基于肺癌患者舌象图像与中医临床症状的肺癌临床分期预测模型

A prediction model for clinical staging of lung cancer based on tongue image parameters and traditional Chinese medicine clinical symptoms of patients

关键词:  肺癌  舌象图像  预测模型  Logistic回归  机器学习

Key words:  lung cancer  tongue image  prediction model  Logistic regression  machine learning

摘要:

[目的] 构建基于肺癌患者基本信息、舌象图像参数与中医症状的肺癌临床分期预测模型。[方法] 参照临床流行病学横断面调查研究方法，探索肺癌舌象图像参数及相关影响因素，对肺癌患者舌象参数与中医症状调查问卷进行自变量筛选，将具有统计学意义的变量纳入二元Logistic回归分析，并应用Logistic回归、支持向量机、随机森林、极端梯度提升、K近邻分类算法、反向传播神经网络机器学习智能算法评价肺癌临床分期模型的预测能力。[结果] 相关性分析结果显示经单因素分析20个变量与肺癌临床分期进展存在相关性，分别为年龄（OR=1.618，P<0.001）、卡氏功能状态（KPS）评分（OR=2.416，P<0.001）、既往病史（OR=2.104，P<0.05）、吸烟史（OR=2.275，P<0.001）、饮酒史（OR=1.357，P<0.05）、病程（OR=1.257，P<0.001）、放疗（OR=0.631，P<0.001）、苔色（CC）-B5（OR=1.807，P<0.001）、腻苔（OR=1.612，P<0.001）、裂纹舌（OR=1.988，P<0.05）、自汗（OR=1.775，P<0.05）、视物昏糊（OR=1.495，P<0.001）、口干（OR=1.691，P<0.001）、干咳少痰（OR=1.443，P<0.01）、胸痛（OR=1.849，P<0.05）、少神（OR=1.561，P<0.05）、面色晦暗（OR=2.081，P<0.001）、唇色淡白（OR=1.184，P<0.05）、皮肤甲错（OR=1.299，P<0.05）、喘息（OR=1.194，P<0.05）。以肺癌临床分期为因变量，判别模型预测概率为自变量绘制受试者工作特征（ROC）曲线，Logistic回归模型ROC曲线下面积（AUC）为0.946；Logistic回归模型Ⅰ~Ⅳ期AUC分别为0.901、0.960、0.953、0.971。Logistic回归预测概率的AUC面积为0.946，95%CI（0.877，0.973）；随机森林（RF）算法预测概率的AUC面积为0.945，95%CI（0.802，0.977）；支持向量机（SVM）预测概率的AUC面积为0.942，95%CI（0.864，0.952）、极端梯度提升（Xgboost）预测概率的AUC面积为0.931，95%CI（0.814，0.948）、反向传播神经网络（BP）神经网络预测概率的AUC面积为0.930，95%CI（0.793，0.965）、K近邻分类算法（KNN）预测概率的AUC面积为0.927，95%CI（0.775，0.946）。[结论] 基于患者基本信息、舌象图像参数与中医症状，应用Logistic回归和机器学习方法构建肺癌临床分期预测模型可行，具有较好预测能力和分类效能，具有促进辅助诊疗、判断预后等临床价值。

Abstract:

[Objective] To construct a clinical stage prediction model of lung cancer based on the basic information of lung cancer patients，tongue image parameters and traditional Chinese medicine（TCM） symptoms. [Methods] With reference to the cross-sectional investigation and research method of clinical epidemiology，the macro and micro characteristics of lung cancer tongue image and related influencing factors were explored. Independent variables were screened for tongue image parameters of lung cancer patients and TCM symptoms questionnaire，and statistically significant variables were included in binary Logistic regression analysis. Logistic regression，support vector machine，random forest，extreme gradient lifting，K-nearest neighbor classification algorithm and backpropagation neural network machine learning intelligent algorithm were used to evaluate the predictive ability of lung cancer clinical stage model. [Results] Correlation analysis results showed that 20 variables were correlated with the clinical stage progression of lung cancer through univariate analysis. They were age（OR=1.618，P<0.001），KPS score（OR=2.416，P<0.001），medical history（OR=2.104，P<0.05），smoking history（OR=2.275，P<0.001），drinking history（OR=1.357，P<0.05），and course of disease（OR=1.257，P<0.001），radiotherapy（OR=0.631，P<0.001），CC-B5（OR=1.807，P<0.001），greasy coating（OR=1.612，P<0.001），cracked tongue（OR=1.988，P<0.05），spontaneous sweating（OR=1.775，P<0.05）Blurred vision（OR=1.495，P<0.001），dry mouth（OR=1.691，P<0.001），dry cough with little sputum（OR=1.443，P<0.01），chest pain（OR=1.849，P<0.05），oligopsia（OR=1.561，P<0.05），dull complexion（OR=2.081，P<0.001），pale lip color（OR=1.184，P<0.05） skin onychia（OR=1.299，P<0.05），wheezing（OR=1.194，P<0.05）. The ROC curve was drawn with the clinical stage of lung cancer as the dependent variable and the prediction probability of the discriminant model as the independent variable. The area under the ROC curve of the Logistic regression model was 0.946. In Logistic regression model，the AUC of stages Ⅰ to Ⅲ were 0.901，0.960，0.953 and 0.971，respectively.The AUC area of Logistic regression prediction probability was 0.946，95%CI（0.877，0.973）. The AUC area of RF algorithm prediction probability is 0.945，95%CI（0.802，0.977）. The AUC area of SVM prediction probability is 0.942，95%CI（0.864，0.952） and XGboost prediction probability is 0.931，95%CI（0.814，0.948） and BP neural network prediction probability is 0.930.The AUC area of 95%CI（0.793，0.965） and KNN prediction probability is 0.927，95%CI（0.775，0.946）. [Conclusion] Based on patients’ basic information，tongue image parameters and traditional Chinese medicine symptoms，it is feasible to construct lung cancer clinical stage prediction model by using Logistic regression and machine learning methods，which has good prediction ability and classification efficiency，and has clinical value of promoting auxiliary diagnosis and treatment，judging prognosis and risk early warning.