Design of a Machine Learning System for Prediction of Chronic Wound Management Decisions

Chronic wounds affect 6.5 million Americans, are complex conditions to manage and cost $28–$32 billion annually. Although digital solutions exist for non-expert clinicians to accurately segment tissues, analyze affected tissues or efficiently document their wound assessment results, there exists a lack of decision support for non-expert clinicians who usually provide most wound assessments and care decisions at the point of care (POC). We designed a machine learning (ML) system that can accurately predict wound care decisions based on labeled wound image data. The care decisions we predict are based on guidelines for standard wound care and are labeled as: continue the treatment, request a change in treatment, or refer patient to a specialist. In this paper, we demonstrate how our final ML solution using XGboost (XGB) algorithm achieved on average an overall performance of F-1 = .782 using labels given by an expert and a novice decision maker. The key contribution of our research lies in the ability of the ML artifact to use only those wound features (predictors) that require less expertise for novice users when examining wounds to make standard of care decisions (predictions).