Background:
Tumour staging, informed by Breslow depth and ulceration, incompletely explains outcome variability in primary cutaneous melanoma. Patient-specific factors have been broadly associated with prognosis but are not routinely considered in clinical risk stratification.
Objectives:
Firstly, to characterize the prognostic associations of 48 tumour- and patient-specific factors, collected at time of diagnosis, with 7-year recurrence in primary cutaneous melanoma. Secondly, to compare the performance of ensemble gradient boosting (XGBoost) and regularized logistic regression at integrating these predictors to discriminate recurrence risk in a blinded testing set.
Methods:
This prospective cohort study recruited from outpatient dermatology, surgery, and pathology services in Queensland, Australia (2010-2014). Overall, 630 participants with T1b-T4b primary cutaneous melanoma had outcome data at a fixed follow-up of 7 years post-diagnosis. Participants were split by year of diagnosis into training (2010-2012; n = 303) and testing sets (2013-2014; n = 327). XGBoost and ridge-regularized logistic regression models were fitted to baseline predictors and trained to predict 7-year recurrence, with binary thresholds developed in the training set. Performance metrics in the testing set were calculated using binary thresholds derived in the training set. Comparator models were trained on tumour-specific factors only.
Results:
Tumour-only logistic regression displayed 58.1% sensitivity and 71.8% specificity at predicting 7-year recurrence in the testing set, despite ridge regularization. By including both tumour- and patient-specific factors, XGBoost displayed 66.7% sensitivity and 82.1% specificity, corresponding to a comparative 21% and 37% reduction in false positives and negatives respectively. Discrimination performance generalized to prediction of 7-year distant metastasis and melanoma-specific mortality; and was retained in a testing set subgroup absent sentinel lymph node biopsy (n = 213).
Conclusion:
Patient-specific factors contribute prognostic information beyond tumour-specific staging. Ensemble machine learning is more amenable to interrelated covariates than conventional regression and may better integrate patient-specific factors to stratify melanoma risk beyond existing prognostic criteria.