Predicting Melanoma Impact on the Swedish Healthcare System from the Adult Population Using Machine Learning on Registry Data

Abstract

Melanoma incidence has increased in Western countries over the past 50 years, leading to significant healthcare costs. In Sweden, comprehensive healthcare registries enable large-scale prediction studies using machine learning. Several machine learning models were evaluated to predict melanoma diagnoses using Swedish registry data, assessing the added value of diagnostic and medication data beyond demographics. The study included all adults in Sweden with continuous residency for 9.5 years (n = 6,036,186). The outcome was a melanoma diagnosis, including melanoma in situ, recorded within 5 years after the index date (31 December 2014). Predictors included age, sex, income, education, marital status, region of birth, diagnoses, and dispensed drugs. Models tested were logistic regression, gradient boosting, random forests, and a neural network. A total of 38,582 individuals (0.64%) developed melanoma. The gradient boosting model using all predictors performed best, with an area under the receiving operating characteristics curve (AUC) of 0.735 (95% confidence interval [CI], 0.725–0.746). When diagnosis and medication data were excluded, AUC dropped to 0.681 (95% CI: 0.670–0.691). The findings highlight that including healthcare codes improves predictive performance, and demonstrate the utility of Swedish registries for computational phenotyping. This approach may support early detection of melanoma and targeted follow-up.