JMIR Publications

ABSTRACT

Background:

Tongue characteristics, widely utilized in traditional Chinese medicine for health assessment, have been shown to correlate with specific respiratory infections. With the ongoing global spread of Human adenoviruses (HAdVs), COVID-19, and other seasonal respiratory viruses, this study aims to enhance the convenience and cost-effectiveness of respiratory infection diagnoses by developing prediction models based on tongue characteristics. Method: This study utilized deep learning to extract features from 280 tongue images collected from COVID-19 patients, HAdVs patients, and healthy individuals. Machine learning diagnostic models were subsequently trained on these tongue characteristics to distinguish between normal cases and those indicative of COVID-19 and HAdVs infections. The key features identified by the machine learning algorithms were further visualized in a two-dimensional space. Result: Nine significant tongue features were identified: tongue coating color (red, green, blue), the presence of tooth marks, tongue coating crack ratio, tongue coating moisture level, texture directionality, texture roughness, and texture contrast. Diagnostic models trained on these features achieved an area under the precision-recall curve exceeding 70%, with the area under the receiver operating characteristic curve surpassing 80% for general performance. The SHAP value revealed that tongue color, moisture level, and texture direction were the most influential features. Conclusion: Our findings demonstrate the potential of tongue diagnosis in identifying pathogens responsible for acute respiratory tract infections at the time of admission. This approach holds significant clinical implications, offering the potential to reduce clinician workloads while improving diagnostic accuracy and the overall quality of medical care.