Cheng Baozhi, Ma Jianli, Chen Xiaolong, Yuan Lingyan
Department of Endocrine, Lanzhou Second People's Hospital, Lanzhou, China.
Department of Pediatrics, Lanzhou Second People's Hospital, Lanzhou, China.
Ann Transl Med. 2022 Sep;10(18):960. doi: 10.21037/atm-22-3580.
A predictive model of facial feature data was established by machine learning to screen the objective parameters of risk factors of facial morphological features of type 2 diabetes mellitus (T2DM) following the theory of traditional Chinese medicine (TCM). In TCM, a facial inspection is an important way to diagnose patients. Doctors can judge the health status of their patients by observing their facial features. However, the lack of description of the objective parameters and quantitative indicators hinders the development of TCM testing research.
In this study, the following diagnostic criteria for diabetes developed by the World Health Organization (WHO) in 1999 were used to determine the inclusion and exclusion criteria for T2DM and non-T2DM. T2DM patients and control participants were enrolled in the study, and their facial images were collected. In this study, two facial inspection risk-factor models were constructed, including the "lambda.min" and "lambda.1se" model.
A total of 81 key points in the facial images were screened, and 18 facial morphological parameters were measured. The least absolute shrinkage and selection operator (LASSO) regression model was used to construct T2DM facial inspection risk-factor models. The area under the curves (AUCs) of the "lambda.min" model and the "lambda.1se" model were 0.799 and 0.776, respectively. The predictive efficiency of the two T2DM risk models selected by the LASSO regression model was relatively high. Among the eight parameters, the width of the jaw was the most important of the defined facial features. According to the receiver operating characteristic (ROC) curve analysis of the two prediction models constructed, the two models had good predictive efficiency for T2DM. The AUCs of the two models were 0.695 and 0.682, respectively. And the reproducibility is good. The prediction model was available, which showed that the objective parameters of the facial features recognized by machine learning have a certain value in the automatic prediction of T2DM.
The influence of facial features is physical factor. Thus, the objective parameters of facial features should be specific to differential diagnosis of T2DM.
遵循中医理论，通过机器学习建立面部特征数据预测模型，以筛选2型糖尿病（T2DM）面部形态特征危险因素的客观参数。在中医中，面部望诊是诊断患者的重要方法。医生可通过观察患者面部特征判断其健康状况。然而，缺乏客观参数和定量指标的描述阻碍了中医检测研究的发展。
本研究采用世界卫生组织（WHO）1999年制定的以下糖尿病诊断标准来确定T2DM和非T2DM的纳入和排除标准。招募T2DM患者和对照参与者，并收集他们的面部图像。本研究构建了两个面部望诊危险因素模型，包括“lambda.min”模型和“lambda.1se”模型。
共筛选出面部图像中的81个关键点，测量了18个面部形态参数。采用最小绝对收缩和选择算子（LASSO）回归模型构建T2DM面部望诊危险因素模型。“lambda.min”模型和“lambda.1se”模型的曲线下面积（AUC）分别为0.799和0.776。LASSO回归模型选择的两个T2DM风险模型的预测效率相对较高。在八个参数中，下颌宽度是所定义面部特征中最重要 的。根据所构建的两个预测模型的受试者工作特征（ROC）曲线分析，这两个模型对T2DM具有良好的预测效率。两个模型的AUC分别为0.695和0.682。并且可重复性良好。该预测模型可用，表明机器学习识别的面部特征客观参数在T2DM的自动预测中具有一定价值。
面部特征的影响是生理因素。因此，面部特征的客观参数应特定于T2DM的鉴别诊断。
