AI burn depth assessment shows variable sensitivity and specificity across multiple imaging modalitiesTrial Shows AI May Help Assess Burn Depth Accuracy

BACKGROUND: AI-based burn depth assessment is rapidly emerging, yet evidence for diagnostic accuracy, generalizability, and deployment readiness remains unclear. METHODS: We performed a PRISMA-DTA-aligned systematic review and meta-analysis. We synthesized accuracy outcomes and extracted deployment-relevant features, including validation design, reference-standard family, analytic unit, and subgroup reporting. Six studies (4,897 observations; mixed image-, wound-, and patient-level units) were included; four studies (N = 2541) formed the prespecified primary DTA dataset. We planned bivariate random-effects models where feasible and conducted exploratory subgroup analyses by imaging modality, skin tone, and age. Two additional studies were included only in exploratory analyses after prespecified approximate 2 × 2 reconstruction. RESULTS: In the primary dataset, the prespecified bivariate model did not converge because of sparse/extreme 2 × 2 patterns and limited study numbers; therefore, no pooled sensitivity or specificity was generated. Study-level sensitivity ranged from 0.50 to 1.00 and specificity from 0.54 to 0.97. In the expanded exploratory dataset (6 studies), pooled sensitivity was 0.924 (95% CI 0.788-0.975) and specificity 0.877 (95% CI 0.701-0.956), but these estimates are hypothesis-generating because they rely partly on approximately reconstructed 2 × 2 data. Exploratory descriptive analyses suggested possible modality-related variation and lower specificity in darker skin, although subgroup evidence was sparse and non-confirmatory; pediatric evidence was limited to a single within-study stratum. CONCLUSIONS: The evidence base is insufficient for deployment-ready use of AI burn depth assessment. The primary dataset did not support hierarchical pooling, and exploratory pooled estimates should be interpreted cautiously because they rely on reconstructed 2 × 2 data and mixed analytic units. More importantly, this review clarifies the evidence gaps separating promising algorithmic performance from deployable clinical decision support. Future studies should prioritize standardized reference standards, patient-level external validation in independent institutions, and reporting that supports safe integration into clinical workflows.