AI Identifies Coronary Heart Disease from Tongue and Face Photos with 98% Accuracy in Chinese Study

The article discusses a Chinese study published in Scientific Reports in which an AI system trained on photographs of participants’ faces and tongues correctly identified coronary heart disease CHD) in a small test. The study enrolled 518 individuals, including 200 CHD patients and 318 healthy volunteers. Each person provided a facial image and a tongue image captured with a purpose-built scanning system. The deep learning model analyzed tongue and face data separately before combining the results to predict CHD status. In a test of 104 individuals, the best version achieved 98% accuracy, correctly classifying all 40 CHD cases but generating two false positives. The study notes AI focused on tongue center, coatings, mouth skin, and cheeks, with links to traditional Chinese medicine concepts and cautions about generalizability. Original publisher: Scientific Reports.

• Tongue and face data combined improve CHD prediction over single modalities
• Potential for low-cost pre-screening, especially where access to specialists is limited
• Not yet validated in diverse populations or settings; associations may reflect confounding factors
• Further research required to compare with existing risk assessments and to test longitudinal outcomes

Overview

The article details a study from China in Scientific Reports where researchers trained a deep learning AI system to predict coronary heart disease CHD) using photographs of participants’ faces and tongues. The dataset consisted of 518 individuals, with 200 CHD patients and 318 healthy volunteers. Each participant provided two image types: a face photograph and a tongue photograph, both captured with a scanning system designed for this purpose. The AI analyzed tongue and facial images separately and then fused the information to predict CHD status. The researchers report that using both data sources together outperformed either source alone, with the top model achieving 98% accuracy in a 104-person test (40 CHD cases, 64 non-CHD), correctly identifying all CHD patients but producing two false positives.

Study design and methods

The study relied on a purposeful data collection protocol in which photographs were taken under controlled lighting using specialized equipment. The tongue analysis focused on features such as the center of the tongue, surface coating, and color, while facial analysis looked at skin around the mouth, cheeks, and pigmentation patterns. A deep learning model processed each image type and then integrated the results to classify CHD presence. The researchers emphasize that the Tongue and Face modalities contributed complementary information to the prediction task, and that combining them yielded better performance than either modality alone.

Key findings and interpretation

At the best performance level, the model achieved 98% accuracy in a test set of 104 individuals, correctly identifying all 40 CHD patients but misclassifying two healthy participants as CHD-positive, resulting in false positives. The AI pinpointed specific tongue features (centre, coating) and facial cues (mouth area, cheeks) that it used for predictions. The researchers note that these areas align with traditional Chinese medicine concepts that associate tongue and facial features with heart health. They also acknowledge that this predictive relationship does not establish causation and could reflect other demographic or health-related differences between groups, such as age or comorbid conditions.

Context and limitations

Experts caution that reading disease from appearance is not unique to traditional Chinese medicine. The study compares patients with CHD to healthy volunteers rather than performing population-wide screening, and the data were collected in a controlled, hospital-like setting. This raises questions about generalizability to broader populations, different hospitals, ethnic groups, and real-world healthcare environments. The model has not been tested across diverse clinical settings, and some of the differences it detects could be confounded by age or other illnesses. Furthermore, even with high accuracy, screening tools can generate substantial false positives in populations where CHD prevalence is relatively low. The authors call for larger, longitudinal studies to assess whether this approach can detect undiagnosed disease, differentiate CHD from other conditions, and complement existing risk assessments such as lipid profiles and blood pressure measurements.

Clinical implications and future directions

If validated in larger and more diverse cohorts, a quick, painless photographic test could help triage individuals who need further CHD investigations such as ECGs, blood tests, or coronary imaging, potentially improving access in resource-limited communities. The study highlights the possibility of integrating AI-assisted appearance-based screening with conventional risk checks, cholesterol testing, and blood pressure monitoring. However, robust external validation and direct comparisons with established risk assessment tools are necessary before clinical adoption. The researchers also stress that AI outputs should be interpreted with caution, given the potential influence of non-disease factors and the need for transparency and explainability in AI-driven healthcare tools.

Conclusion

The reported findings illustrate a compelling demonstration of AI-augmented imaging in CHD risk assessment, showing that tongue and face data can jointly signal cardiovascular disease in a controlled study. Yet the authors and commentators rightly emphasize the preliminary nature of the results and the substantial work required to determine real-world utility, generalizability, and impact on health outcomes. Further research will determine whether this approach can become a scalable, low-cost screening adjunct to conventional CHD risk checks.