Evaluating deep learning techniques for identifying tongue features in subthreshold depression: a prospective observational study

Abstract

ObjectiveThis study aims to evaluate the potential of using tongue image features as non-invasive biomarkers for diagnosing subthreshold depression and to assess the correlation between these features and acupuncture treatment outcomes using advanced deep learning models.MethodsWe employed five advanced deep learning models—DenseNet169, MobileNetV3Small, SEResNet101, SqueezeNet, and VGG19_bn—to analyze tongue image features in individuals with subthreshold depression. These models were assessed based on accuracy, precision, recall, and F1 score. Additionally, we investigated the relationship between the best-performing model’s predictions and the success of acupuncture treatment using Pearson’s correlation coefficient.ResultsAmong the models, SEResNet101 emerged as the most effective, achieving an impressive 98.5% accuracy and an F1 score of 0.97. A significant positive correlation was found between its predictions and the alleviation of depressive symptoms following acupuncture (Pearson’s correlation coefficient = 0.72, p<0.001).ConclusionThe findings suggest that the SEResNet101 model is highly accurate and reliable for identifying tongue image features in subthreshold depression. It also appears promising for assessing the impact of acupuncture treatment. This study contributes novel insights and approaches to the auxiliary diagnosis and treatment evaluation of subthreshold depression.