带状疱疹后神经痛发生的治疗相关因素分析及其XGBoost临床预测模型的构建

Analysis of treatment-related factors of postherpetic neuralgia and construction of XGBoost clinical prediction model

目的分析影响带状疱疹后神经痛(PHN)发生的危险因素,特别是治疗相关因素,择优选择机器学习算法构建PHN临床预测模型。方法选取2023年5月至10月于西京医院疼痛医学中心门诊就诊的434名带状疱疹患者的病历资料。收集人口学因素、疱疹相关因素、治疗相关因素及合并疾病等指标。在患者病程满3个月后,根据疼痛VAS判断是否发生PHN,分为PHN组(n=197)和非PHN组(n=237)。使用单因素分析、logistic回归挑选变量,随后使用LASSO回归对挑选出的因素进行筛选和降维,选择最终纳入模型的变量。比较传统logistic回归模型和XGBoost、SVM两种机器学习模型的区分性能,选择最佳算法构建模型并进行验证和评价。结果研究通过LASSO回归筛选发现神经节段、年龄、急性期VAS、疱疹面积、神经阻滞治疗开始时间和疼痛性质是PHN发生的独立影响因素。logistic回归、XGBoost模型和SVM模型在训练集的ROC-AUC均值分别为0.82、0.95、0.77,在验证集的均值分别为0.81、0.81、0.76,提示XGBoost模型的预测性能最佳。使用XGBoost建模,训练集、验证集的ROC-AUC均值及95%CI分别为0.94(0.92~0.97)、0.86(0.79~0.94),提示模型区分度较好。Hosmer-Lemeshow拟合优度检验P>0.05,校准曲线比较接近理想曲线,说明模型预测性能良好。决策分析曲线显示模型具有良好的临床净收益。结论神经受累节段和神经阻滞治疗开始时间这两个治疗相关因素是PHN发生的重要独立影响因素。采用神经节段、年龄、急性期VAS、疱疹面积、神经阻滞治疗开始时间、疼痛性质六个变量构建的XGBoost临床预测模型性能优良,具有较好的区分度和校准度;对于早期甄别PHN高危患者,及时进行针对性治疗具有重要临床意义。

Objective To analyze the risk factors affecting the occurrence of postherpetic neuralgia(PHN),especially the treatment-related factors,and select the optimal machine learning algorithm to construct the PHN clinical prediction model.Methods The medical records of 434 patients with herpes zoster treated in the outpatient clinic of Pain Medicine Center of Xijing Hospital from May to October 2023 were selected.Demographic factors,herpes-related factors,treatment-related factors,and co-morbidities were collected.After 3 months of disease course,patients were divided into PHN group(n=197)and non-PHN group(n=237)according to the pain VAS.Univariate analysis and logistic regression were used to select variables,and then LASSO regression was used to screen and dimensionally reduce the selected factors to determine the final variables to be included in the model.The differentiation performance of traditional logistic regression model and two machine learning models(XGBoost and SVM)was compared,and the optimal algorithm was selected for model construction and verification and evaluation.Results LASSO regression screened ganglion segment,age,acute phase VAS,herpes area,time to start nerve block therapy,and nature of pain as independent influencing factors for PHN.The mean ROC-AUC values of logistic regression model,XGBoost model,and SVM model in the training set were 0.82,0.95,and 0.77,respectively,and those in the validation set were 0.81,0.81,and 0.76,respectively,suggesting that the XGBoost model had the best prediction performance.Using XGBoost to construct the final prediction model,the mean ROC-AUC values and 95%CI of the training set and validation set were 0.94(0.920.97)and 0.86(0.790.94),respectively,indicating a better model differentiation.Hosmer-Lemeshow goodness-of-fit test(P>0.05)showed that the calibration curve was close to the ideal curve,indicating that the model has good prediction performance.The decision curve analysis showed that the model had an excellent net clinical benefit.Conclusion The two treatment-related factors,the affected nerve segment and the time to start nerve block therapy,are critical independent influencing factors for the occurrence of PHN.The XGBoost clinical prediction model constructed by ganglion segment,age,acute phase VAS,herpes area,time to start nerve block therapy,and nature of pain performs well with good differentiation and calibration.For early identification of PHN high-risk patients,timely targeted treatment has important clinical significance.