AI-based tongue imaging could help non-invasive detection of coronary heart disease

Coronary heart disease is the most common cause of death from disease worldwide. According to the World Health Organization, coronary heart disease is responsible for 17.9 million deaths per year worldwide, or nearly one-third of all deaths from disease each year.

Coronary angiography is currently the best method to confirm a diagnosis of coronary artery disease, but it is expensive and invasive, poses risks to patients, and is not suitable for early diagnosis and assessment of disease risk.

In search of a safer, cheaper and more effective diagnostic method, a research team from the School of Traditional Chinese Medicine, Beijing University of Chinese Medicine, the School of Life Sciences, Beijing University of Chinese Medicine and the School of Traditional Chinese Medicine, Hunan University of Chinese Medicine used artificial intelligence (AI) to develop a diagnostic algorithm based on tongue imaging. Their work is published in The Frontiers of Cardiovascular Medicine.

Recent research on the development of diagnostic models for coronary artery disease has focused on clinical risk factors as variables. Recent studies have also established that additional biological components such as pulse waves and facial appearance can play a considerable role in establishing a diagnosis of coronary artery disease.

However, while a patient’s symptoms and signs form the basis of a clinical diagnosis, researchers note that traditional Chinese medicine does things differently, albeit quite effectively.

How Traditional Chinese Medicine Sheds Light on This New Work

“Traditional Chinese medicine (TCM) uses unique and effective diagnostic strategies, including observing patients’ external conditions. TCM theory posits that ‘internal diseases manifest externally,’ allowing practitioners to assess the severity of diseases through observation,” the team writes in the new study.

The key to observation in TCM is tongue diagnosis, which includes assessing the color, coating, and shape of the tongue. The tongue, filled with nerves and blood vessels, functions as an integral part of the cardiovascular system, and its appearance can change as systemic diseases and conditions develop, especially those that affect blood circulation. Additionally, at least 14 studies conducted since 2019 have established that tongue observation is an effective means of diagnosing diseases.

With this knowledge, the team sought to determine whether tongue images could serve as a critical basis for non-invasively maximizing the diagnosis of CAD.

Development of a CAD diagnostic model

To build their machine learning framework, the team selected the ResNet-18 network, pre-trained on an ImageNet dataset. They built two CAD diagnostic frameworks: one based solely on CAD risk factors, and a second integrating deep features from tongue images with CAD risk factors. The algorithm they selected for tongue feature extraction, based on the Deeplab V3+ framework, demonstrated an overall accuracy of over 99%.

The team evaluated several machine learning algorithms and ultimately selected XGBoost, which performed best on a classification task involving “deep features of tongue images” and risk factors.

Furthermore, they explain that when they compared machine learning algorithms – some programmed with only risk factors and others with risk factors and tongue image features – they determined that “the inclusion of tongue image features significantly improved the performance of the algorithm, indicating that adding tongue features as input variables contributes positively to the optimization of the algorithm.”

Between March 2019 and November 2022, the researchers recruited hypertensive patients aged 18 to 85 from four different hospitals, ultimately compiling a study cohort of 244 patients with hypertension and an additional 166 patients with hypertension associated with coronary artery disease.

Results and limitations

Among the notable results, the team’s coronary artery disease diagnostic algorithm performed particularly well in subjects aged 65 and older, and yielded similar results for men and women, demonstrating good generalizability. It also showed greater accuracy of judgment in cases with three or more risk factors, “underscoring the importance of considering multiple risk factors in the diagnosis of coronary artery disease,” the researchers write.

Limitations of the study include the lack of patients from different countries and ethnicities, a small sample based only on people with hypertension, and the use of only one type of equipment to collect tongue images, which limits the applicability of the diagnostic model with alternative collection devices and in varying situations.

The researchers suggest that future work should include a larger and more comprehensive study population to validate and optimize their diagnostic model. They also note that incorporating additional biomarkers into tongue images could allow for the creation of a broader model. However, this research serves as a useful foundation for next steps.

“Our work opens a new perspective, suggesting that tongue images have applicable diagnostic value for the diagnosis of coronary heart disease,” the researchers conclude. “The characteristics of tongue images could become new indicators of coronary heart disease risk, demonstrating the feasibility of integrating traditional Chinese medicine theories with modern technology.”

© 2024 Science X Network