Liu Yongyue, Fan Linmiao, Zhao Mei, Wei Dongshen, Zhao Menglan, Dong Yihang, Zhang Xiaoqing
School of Life Sciences, Beijing University of Chinese Medicine, No. 11, North 3rd Ring Road East, Beijing, 100029, China.
Tsinghua University Institute of Advanced Equipment (Tianjin), Tianjin, 300300, China.
Chin Med. 2025 Jun 12;20(1):84. doi: 10.1186/s13020-025-01126-w.
The objective of this study was to develop a quantitatively analyzed Traditional Chinese Medicine (TCM) constitution recognition model utilizing tongue fusion features and deep learning techniques.
A prospective investigation was conducted on participants undergoing TCM constitution assessment at two medical centers. Tongue images and corresponding TCM constitution data were collected from 1374 participants using specialized equipment. Both traditional and deep features were extracted from these images. Significant features associated with constitutional characteristics were identified through LASSO regression and Random Forest (RF). Eight machine learning algorithms were employed to construct and evaluate the efficacy of the models. The highest-performing model was selected as the foundational classifier for developing an integrated tongue image feature model. Model performance was comprehensively evaluated using accuracy, precision, recall, F1 score, and area under the curve (AUC).
Analysis revealed 11 critical traditional tongue image features and 26 deep tongue image features. Three datasets were constructed: traditional tongue image features, deep tongue image features, and a fusion feature dataset incorporating both. The multilayer perceptron (MLP) model combining traditional and deep features demonstrated superior performance in TCM constitution classification compared to single-feature models. In the training phase, the model achieved an accuracy (ACC) of 0.893 and an AUC of 0.948. On the test set, it achieved an ACC of 0.837 and an AUC of 0.898, with sensitivity and specificity of 0.680 and 0.930, respectively, indicating excellent generalization ability.
This study successfully developed an intelligent TCM constitution recognition model that overcomes the limitations of traditional methods and validates the value of tongue images for accurate constitution recognition.
本研究的目的是利用舌融合特征和深度学习技术开发一种定量分析的中医体质识别模型。
对两个医疗中心接受中医体质评估的参与者进行前瞻性调查。使用专门设备从1374名参与者收集舌图像和相应的中医体质数据。从这些图像中提取传统特征和深度特征。通过LASSO回归和随机森林（RF）识别与体质特征相关的显著特征。采用八种机器学习算法构建和评估模型的有效性。选择性能最佳的模型作为开发集成舌图像特征模型的基础分类器。使用准确率、精确率、召回率、F1分数和曲线下面积（AUC）全面评估模型性能。
分析揭示了11个关键的传统舌图像特征和26个深度舌图像特征。构建了三个数据集：传统舌图像特征、深度舌图像特征以及包含两者的融合特征数据集。与单特征模型相比，结合传统特征和深度特征的多层感知器（MLP）模型在中医体质分类中表现出卓越性能。在训练阶段，该模型的准确率（ACC）达到0.893，AUC为0.948。在测试集上，其ACC为0.837，AUC为0.898，灵敏度和特异性分别为0.680和0.930，表明具有出色的泛化能力。
本研究成功开发了一种智能中医体质识别模型，克服了传统方法的局限性，并验证了舌图像在准确体质识别中的价值。
