A new study from the University of Gothenburg suggests that artificial intelligence could quietly transform how doctors identify patients at risk for the deadliest form of skin cancer before any skin lesions appear.

Melanoma diagnoses have steadily increased across Western countries over the past 50 years. In the U.S. alone, they’re expected to increase by approximately 11% this year. Although screening is largely reactive, patients typically see a dermatologist after noticing suspicious changes. 

A team of Swedish researchers propose a more proactive approach.

Researchers analyzed registry data from over 6 million Swedish adults, pulling information from health systems that have routinely collected demographic and socioeconomic variables, medical diagnoses and medication use data. Of the 6 million individuals tracked, over 38,000 developed melanoma over the five-year study period.

The findings demonstrated that the best performing AI model, which used a method called gradient boosting, correctly distinguished people who later developed melanoma from those who did not in about 73% of cases. Using only sex and age yielded roughly 64% accuracy. 

More strikingly, the model made it possible to identify a small subset of individuals whose five-year melanoma risk was approximately 33%.

Unlike imaging technologies or genetic screening, the AI draws only on historical records maintained in the health systems, making it relatively straightforward to deploy at scale. Martin Gillstedt, a doctoral student at Gothenburg’s Sahlgrenska Academy who conducted much of the analysis, said, “Our results give a clear signal that registry data can be used more strategically in the future,” even if this kind of support is not yet part of routine care.

The findings point toward a model of precision medicine, one where screening for diseases such as melanoma is directed at smaller, higher-risk groups rather than the broader population. Sam Polesie, an associate professor of dermatology and venereology  at the University of Gothenburg who led the study, said the approach could improve both detection accuracy and the efficient use of health care resources.

More studies and policy decisions are needed before the AI-generated risk scores can enter a consultation room as the researchers themselves noted that expanding screening carries the risk of overdiagnosis. 

The study raises an important question that extends well beyond melanoma: If AI algorithms trained on routine health data can reliably predict who develops cancer, how much of early diagnosis should be left to algorithms?