Abstract:
Aim: The aim of this study is to develop a clinically reliable prediction model for skin graft failure in skin cancer patients undergoing full-thickness skin graft (FTSG) or split-thickness skin graft (SSG). The primary objective is to identify key predictors of graft failure and evaluate the performance of different machine learning models.
Methods: Data was retrospectively collected from 104 skin cancer patients who underwent either FTSG or SSG between 1st July and 30th September 2024. Six key predictors were identified using Recursive Feature Elimination (RFE) with cross-validation: BMI, comorbidities (e.g., diabetes), graft longest dimension, donor site, postoperative day of first graft failure sign, and dressing used. The dataset was balanced using Synthetic Minority Over-sampling Technique (SMOTE), and three models: Random Forest (RF), Logistic Regression (LR), and XGBoost were trained to predict graft failure, with a focus on sensitivity (recall).
Results: RF achieved the highest sensitivity (recall = 0.88), with an overall accuracy of 95% and the highest cross-validated F1-score (0.944 ± 0.046). LR also showed high sensitivity (recall = 0.88) but with lower precision and accuracy (86%). XGBoost matched RF in sensitivity and demonstrated strong precision (0.88) and accuracy (95%), with a cross-validated F1-score of 0.929 ± 0.051.
Conclusion: With high sensitivity and accuracy, the RF and XGBoost models showed excellent predictive performance. These models can help identify at-risk patients, allowing for peri-operative optimisation and consideration of alternative treatments, such as radiotherapy. The later alleviates patient burden, including time and anxiety associated with repeat procedures, and optimise resource use.
