初次甲状腺癌术后声带麻痹模型预测分析

Prediction and analysis on model of vocal cord paralysis after primary thyroid cancer surgery

摘要:

目的 构建初次行甲状腺癌手术患者术后声带麻痹情况的预测模型，预测声带麻痹发生几率。方法 回顾性分析2012年1月—2022年2月收集的394例甲状腺癌患者的临床资料，根据术后有无声带麻痹分为声带麻痹组与无声带麻痹组。根据时间顺序，将2012年1月—2021年5月收集的358例患者作为建模组，2021年6月—2022年2月收集的36例患者作为验证组，运用单因素及多因素，分析两组临床患者基本资料，选择最合适的自变量构建Logistic回归模型，分析甲状腺癌术后声带麻痹的影响因素。内部通过十折交叉验证，外部验证通过比较建模组及验证组的区分度、校准度及临床有效性，绘制受试者工作特征(ROC)曲线、校准曲线及临床决策曲线，评估Logistic回归模型预测价值，计算约登指数、灵敏度、特异度、预测概率P值，根据P=1/1+e^-y，计算出Y值，最后绘制Nomogram图。结果 394例患者均一次完成甲状腺癌手术切除，术后41例患者发生声带麻痹，声带麻痹发生率为10.4%。Logistic回归模型可知喉返神经入喉处是否粘连(OR=11.804，95%CI为3.078~45.273)、术前Tg(OR=0.021，95%CI为0.002~0.202)、是否贴近喉返神经(OR=20.984，95%CI为2.058~214.007)、手术时间(OR=2.768，95%CI为1.122~6.829)是甲状腺癌术后声带麻痹的独立预测因素。十折交叉验证显示ROC曲线下面积为0.7284，建模组ROC曲线下面积为0.794 3(95%CI为0.716~0.872)，验证组ROC曲线下面积为0.772 2(95%CI为0.526~1)；Hosmer-Lemeshow拟合优度检测显示模型拟合较好，建模组Chi-Square=1.1，P=0.981 6，验证组Chi-Square=3.87，P=0.567 7；约登指数最大为0.188，此时灵敏度为0.57，特异度为0.62，ROC曲线下面积为0.59，P=0.216，为最佳临界值，根据P=1/1+e^-y，此时Y=-1.25。结论 本回归模型预测准确度较好，对甲状腺癌术后预防声带麻痹的发生具有一定的参考意义。

关键词:甲状腺肿瘤;声带麻痹;风险因素;预测模型

Abstract:

Objective To construct a prediction model of vocal cord paralysis after primary thyroid cancer surgery so as to predict the incidence of postoperative vocal cord paralysis.Methods Clinical data of 394 patients undergoing primary thyroid cancer surgery in our hospital from Jan 2012 to Feb 2022 were collected and analyzed retrospectively. They were divided into vocal cord paralysis group and non-paralysis group based on the postoperative activity of vocal cord. According to the time sequence, 358 patients hospitalized between Jan 2012 and May 2021 were selected as the modeling group and 36 patients from June 2021 to Feb 2022 as the validation group. The basic data of patients in the two groups were analyzed by univariate and multivariate analyses to screening out the most appropriate independent variables for construction of a logistic regression model to analyze the influencing factors of vocal cord paralysis after thyroid cancer surgery. The internal was verified by ten fold cross validation, and the external by comparing the differentiation, calibration and clinical effectiveness of both groups. The receiver operating characteristic (ROC) curve, calibration curve and clinical decision-making curve were drawn to evaluate the predictive value of the logistic regression model. Jordan index, sensitivity, specificity and prediction probability p value were used to calculate the Y value according to P=1/1+e-y. And finally the nomogram was drawn.Results All the 394 patients completed primary surgical resection of thyroid cancer, and postoperative vocal cord paralysis occurred in 41 with an incidence of 10.4%. Logistic regression model showed that the recurrent laryngeal nerve conglutination entering the larynx (OR=11.804, 95%CI:3.078~45.273), preoperative TG (OR=0.021, 95%CI：0.002~0.202), getting close to the recurrent laryngeal nerve (OR=20.984,95%CI:2.058~214.007), and operation time (OR=2.768, 95%CI:1.122~6.829) were independent predictors of vocal cord paralysis after thyroid cancer surgery. Ten fold cross validation showed that the area under ROC was 0.7284, the area under ROC in the modeling group was 0.7943 (95%CI=0.716~0.872), and that in the validation group was 0.772 2(95%CI=0.526~1). Hosmer-Lemeshow goodness of fit test showed that the model fitted well, with chi square=1.1, P=0.9816>0.05 in the modeling group and chi square=3.87, P=0.567 7>0.05 in the validation group. The maximum Jordan index was 0.188, the sensitivity was then 0.57, the specificity was 0.62, the area under ROC was 0.59, P=0.216, which was the best critical value. According to P=1/1+e-y, y=-1.25.Conclusion With good prediction accuracy, this regression model has certain reference significance for the prevention of vocal cord paralysis after thyroid cancer surgery.

Key words:Thyroid neoplasm;Vocal cord paralysis;Risk factor;Predictive model