Abstract
Background: Skin graft failure is a significant complication in burn care that prolongs recovery, compromises surgical outcomes, and necessitates repeat operations. Accurate risk assessment is challenging due to a complex interplay of patient and injury factors, underscoring the need for objective, data-driven predictive tools. We developed and validated an interpretable machine learning (ML)-driven tool to provide individualized, data-driven risk assessment for skin graft loss in patients with burn injuries. Methods: A cohort study of patients who had autologous skin grafting from the American Burn Association Burn Care Quality Platform was conducted from 2013 to 2022. A Random Forest ML model was developed and validated on independent training (70%), validation (15%), and testing (15%) sets. Model performance was assessed using the area under the receiver operating characteristic curve (AUC-ROC). Results: Of 286 479 patients during the study period, 24 046 underwent autologous skin grafting. Machine learning demonstrated discriminatory predictive performance with AUC-ROC of 0.794. Burns ≥ 40% total body burn surface area (odds ratio [OR] 8.16; P < .001), history of alcohol abuse (OR 1.86; P < .001), active tobacco use (OR 1.45; P = .004), age > 60 years (OR 1.70; P = .013), hypertension (OR 1.52; P = .003), male sex (OR 0.74; P = .009), diabetes mellitus (OR 1.45; P = .030), and contact burns (OR 1.70; P = .006) emerged as independent predictors of graft loss. The model was translated into an interactive, interpretable web-based clinical tool. Conclusions: The Skin Graft Loss Risk Estimator provides a transparent, interpretable and personalized risk assessment of skin graft loss to enhance clinical judgment, facilitate effective patient counseling and surgical planning in burn care.

This publication has 34 references indexed in Scilit: