Diabetes is caused by the body’s inability to produce or properly use insulin, and if it is caught late it leads to serious complications. In the Philippines, the number of adults with diabetes is projected to reach 5,456,300 by 2030 and 7,503,900 by 2045. The standard screening test, fasting plasma glucose, needs a blood draw, which can mean bleeding, dizziness, and a small risk of infection, and its cost and the distance to a clinic discourage regular testing.
A paper by Tricia Z. Pulmano, Lester Niel S. J. Salvador, Maria Leonora C. Guico, and Jan Kevin A. Galicia, presented at the IEEE Region 10 Conference (TENCON 2024) in Singapore in December 2024, asks whether a phone camera could do a first pass instead.
The approach
The idea builds on earlier work by Garcia et al. showing that the texture of skin in specific facial regions changes with type 2 diabetes. The app takes a photo, detects 68 facial landmarks, and extracts three blocks: the forehead, the areas around the eyes, and the bridge of the nose. Following a related system called the Computer-Assisted Non-invasive Diabetes Mellitus Detection System (C-NDMDS), it describes each block with local binary patterns and Gabor filters, then feeds those features to a support vector machine that decides between diabetic and non-diabetic.
The dataset was small: ten healthy individuals and ten with type 2 diabetes, all aged 21 to 40. The classifier was evaluated with five-fold cross-validation.
What they found
The Gabor filter analysis did show higher texture values in the diabetic subjects, in the 180 to 320 range, which suggests there is a measurable skin signal. Among the regions, the forehead contributed the most to classification, while the left cheek and nose contributed weakly or negatively.
The classifier itself was only moderate. Five-fold cross-validation gave 60 percent accuracy, 35 percent sensitivity, and 40 percent specificity, which the authors attribute to the small dataset and inconsistent photo conditions. On the engineering side, the Android app processes an image in 0.659 seconds, so the workflow of uploading a photo and getting a result is quick.
What comes next
The authors conclude that facial texture analysis with machine learning has potential as a non-invasive, accessible screening step, but not yet as a reliable one. They call for larger and more diverse datasets, standardized capture conditions (lighting, angle, resolution), and better feature extraction and models before the method could be trusted for early screening.