病原学阴性初治肺结核临床特征分析及预测模型构建

Analysis of clinical characteristics of microbiologically-negative new-onset pulmonary tuberculosis and a nomogram prediction model

目的分析病原学阴性初治肺结核临床资料,构建并评价预测模型,提高对病原学阴性初治肺结核诊断的准确率及规范性。方法本研究为回顾性队列研究,纳入2019年1月1日至2021年12月31日期间于新乡医学院第一附属医院就诊,入院诊断为肺部阴影的9356例患者。根据预先设定的纳入和排除标准进行筛选,最终纳入785例患者,利用R语言软件,通过设置随机种子数,按7∶3的比例将患者数据随机分为训练集和验证集,此方法为简单随机抽样。利用最小绝对收缩和选择算子进行自变量初步筛选,并应用logistic回归构建模型;通过受试者操作特征(ROC)曲线下面积、P-R曲线评价区分度,Hosmer-Lemeshow检验评估拟合优度,校准曲线、Bootstrap交叉验证、决策曲线分析、临床影响曲线进行综合评价。结果共纳入785例患者,男531例,女254例,年龄15~60岁,其中病原学阴性初治肺结核440例,非结核肺部疾病345例,筛选出9个自变量。在训练集ROC曲线截断值为0.530时,敏感度、特异度分别为0.935、0.953,P-R曲线精准率、召回率分别为0.965、0.935,Hosmer-Lemeshow检验结果为χ^(2)=2.238,P=0.327,Bootstrap交叉验证准确率和Kappa值分别为0.931、0.860。在验证集ROC曲线中截断值为0.530时,敏感度、特异度分别为0.891、0.920,P-R曲线精准率、召回率分别为0.922、0.891,Hosmer-Lemeshow检验结果为χ^(2)=3.351,P=0.187。结论本研究利用机器学习技术构建了病原学阴性初治肺结核诊断预测模型,该模型在训练集和验证集中均表现出较高的区分度和拟合优度,可有效预测病原学阴性初治肺结核的风险大小。该模型为提高病原学阴性初治肺结核诊断准确度提供了新的辅助工具,有助于改善结核病的早期诊断。

Objective To analyze clinical data of microbiologically-negative new-onset pulmonary tuberculosis,and to create a nomogram to enhance the accuracy and standardization of clinical diagnosis.Methods This was a retrospective cohort study.Totally 9356 patients diagnosed with pulmonary shadows in the First Affiliated Hospital of Xinxiang Medical University between January 1,2019,and December 31,2021,were included.Based on the inclusion and exclusion criteria,785 patients were ultimately enrolled.Using R package with a random seed number,patients were randomly divided into training and validation sets at a ratio of 7∶3 by a simple random sampling.The least absolute shrinkage and selection operator were utilized for an initial variable selection,followed by the construction of a logistic regression model.The discrimination performance of the nomogram was evaluated using the receiver operating characteristic(ROC)curve,the area under the curve(AUC)and the precision-recall(P-R)curve.The Hosmer-Lemeshow test was performed to assess the goodness of fit.The performance of the nomogram was comprehensively assessed by calibration curves,Bootstrap cross-validation,decision curve analysis,and clinical impact curves.Results A total of 785 patients were included,with 531 males and 254 females,aged 15 to 60 years.Among them,440 were new onset of microbiologically negative pulmonary tuberculosis and 345 had non-tuberculous pulmonary diseases.Nine independent variables were selected.In the training set,the optimal cut-off was 0.530,and the sensitivity,specificity,precision and recall of the P-R curve were 0.935,0.953,0.965 and 0.935,respectively.The Hosmer-Lemeshow test result of the nomogram in the training set yielded χ^(2)=2.238,and P=0.327.The Bootstrap cross-validation accuracy and Kappa value were 0.931 and 0.860,respectively.In the validation set,the optimal cut-off was 0.530,and the sensitivity,specificity,precision and recall of the P-R curve were 0.891,0.920,0.922 and 0.891,respectively.The Hosmer-Lemeshow test result was χ^(2)=3.351,P=0.187.Conclusions This study constructed a diagnostic nomogram for new-onset microbiologically negative pulmonary tuberculosis using machine learning techniques,showing high discrimination and goodness of fit in both the training and validation sets.It effectively predicts the risk of tuberculosis in patients with microbiologically-negative new-onset pulmonary tuberculosis.This model provides a new auxiliary tool to improve the diagnostic accuracy of microbiologically-negative new-onset pulmonary tuberculosis and contributes to the improvement of early diagnosis and treatment of tuberculosis.