AI Flagged About Three in Four Diabetic Foot Ulcers Unhealed at 12 Weeks Using First Visit Thermal Images

The prototype correctly classified just 53% of ulcers that healed

foot surgery
Researchers used thermal images from the first clinic visit to predict wound healing at 12 week Viktors Duks/Pexels

An AI prototype using first-visit thermal images flagged an average of 75.8% of diabetic foot ulcers classified as unhealed at 12 weeks. In the same tests, it correctly classified only 53% of the ulcers that healed. The Journal of Diabetes Science and Technology reported both figures.

For a patient whose wound is slow to close, an early flag could mean a clinician takes a closer look sooner. But a flag could also send a patient for an extra review even if the ulcer would heal with existing care. The study tested the predictions, not what happened when clinicians acted on them.

A First-Visit Image Captures Heat Patterns Across the Wound

TherMUL pairs a smartphone with a clip-on infrared camera. At the initial assessment, researchers held the camera about 20 cm from the foot. They did not need to touch the wound.

A colour image helped researchers locate the ulcer. Thermal imaging showed how temperatures varied across it. The model used the pattern of warmer and cooler areas rather than the average temperature of the wound.

Clinicians assessed wound healing independently at week 12. An ulcer counted as healed only if new skin covered the entire wound and no area remained open. Researchers developed and tested the model after those outcomes were known.

The predictions drew on images from the first visit, not later ones. The protocol also included images at weeks two and 12 to document how the wounds progressed.

A Separate Seven-Ulcer Check Shows Both Kinds of Error

The development dataset covered 68 ulcers from 56 participants at two specialist foot services in Melbourne. By week 12, 44 ulcers had healed and 24 had not.

Researchers repeatedly divided the dataset into groups, fitting the model on one portion and testing it on the rest. The headline figure is an average across those tests, not a count out of 24.

The researchers then ran a separate check on seven ulcers whose images the model had not seen. Five predictions matched the clinical outcome. Of the other two, the model missed one ulcer that stayed open and flagged one that healed. Seven cases cannot establish how often those errors would occur in another clinic.

An Earlier Flag Has Yet to Prove Its Value in Care

One benchmark clinicians use is how much a wound shrinks over the first four weeks of care. A thermal image is available before that comparison can be made. But the study did not establish whether the heat pattern tells clinicians something they would not already pick up at the first visit.

The study focused on ulcers with little or no infection or restriction in blood flow, so it says less about more severe cases. Room temperature, humidity and airflow can affect thermal readings. The study did not account for those factors in its analysis.

A clinic would also need to know what happens when the model gives no warning. If staff treat that result as an all-clear, a wound that stays open could get less attention than it needs.

The practical test is whether the flag brings forward a review for a wound that needs one and would otherwise have waited. A clinic would have to weigh those earlier reviews against the extra appointments for patients whose wounds would heal without them.