Artificial intelligence could make screening for hypertension and diabetes faster and more accessible by analyzing short facial videos. Researchers in Japan have developed a machine learning algorithm capable of detecting both conditions from recordings lasting as little as five seconds. The findings will be presented at ESC Congress 2026 in Munich and published in the European Heart Journal.
Hypertension and diabetes are major modifiable risk factors for cardiovascular disease, yet many people remain undiagnosed. Conventional screening usually requires blood pressure measurements, blood tests or dedicated healthcare visits. Researchers from the University of Tokyo and Institute of Science Tokyo investigated whether contactless video analysis could provide an alternative approach for large scale screening.
Detecting disease without physical contact
The prospective, single center study included 215 participants, consisting of diagnosed patients and healthy volunteers. Researchers recorded short, high speed videos of participants’ faces and palms using a spectroscopic camera. A machine learning algorithm analyzed characteristics that cannot simply be assessed by looking at the video. These included pulse wave dynamics, which provide information about arterial stiffness, patterns in blood flow through the skin and spectral characteristics associated with skin coloration. Participants also underwent conventional assessments for hypertension and diabetes, allowing researchers to compare the AI results with established diagnostic methods.
Previous research with the technology showed that a 30 second recording of the face and palms could identify hypertension with 95 percent accuracy. Sensitivity was 100 percent for normal blood pressure and 89.2 percent for hypertension. When the recording was shortened to just five seconds, accuracy remained 90.3 percent.
Diabetes detected through facial blood flow
The latest research extended the approach to diabetes. By analyzing facial blood flow patterns, the algorithm identified diabetes with an accuracy of 88.2 percent from a 30 second video. With only five seconds of footage, accuracy was 81.2 percent. The researchers also investigated whether the system could estimate blood pressure without using a conventional cuff. Based solely on facial video, the mean absolute percentage error for systolic blood pressure was 8.6 percent.
The average error was minus 2.6 mmHg, which fell within the Association for the Advancement of Medical Instrumentation limit of plus or minus 5 mmHg. However, variability remained too high: the standard deviation was plus or minus 12 mmHg, compared with the AAMI criterion of plus or minus 8 mmHg. The researchers therefore emphasize that further development is necessary before the technology could be considered for routine screening. Larger multicenter datasets and further optimization of the algorithm should help improve the consistency of blood pressure estimates.
Large scale screening
According to researcher Ryoko Uchida, the goal is to develop contactless screening that can be used in everyday environments. If validated in larger and more diverse populations, the technology could potentially identify people at increased risk without blood sampling, a blood pressure cuff or a dedicated clinic appointment. Such an approach could be particularly relevant given the scale of hypertension and diabetes worldwide. An estimated 1.4 billion adults aged 30 to 79 have hypertension, while approximately 589 million people live with diabetes.
Nico Bruining, program co-chair of the ESC Digital and AI Summit and editor-in-chief of the European Heart Journal – Digital Health, sees potential for screening beyond hospitals. Because video based assessment is quick and contactless, it could eventually reach people who are not routinely screened. The technology is not yet ready to replace conventional diagnostic methods. The researchers first plan to validate the algorithm in larger cohorts representing more diverse populations. If those studies confirm the current findings, facial video analysis could eventually become a low threshold first step for identifying people who require further medical assessment.
AI based diabetes research
In February, an international research consortium led by partners of the German Center for Diabetes Research developed an AI based method to identify subtle pancreatic tissue changes associated with type 2 diabetes. Researchers combined high resolution gigapixel imaging, deep learning and explainable AI to detect morphological patterns that are difficult to recognize with conventional microscopy.
The AI models accurately distinguished pancreatic tissue from people with and without type 2 diabetes based solely on histological characteristics. Explainable AI also revealed which structures contributed most strongly to these predictions. Potential new morphological biomarkers included changes in the islets of Langerhans and α cells, alterations in neuronal axons and the proximity of adipocyte clusters to pancreatic islets.
The study provided new insights into pancreatic changes linked to beta cell dysfunction and impaired insulin regulation. The approach could support further research into disease progression and, eventually, contribute to more precise diagnostics and targeted interventions.
Add ICT&health on Google
Show more content from ICT&health in Google Search.
