS相关性死亡风险的AUC均高于0.70,两者的校准度和区分度均较好.预测模型2的区分度优于预测模型1.基于预测模型2预测的风险值进行风险分层,死亡风险值＜2.37为低风险组,2.37～＜26.83为中风险组,≥26.83为高风险组,低、中和高风险组间生存率的差异有统计学意义(log-rank P＜0.001).结论 本研究建立的AIDS相关性死亡预测模型包含年龄、CD4+T淋巴细胞计数和Hb共3个预测因子,预测1～3年AIDS相关性死亡的校准度和区分度较好,对预测PLWHA个体AIDS相关性死亡风险及早期干预有潜在价值.
Objective To establish and verify the prediction model for AIDS-related death in persons living with HIV/AIDS(PLWHA)initiating antiretroviral therapy(ART).Methods Eligible PLWHA and their data were extracted from Chinese HIV/AIDS comprehensive prevention and control information system during 2003-2019 to establish a retrospective cohort.Random over-sampling minority examples(ROSE)technique and random survival forest algorithm were used to establish and validate a prediction model of AIDS-related death.The prediction model was evaluated by time-dependent Brier score and time-dependent area under the ROC curve(AUC).The threshold of risk stratification was set in the validation set according to the tertiles of the risk values in the training set,and survival rates among the three risk groups were compared using Kaplan-Meier curve and log-rank test.Results A total of 360 subjects were included,with a median follow-up time of 36.72(10.62,60.69)months,during which 120 subjects experienced AIDS-related deaths.Based on the importance of candidate predictors,a prediction model 1 was established including age,CD4+T lymphocyte count,hemoglobin(Hb),and HCV infection.Based on prediction model 1,a simplified prediction model 2 was developed by eliminating HCV infections with low importance.In the validation set,both prediction models showed Brier scores below 0.15 for predicting 1-5 year AIDS-related mortality risk and AUCs above 0.70 for predicting 18-42 month AIDS-related mortality risk,indicating good calibration and discrimination.Prediction model 2 demonstrated superior discrimination compared to prediction model 1.Based on the risk values by prediction model 2,risk stratification was performed,with a death risk value＜2.37 categorized as the low-risk group,2.37-＜26.83 as the medium-risk group,and ≥26.83 as the high-risk group.The differences in survival rates among the low,medium,and high-risk groups were statistically significant(log-rank P＜0.001).Conclusions The prediction model for AIDS-related death established in this study,comprising age,CD4+T lymphocyte count,and Hb,demonstrates good calibration and discrimination for predicting 1-3 year AIDS-related mortality.It has potential value for predicting individual AIDS-related mortality risk among PLWHA and guiding early interventions.
NETL
NSTL
万方数据
