A machine-learning analysis of tongue images could accurately distinguished euthyroid from hypothyroid states in patients with Hashimoto’s thyroiditis, according to a retrospective multicenter study conducted in China.
Among four training methods, the support vector machine (SVM) showed the strongest diagnostic performance, with an area under the curve of 0.89, sensitivity of just over 80%, and specificity of nearly 94% in the internal test set. In external validation, the SVM maintained an area under the curve of 0.88 and sensitivity of 93%. Logistic regression, random forest, and decision-tree models also demonstrated high discriminatory capacity, each exceeding an area under the curve of 0.82.
Key predictive features included tongue texture uniformity, color depth, and morphologic features. Quantitative image data derived from gray-level co-occurrence and dependence matrices, along with first-order statistical features such as entropy and range, contributed to classification accuracy. “The external validation set is an important criterion for assessing the generalization ability of the models,” wrote lead study author Ting Ruan, of the Liaoning University of Traditional Chinese Medicine in Shenyang, China, and colleagues.
The investigators analyzed 555 standardized tongue images from 160 patients with Hashimoto’s thyroiditis recruited from two hospitals. The internal data set included 120 patients—60 in the euthyroid group and 60 in the hypothyroid group—while 40 additional patients formed the external validation cohort. Each participant provided four tongue images taken under uniform lighting and imaging conditions. All photographs were captured under fixed light intensity and camera distance.
A total of 1,125 quantitative features were extracted from each image. Following normalization and dimensionality reduction using the minimum-redundancy maximum-relevance algorithm, the four models were trained and tested. Statistical comparisons used t-tests or Mann–Whitney U tests for quantitative data and χ² or Fisher exact tests for categorical variables.
Patient demographics didn't differ significantly between the euthyroid and hypothyroid groups. As expected, the patients with hypothyroidism showed lower free triiodothyronine and thyroxine levels and higher thyroid-stimulating hormone concentrations across cohorts.
The investigators acknowledged the study’s retrospective design, limited sample size, and single-ethnic population as constraints. The external validation cohort, collected from Lixin County People’s Hospital in Anhui Province, was independent of the internal data set sites. They wrote that larger, more diverse populations and combined analyses of tongue-image, biochemical, and clinical variables may refine predictive performance.
In conclusion, the team reported that integrating tongue imaging with machine-learning analysis may provide a noninvasive, low-cost tool to assist in evaluating thyroid function in patients with Hashimoto’s thyroiditis. They indicated that such image-based screening could complement existing diagnostic approaches and facilitate earlier identification of thyroid dysfunction in resource-limited settings.
This study was supported by the Liaoning Science and Technology Department and the Beitun Science and Technology Bureau of the 10th Division of the Xinjiang Production and Construction Corps.
Ethical approval was obtained from institutional review boards at participating centers, with informed consent waived because of the retrospective design.
Source: Frontiers in Medicine
