Machine Learning-Assisted Tongue Image Analysis for the Diagnosis of Hashimoto's Thyroiditis

Ting Ruan¹Wenjun Wu²Mingji Piao³Yihan Sun⁴Xingai Jv⁵Mengyou Liu⁶Li Lu⁷Bo Zhang⁸Yifei Zeng⁹Dongxiao Zhang⁹Yongxin Li^10*Jianchun Cui^11*

• ¹Liaoning University of Traditional Chinese Medicine, Shenyang, China
• ²Changchun University of Chinese Medicine, Changchun, China
• ³Department of General Surgery, Liaoning University of Traditional Chinese Medicine Affiliated Hospital, Shenyang, China
• ⁴Department of Cardiology, People's Hospital of Liaoning Province, Shenyang, China
• ⁵Department of General Medicine, People's Hospital of Liaoning Province, Shenyang, China
• ⁶Department of Thyroid and Breast Surgery, The People's Hospital of Lixin County, Bozhou, China
• ⁷Department of Endocrinology, People's Hospital of Liaoning Province, Shenyang, China
• ⁸The 10th Division Beitun Hospital, Xinjiang Production and Construction Corps, Urumqi, China
• ⁹Beijing Hospital of Traditional Chinese Medicine, Beijing, China
• ¹⁰Yizhun Medical AI Co Ltd Department of Research & Development, Beijing, China
• ¹¹Department of Thyroid and Breast Surgery, People’s Hospital of China Medical University (Liaoning Provincial People’s Hospital),, Shenyang, China

Objective: This study aims to evaluate the value of a machine learning model based on tongue features in the adjunctive diagnosis of Hashimoto's thyroiditis (HT) and its concomitant hypothyroidism. Methods: Tongue images and related clinical data were retrospectively collected from 120 HT patients (60 each from the euthyroid group and the hypothyroidism group), and the tongue region was segmented by preprocessing, and the image feature dimensions were extracted with 1125 dimensions. Therefore, four methods, namely, random forest (RF), logistic regression (LR), support vector machine (SVM), and decision tree (DT), were utilized for model training, and 80 tongue images of 40 patients from Lixin County People's Hospital in Anhui Province were utilized for external validation. The model performance evaluation indexes included AUC (Area Under the Curve), Sensitivity, Specificity, Positive Predictive Value (PPV), and Negative Predictive Value (NPV). Results: t-Distributed Stochastic Neighbor Embedding (t-SNE) visualization based on the test set revealed a distinguishable clustering trend between the two groups. The key classification features included tongue texture uniformity, body morphological features, and color depth. The AUC of the four models was higher than 0.82, confirming that the tongue image features have significant predictive value for HT, and the lower limit of 95% CI for all models was higher than 0.75, indicating that the models had stable differentiation ability. The AUC of SVM(0.894) was the best, significantly higher than the other models (RF:0.857, LR: 0.876, and DT:0.828), indicating that the SVM possesses the strongest ability to classify patients with and without HT and the highest stability. The SVM exhibited balanced performance, with a sensitivity of 0.804 and specificity of 0.936. Consequently, it represents the optimal model for achieving an equilibrium between recall and precision. In external validation, the efficacy of the four models is notable, and the trend is consistent with the test set. SVM still demonstrates notable performance and possesses the best generalization ability among the four models. Conclusion:The tongue image-based machine learning model can effectively assist in distinguishing euthyroid from hypothyroidism in HT, offering a non-invasive, low-cost, and intelligent tool for auxiliary diagnosis and disease risk monitoring in primary care settings.