Using machine learning to predict short-term mortality in Icelandic skin cancer patients: A nationwide retrospective cohort analysis

Summary: Abstract Body: Early detection and risk stratification remain key challenges in skin cancer mortality. Recent advances in machine learning (ML) have led to success in predicting cancer mortality, including melanoma and Merkel cell carcinoma. Our research aims to similarly develop a model to screen for 5-year mortality in Icelandic patients with skin cancer. The Icelandic Cancer Registry, a nationwide comprehensive database with no loss to follow-up given its design, was used to identify all patients with cutaneous malignancies according to ICD-O-3 codes (n=18300, 1949-2023). An XGBoost-based ML model was developed to predict 5-year mortality in our cohort. The dataset included patient demographics, tumor characteristics, and family history. Preprocessing involved imputation, normalization, and one-hot encoding. A 50:50 class-weighted approach for class imbalance, 4:1 train-test ratio, hyperparameter tuning using grid search, and 5-fold cross-validation were utilized. The ML model achieved an overall accuracy of 98.4% with a ROC-AUC of 0.984. For the majority class (low-risk cases), the precision, recall, and F1 scores were 99.9%, 98.5%, and 99.2%, respectively. Performance for the mortality class yielded a precision of 26.4%, recall of 82.6%, and F1-score of 40%. The balanced accuracy was 90.6%, and specificity reached 98.6%. Feature importance analysis identified Breslow thickness as the most significant predictor, followed by tumor subtype and staging variables. Loss curves showed the absence of overfitting for strong generalization across datasets. The ML model showed excellent discriminatory ability in predicting mortality. To maintain a high recall, lower precision was justified by the low-harm clinical action of recommending increased screening for flagged patients. With minimal cost or risk, the screening tool could utilize health record data to save lives with earlier interventions for high-risk cases and ensure fewer patients with skin cancer go undetected. Dany Alkurdi¹, Lara Shqair¹, Saniya Tariq¹, Omar Alani¹, Jonas Adalsteinsson¹ 1. Icahn School of Medicine at Mount Sinai, New York, NY, United States. Pigmentation, Melanoma, and Melanoma Immune Surveillance