674 Machine Learning and Automation in Burn Care: A Systematic Review

Abstract

Abstract Introduction Current methods of burn evaluation and treatment are subjective and dependent on surgeon experience, with high rates of inter-rater variability leading to inaccurate diagnoses and treatment. Machine learning (ML) and automated methods are being used to develop more objective and accurate methods for burn diagnosis and triage. Defined as a subfield of artificial intelligence that applies algorithms capable of knowledge acquisition, machine learning draws patterns from data, which it can then apply to clinically relevant tasks. This technology has the potential to improve burn management by quantitating diagnoses, improving diagnostic accuracy, and increasing access to burn care. The aim of this systematic review is to summarize the literature regarding machine learning and automated methods for burn wound evaluation and treatment. Methods A systematic review of articles available on PubMed and MEDLINE (OVID) was performed. Keywords used in the search process included burns, machine learning, deep learning, burn classification technology, and mobile applications. Reviews, case reports, and opinion papers were excluded. Data were extracted on study design, study objectives, study models, devices used to capture data, machine learning, or automated software used, expertise level and number of evaluators, and ML accuracy of burn wound evaluation. Results The search identified 592 unique titles. After screening, 35 relevant articles were identified for systematic review. Nine studies used machine learning and automated software to estimate percent total body surface area (%TBSA) burned, 4 calculated fluid requirements, 18 estimated burn depth, 5 estimated need for surgery, 6 predicted mortality, and 2 evaluated scarring in burn patients. Devices used to estimate %TBSA burned showed an accuracy comparable to or better than traditional methods. Burn depth estimation sensitivities resulted in unweighted means >81%, which increased to >83% with equal weighting applied. Mortality prediction sensitivity had an unweighted mean of 96.75%, which increased to 99.35% with equal weighting. Conclusions Machine learning and automated technology are promising tools that provide objective and accurate measures of evaluating burn wounds. Existing methods address the key steps in burn care management; however, existing data reporting on their robustness remain in the early stages. Further resources should be dedicated to leveraging this technology to improve outcomes in burn care.