Multimodal AI models assessing burn photographs, from detection to treatment recommendation

Source article: Reliable Detection, Unreliable Clinical Decision-making: Artificial Intelligence Models in Burn Assessment

Background: Multimodal artificial intelligence (AI) systems are increasingly used for visual interpretation tasks. Whether commercially available models can be applied to burn surgery, and to what extent, remains unclear. We evaluated a clinical chain from visual recognition (distinguishing burns from nonburned skin) to treatment decision-making (debridement versus immediate skin grafting), including accuracy, output consistency, and skin tone subgroup performance. Methods: Four multimodal large language models…

By the Truvace deskOctober 15, 20262 min read1 sourcePeer-reviewedTRV-2026-1394· cite & verify

Reliable Detection, Unreliable Clinical Decision-making: Artificial Intelligence Models in Burn Assessment — Anindya Islam Amit, State Minister, visits burn injury victims, Ghera power plant explosion, Dhaka, 2026-05-14 by Press Information Department. Public domain

Trace impact readingNegative state

P 73

Both sides are scored from claims and sources, not community votes.

G 68

In brief

Researchers tested four commercially available multimodal language models on 50 cropped clinical photographs, 25 showing burns and 25 showing nonburned skin, repeating each assessment five times for 4000 assessments. Models had to say whether a burn was present and then recommend debridement or immediate skin grafting, judged against a consensus of plastic surgeons.

The pattern suggests that a model that looks competent at seeing may still be unsafe at deciding, so strong recognition scores should not be read as readiness for treatment guidance. The study used a small set of cropped images and named model versions, so how far the findings carry to real clinical workflows remains open. The authors see detection as a possible base for future triage support.

Main points

Burn detection accuracy ranged from 97.2% to 100% across the four models, with sensitivity of 94.4%-100% and specificity of 100%.
Detection performed equivalently across Fitzpatrick skin tone strata.
Treatment accuracy in burn cases ranged from 24.8% to 72.8%, and three models fell below the majority-class baseline.
Agreement on treatment decisions fell to κ = 0.132-0.385 in three models, versus κ = 0.896-1.000 for detection.
Error patterns were systematic and opposed across models.

The gain

Four multimodal models distinguished burns from nonburned skin with 97.2%-100% accuracy, with equivalent performance across Fitzpatrick skin tone strata.

The problem

The same models gave unreliable treatment recommendations for burns (debridement versus immediate skin grafting), with accuracy of 24.8% to 72.8% and inconsistent outputs for identical images.

The rundown

The evaluation covered GPT-5.4 Pro, Grok 4.1, Gemini 3.1 Pro and Claude Opus 4.6. Each model assessed the same 50 images across five independent runs, and plastic surgeons' expert consensus set the ground truth. Reproducibility was measured with Fleiss κ.

Visual recognition was both accurate and consistent, with κ between 0.896 and 1.000. Treatment decisions were far less stable: in three models the κ dropped to 0.132-0.385, meaning identical images often drew different recommendations.

The authors describe the clinical chain as breaking at the transition from perception to decision, which limits current applicability in burn surgery. They suggest that reliable visual recognition may be a foundation for future AI-assisted triage in burn care.

What this doesn’t fix

The study tested a small set of 50 cropped photographs and concluded that current applicability in burn surgery is limited, so results do not establish performance in real-world clinical decision-making.