Researchers at Hainan Medical University, led by Yinhai Wang and Aimin Gong, published a study on September 18, 2026, in Scientific Reports detailing an AI model capable of screening for coronary heart disease using only facial and tongue images. The team photographed 518 individuals—200 patients previously diagnosed with heart disease and 318 healthy volunteers—under standardized lighting. The AI was tasked with predicting the presence of disease based solely on these images.1

Key Findings
  • The model analyzed specific regions of the face and tongue; the biological connection to heart health remains subject to further investigation.
  • A correlation between yellowish eyelid deposits and heart attack risk was previously confirmed by a 33-year study of over 12,000 people in Copenhagen.
  • When a similar concept was tested across nine hospitals six years ago, results were significantly less precise.
  • A false positive rate of 3 percent can lead to an unmanageable number of false alarms in large-scale population screenings.

Which facial and tongue features does the AI analyze?

Coronary heart disease typically begins with the buildup of fatty plaques that narrow the arteries supplying the heart muscle. Because many patients are unaware of their condition until they experience chest pain or a cardiac event, a quick and non-invasive screening tool is a long-sought goal in cardiology. The Hainan team utilized a dual-approach model: a neural network mapped the borders of the tongue, while Google’s open-source MediaPipe identified facial landmarks to align each image. When combined, the analysis of these features outperformed either the tongue or facial imagery alone.1 Heat maps indicated the model focused on the center and coating of the tongue, as well as the skin surrounding the lips and cheeks. While the researchers noted these areas mirror concepts in traditional Chinese medicine—where the tongue and face are viewed as reflections of cardiac health—the study did not investigate the underlying physiological mechanisms connecting these specific tissue textures to arterial narrowing.

What are the known facial indicators of heart disease?

The use of visual cues to detect illness is not unique to Eastern traditional practices. Cyanosis, or bluish discoloration of the lips, can signal low blood oxygen, while xanthelasma—yellowish lipid deposits on the eyelids—is often a marker of high cholesterol. A notable study, the Copenhagen City Heart Study, tracked this connection over 33 years. Among 12,745 participants who were free of vascular disease at the start of the study (1976-1978), 563 individuals, or 4.4 percent, displayed xanthelasma. By May 2009, researchers found that these individuals faced a 48 percent higher risk of heart attack, even after accounting for traditional risk factors like cholesterol and triglyceride levels.3 However, most of the 1,872 participants who suffered heart attacks during the follow-up did not have these visible deposits.

Xanthelasma on eyelids
Xanthelasma, yellowish lipid deposits on the inner corner of the eyelid, serves as a recognized warning sign for potential cholesterol issues. Photo: Klaus D. Peter, CC BY 3.0 DE, Wikimedia Commons

How reliable is AI diagnosis in a clinical environment?

The high accuracy observed in the Hainan study—where 40 of 40 patients were correctly identified in a test group—may be inflated by the distinct demographic differences between the "patient" and "volunteer" groups. Factors such as age, smoking status, medication, and lighting can influence appearance, and the model may have inadvertently learned these variables rather than the indicators of heart disease. Furthermore, the model has not been tested in diverse clinical settings. Previous studies have shown that AI can struggle with generalization; for instance, a 2018 study found that pneumonia-detection models trained on data from specific hospitals were better at identifying the hospital of origin than the actual disease.4 A more rigorous test occurred in 2020, involving 1,013 patients across nine hospitals. While the model correctly identified 80 percent of patients, it flagged many healthy individuals incorrectly, achieving an area under the curve (AUC) of 0.730, which still outperformed traditional risk scoring methods but fell short of the near-perfect results seen in controlled settings.5

Notes and Sources

  1. Wang, Y., et al. (2026). Explainable multimodal deep learning fusion of tongue and facial images for noninvasive coronary heart disease screening. Scientific Reports. doi.org/10.1038/s41598-026-71871-x
  2. Christoffersen, M., et al. (2011). Xanthelasmata, arcus corneae, and ischaemic vascular disease and death in general population: prospective cohort study. BMJ, 343, d5497. doi.org/10.1136/bmj.d5497
  3. Zech, J. R., et al. (2018). Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs. PLOS Medicine, 15(11), e1002683. doi.org/10.1371/journal.pmed.1002683
  4. Lin, S., et al. (2020). Feasibility of using deep learning to detect coronary artery disease based on facial photo. European Heart Journal, 41(46), 4400-4411. doi.org/10.1093/eurheartj/ehaa640