Human adenoviruses (HAdVs) are prominent respiratory pathogens affecting individuals of all ages, leading to acute upper and lower respiratory tract diseases, such as pneumonia and bronchitis [1]. Nearly half (45.7%) of acute respiratory tract infection (ARTI) outbreaks between 2009 and 2020 were attributed to HAdV-7 in China [2]. Since 2020, a temporary decline in HAdVs activity has been noted due to the rapid global spread of COVID-19 and subsequent public health measures [3,4]. As COVID-19 persists in its global circulation, a resurgence of HAdVs and other seasonal respiratory virus infections has occurred [5]. The clinical presentations of ARTI caused by HAdVs and COVID-19 are similar, often presenting as mild symptoms such as fever, rhinorrhea, cough, and sore throat, which poses a significant diagnostic challenge in the absence of viral testing. HAdVs are more prone than COVID-19 to causing acute respiratory infections in children, including pharyngitis, tonsillitis, pharyngoconjunctival fever, bronchitis, and pneumonia [6]. Moreover, recent studies indicate that mixed respiratory viral infections may lead to more severe disease outcomes than single infections [7]. Given the tendency of many patients with ARTI to self-medicate, it is essential to develop methods to differentiate between these common viral infections. Such differentiation could promote timely medical consultation and help reduce the transmission of adenovirus to children.

Tongue image diagnosis is a straightforward, noninvasive, and valuable diagnostic method used in traditional Chinese medicine (TCM), which assesses key features, such as the color, size, and shape of the tongue, along with the color, thickness, and moisture level of the tongue coating [8]. In respiratory infections, TCM theory associates specific tongue coatings with distinct syndromes [9]. Epidemiological studies have also identified distinctive tongue features in patients with acute respiratory infections. Patients with COVID-19 have been reported to present features, such as fissured tongue and strawberry tongue [10-13], whereas tongue signs in HAdV infections remain scarcely documented, limiting comprehensive understanding. Given the convenience and cost-effectiveness of tongue imaging, its application for differentiating COVID-19 from HAdV infections shows promising potential.

Tongue diagnosis traditionally relies heavily on the experience and observational skills of TCM practitioners to interpret tongue features. However, advances in artificial intelligence (AI) technologies have enabled the extraction of tongue image characteristics, such as texture, color, and coating, facilitating more objective and intelligent tongue diagnosis. Recent studies have applied AI-assisted tongue image analysis in various diseases, including diabetes [14], fatty liver [15], non-small cell lung cancer [16], and gastric cancer [17], combining tongue image features with clinical indicators to develop exploratory disease risk prediction models. Similarly, Zhou et al [18] proposed an automatic multiview disease detection system using tongue images, achieving an average classification accuracy exceeding 95% for breast tumors, heart disease, fatty liver, and lung tumors. Given the limited research on AI-assisted tongue diagnosis for respiratory infections, there is a clear need to explore its potential in this domain. Therefore, this study aims to investigate the application of AI-based tongue image analysis to distinguish COVID-19 from adenoviral respiratory infections, with the goal of improving diagnostic accuracy and integrating traditional diagnostic methods with contemporary medical technologies.

In this study, we developed and validated an AI-based prediction model to analyze tongue images for the differential diagnosis of respiratory infections caused by COVID-19 and human adenoviruses. Our models demonstrated high diagnostic accuracy, with key tongue image features, such as color, coating, moisture, and texture serving as significant discriminators. These findings highlight the potential of noninvasive tongue image analysis as an effective and cost-efficient diagnostic tool, which may facilitate early and accurate differentiation of respiratory viral infections and support clinical decision-making.

In recent years, the concurrent prevalence of respiratory infectious diseases, such as COVID-19 and HAdVs, has made it challenging to establish a differential diagnosis without pathogenic testing, thereby impacting the efficacy of the treatment [29]. Tongue image diagnosis has emerged as an effective, noninvasive method for auxiliary diagnosis that can be conducted in various settings, catering to the global demands of primary health care systems [30]. In this study, we developed machine learning models that use tongue images to predict acute respiratory tract infection (ARTI) pathogens. The accuracy of these models exceeded 80%, with feature importance plots revealing tongue color, moisture level, and tongue texture direction as pivotal variables. We then projected the tongue image features of the 3 groups of participants into a lower-dimensional space by t-SNE plot. The resulting clusters of COVID-19 and adenovirus were notably distinct, suggesting that the tongue image parameters we extracted contribute significantly to the diagnosis of ARIs. This noninvasive, convenient, and rapid approach has the potential to mitigate unnecessary diagnostic procedures and reduce health care costs.

The constructability of a risk warning model using tongue diagnosis data has been validated in research on metabolic diseases and cancers [16,31]. Our study is pioneering in applying AI deep learning techniques to investigate the diagnostic value of tongue images in acute respiratory tract infection diagnosis and can be integrated into clinical practice to aid decision-making and alleviate physician burden. Previous research by Mai and Krauthammer [32] aimed to predict common respiratory viruses in the United States by combining natural language processing tools with machine learning techniques. However, the model’s prediction performance of HAdVs was moderate, with an area under the receiver operating characteristic curve (AUROC) of 0.53 [32]. Chang et al [33] used an XGBoost model, incorporating demographic, physical examination, laboratory, and vital sign data, to predict HAdVs in hospitalized children with respiratory symptoms, achieving an AUROC of 0.82. Our model exhibited superior performance, achieving an AUROC of 0.87, following the inclusion of various features in its development. In contrast to models established in previous studies, the model developed in this paper benefited from the usage of noninvasive tongue image data, obviating the need for laboratory examination indicators. By using the YOLOv4 object detection framework for image feature extraction and integrating ensemble learning algorithms (ie, GBM and XGBoost), we have demonstrated the clinical utility of tongue images. All 4 models exhibited discriminative validity over tongue images, indicating that tongue images can serve as a reliable tool for ARTI diagnosis and are robust across different AI deep learning model types. Notably, tree-based models, such as GBM, XGBoost, and random forest, outperform logistic regression due to their ability to capture complex nonlinear relationships and interactions among features. Tongue image-derived features often display nonlinear patterns that contradict the linear assumptions of logistic regression. In contrast, tree-based models are nonparametric and resistant to multicollinearity, making them ideal for complex datasets derived from image analysis.

Previous research has validated the use of objective tongue image acquisition equipment, methods, and data analysis techniques [34]. Research on AI in TCM tongue diagnosis has predominantly focused on standardizing tongue diagnosis to minimize human errors [34]. Xu et al [35] developed a multitask joint learning model for segmenting and classifying tongue images, using a deep neural network to optimally extract tongue image features. Meng et al [36] proposed a novel feature extraction framework, termed constrained high dispersal neural networks, to extract unbiased features and minimize human labor in TCM tongue image diagnosis. This study used the YOLOv4 framework, which excelled in extracting high-level image features, thereby enhancing the model’s capacity to identify intricate patterns and subtle variations [21]. The model was initially pretrained on the VOC 2007 and VOC 2012, which are effective for capturing general low-level visual features but lack domain-specific characteristics relevant to medical imagery, such as tongue images. To overcome the challenge of transferring features from natural datasets to specialized domains, domain-specific fine-tuning was performed using collected samples and synthetic data. Although the limited sample size may restrict the model’s ability to capture complex patterns in tongue imagery, the fine-tuned feature extractor achieved approximately 95% classification accuracy across 20 defined groups, demonstrating effective adaptation. This integration, coupled with the fusion of ensemble learning algorithms, harnesses the complementary strengths of clinical and image-derived features, effectively addressing the limitations of each modality in isolation [37]. Consequently, the model gains enhanced discriminative power, leading to more accurate predictions of ARTI.

The analysis of feature importance in machine learning models (GBM, XGBoost, and random forest) revealed that a red tongue coat is most indicative of HAdVs and control group, while the green color is more suggestive of COVID-19 cases. Furthermore, statistical analysis of the tongue images of the 3 groups showed that the COVID-19 group had a higher moisture level and lower contrast rate, suggesting that the patients with COVID-19 are more likely to have a thicker and greasier tongue coating compared to the other 2 groups. Our findings align with the TCM theory. According to TCM, COVID-19 is classified as a “cold-dampness” disease, characterized by a tongue that is pale red or dark and a thick, white greasy coating. HAdVs are considered “warm” diseases in TCM, typically presenting with red and dry tongues. The thickness and dryness of the tongue coating can also indicate the severity of the disease and the extent of fluid damage. Epidemiological studies conducted in China, the United Kingdom, and Ukraine have yielded analogous findings; the predominant tongue colors observed were pale pink and dark red, with the most frequently encountered tongue coating being thin and greasy [38-40]. Therefore, a patient presenting with acute respiratory symptoms and a red tongue devoid of a thick, greasy coating is more likely to have an HAdVs infection. Such observations could guide health care professionals in suggesting additional diagnostic tests and preventive measures, like isolation or improved hygiene, to avert transmission to susceptible household members. Considering its noninvasive nature and rapid assessment capability, AI-assisted tongue image analysis could be particularly valuable as a screening or triage tool in primary care clinics and emergency departments, where timely differentiation of respiratory infections is essential. The advancement of large language models and multimodal AI has facilitated the development of intelligent diagnostic systems that enhance tongue image interpretation and provide real-time decision support. These technologies not only assist clinicians in prioritizing patients for confirmatory testing and early intervention, optimizing health care resource allocation but also enable remote patient monitoring and telemedicine applications.

Hypothetically, the SARS-CoV-2 virus may induce alterations in the expression levels of genes coding for apoptosis and necroptosis of epithelial cells [41,42], resulting in the accumulation of oral epithelial cells and increasing tongue coating thickness. However, these proposed mechanisms remain preliminary and require further experimental validation to clarify their roles in oral manifestations of COVID-19. Supporting the clinical relevance of tongue changes, Wang et al [43] found a correlation between tongue coating thickness in patients with COVID-19 and levels of white blood cells as well as the neutrophil-to-lymphocyte ratio. Conversely, in patients with fever who tested negative for COVID-19, the presence of slimy or greasy tongue fur was associated with the level of C-reactive protein [43]. In addition, studies have indicated that greasy tongue fur is associated with higher blood fibrinogen levels in patients with stroke and with increased activity of glossal epithelial cells and vascular permeability in rodent models [44,45].

It is important to acknowledge several limitations of this study that may impact the generalizability of its findings. Although our predictive models demonstrate the potential of tongue image features for the differential diagnosis of acute respiratory tract infections, their accuracy remains modest, with area under the precision-recall curve (AUPRC) values below 80% across all models (Multimedia Appendix 7). Visualization of the data using t-SNE projected the 3 groups into generally distinct but overlapping clusters in 2D feature space, with considerable overlap, especially between the COVID-19 and adenovirus groups. This overlap is further reflected in our confusion matrix analyses, which consistently show misclassifications between these 2 groups across all classification models. Such overlap between the COVID-19 and adenovirus groups observed in the t-SNE projection may be partly attributable to the similarity of tongue image features shared by these 2 patient populations. In addition, the dimensionality reduction inherent in t-SNE can lead to loss of critical discriminative information, further exacerbating this overlap. Together with the limited sample size used for model development and validation, as well as the exclusion of additional clinical information, such as patient symptoms and other relevant features, these factors collectively contribute to the modest accuracy of our predictive models. Furthermore, the COVID-19 pandemic’s operational constraints necessitated data collection at a single tertiary hospital. Nevertheless, as a national referral center serving Beijing and surrounding provinces, our institution treats patients from geographically diverse regions, which may partially mitigate concerns regarding population representativeness. Future studies should incorporate larger and more diverse cohorts, combined with the integration of comprehensive clinical data, which are warranted to improve the discriminative power and generalizability of the models.