Assessment of machine-learning models tested on Swedish registry data enabled more accurate melanoma diagnosis prediction, with added health-care code, age, sex, and medication information for improved performance, according to the results of a study published in Acta Dermato-Venereologica.
“Our study shows that data which is already available within health-care systems can be used to identify individuals at higher risk of melanoma,” stated Martin Gillstedt, MSc, a doctoral student at the University of Gothenburg’s Sahlgrenska Academy and a statistician in the Department of Dermatology and Venereology at Sahlgrenska University Hospital, Sweden. “This is not a form of decision support that is currently available in routine health care, but our results give a clear signal that registry data can be used more strategically in the future.”
Study Methods
Researchers used registry data to evaluate some machine-learning models' ability to predict melanoma diagnoses. They wanted to explore added performance value for diagnostic and medication data, on top of demographic data, for the artificial intelligence models.
The study included all adults in Sweden who had continuously resided in the country for 9.5 years (n = 6,036,186), with a focus on melanoma diagnosis within 5 years of the index date of December 31, 2014.
Predictive factors included age, sex, income, education, marital status, region of birth, diagnoses, and dispensed drugs.
The machine-learning techniques examined included logistic regression, gradient boosting, random forests, and a neural network.
Key Findings
A total of 38,582 individuals developed melanoma (0.64%).
Of all tested models, the gradient boosting model using all predictive factors performed best and demonstrated an area under the receiving operating characteristics curve (AUC) of 0.735 (95% confidence interval [CI] = 0.725–0.746). Without diagnostic and medication data, the AUC dropped to 0.681 (95% CI = 0.670–0.691).
The combined data allowed the identification of small groups with a higher risk for developing melanoma within 5 years (33%).
“Our analyses suggest that selective screening of small, high-risk groups could lead to both more accurate monitoring and more efficient use of health-care resources. This would involve bringing population data into precision medicine and supplementing clinical assessments,” said lead researcher Sam Polesie, MD, PhD, Associate Professor of Dermatology and Verereology at the University of Gothenburg and a dermatologist at Sahlgrenska University Hospital.
The study authors concluded that this machine-learning approach could support early identification of melanoma and more targeted patient follow-up. However, they suggested that more research and policy frameworks are needed before this can be applied to clinical settings.
DISCLOSURES: For full disclosures of the study authors, visit medicaljournalssweden.se.
