Leveraging machine learning to predict de novo skin malignancy following lung transplantation

Abstract Aims This study aimed to predict post-transplant malignancy risks at multiple levels among lung transplant recipients using machine learning (ML) and to identify key clinical and immunogenetic predictors. Materials and methods A dataset of 30,917 lung transplant recipients with no prior cancer history was analyzed using pre-, peri-, and post-transplant variables. Multiple ML algorithms—gradient boosting, random forest, neural networks, and logistic regression—were applied to predict: (1) overall de novo malignancies (DNM), (2) skin versus non-skin cancers, and (3) skin cancer subtypes, including basal cell carcinoma (BCC) and squamous cell carcinoma (SCC). Results Gradient boosting achieved the highest AUC for overall malignancies (0.746) and skin versus non-skin cancers (0.642), while random forest performed best for BCC versus SCC classification (AUC = 0.726). Significant predictors included HLA-DR alleles (DR52, DR1, DR53), A locus mismatch, recipient ethnicity, BMI, serum albumin, CMV/EBV serostatus, and cardiac-related measures (LV remodeling, cardiac output, prior cardiac surgery). Additional subtype predictors included peak PRA Class I sensitization, insulin signaling, donor-derived transfusions, and waiting list duration. Conclusions ML-driven predictive modeling enables personalized assessment of post-transplant malignancy risk, supporting early detection, targeted surveillance, and optimized long-term care for lung transplant recipients.