Five Years Early

The Stakes

Why Catching Melanoma Early Is Almost Everything

Melanoma is the deadliest form of skin cancer. It accounts for only about one percent of skin cancer cases but is responsible for the large majority of skin cancer deaths. The most important factor in survival is not the treatment -- though treatments have improved dramatically -- but the stage at which the cancer is caught.

When melanoma is diagnosed at its earliest, localised stage, the five-year survival rate in the United States is approximately 99 percent. When it has spread to distant organs, that figure falls to around 29 percent. The difference between those two outcomes is almost entirely determined by how early the cancer is detected. No drug, immunotherapy, or surgical advance can close that gap as effectively as catching the disease before it spreads.

Current screening approaches are not systematic in most countries. Dermatologists examine skin visually, often aided by dermoscopy, and refer suspicious lesions for biopsy. This works well for patients who are already motivated to seek skin checks, but it depends on patients presenting -- and on clinicians having the time and context to identify who most needs a close look. The Swedish study proposes something different: using existing data to find the people who most need to be in the clinic before they have any symptoms at all.

The Study

Six Million People, Routine Records, and a Machine That Reads Risk

The study, led by researchers at the University of Gothenburg in collaboration with Chalmers University of Technology, drew on Sweden's exceptional national health registry infrastructure. Sweden maintains comprehensive, linked registries that cover virtually its entire population -- medical diagnoses, prescribed medications, inpatient and outpatient records, and socioeconomic data -- at an individual level and over long time periods. This infrastructure made it possible to assemble a cohort of 6,036,186 adults and follow them over five years, identifying who developed melanoma and who did not.

Of those six million people, 38,582 (approximately 0.64 percent) developed melanoma during the follow-up period. The study then used this cohort to train and evaluate machine learning models that attempted to predict, from information available at the start of the observation period, which individuals would develop melanoma within five years. The predictive variables included age, sex, medical diagnoses recorded in the registry, medications in use, and socioeconomic indicators.

The most sophisticated model distinguished future melanoma cases from non-cases with an AUC of approximately 0.73 -- meaning that in roughly 73 percent of pairwise comparisons between a person who developed melanoma and one who did not, the model correctly ranked the melanoma case as higher risk. This compares to approximately 64 percent when using only age and sex -- the kind of blunt demographic stratification that traditional screening guidelines rely on. The improvement is modest in absolute AUC terms, but its implications for targeted screening are significant.

More concretely: the model identified a small, high-risk subset of the population for whom the estimated probability of developing melanoma within five years was approximately 33 percent. In a general population where baseline risk is 0.64 percent, identifying a group with a 33 percent five-year risk represents a roughly 50-fold enrichment. This is the group that a systematic screening programme could prioritise.

Wider Context

AI in Dermatology: What Has Already Been Shown

The Swedish population study sits within a broader body of AI dermatology research that has been accumulating for several years. The direction of travel is consistent: well-designed AI systems, given appropriate training data, can match or in some settings exceed dermatologist-level performance on specific diagnostic tasks.

The Swedish study addresses a different but complementary question. Where image-based AI asks "does this lesion look like melanoma?", the registry-based approach asks "does this person's overall health profile suggest they are likely to develop melanoma in the next five years?" These are distinct screening modalities. The first is triggered by a lesion that is already visible. The second identifies who to look for before any lesion appears.

A separate analysis published by ASCO Post in April 2026 reviewed AI melanoma diagnosis accuracy more broadly, confirming dermatologist-level accuracy in multiple settings while noting that external validation -- testing models trained in one population on a different population -- remains necessary before widespread clinical deployment. The Swedish study is notably strong on this dimension: training on millions of population-representative individuals reduces the risk of overfitting to a narrow clinical sample.

Implications

Precision Screening and the Data That Already Exists

The most significant aspect of this study may not be the algorithm itself, but what it says about the information that healthcare systems are already sitting on. Sweden's registry infrastructure is unusually comprehensive, but the underlying data types -- age, diagnoses, medications, socioeconomic factors -- exist in fragmented or aggregated form in most developed healthcare systems. The question is whether they can be assembled and deployed at the individual level.

The researchers themselves framed the implication as a shift toward what they called "selective screening of small, high-risk groups." Rather than attempting to screen an entire population for melanoma -- which is cost-prohibitive and associated with significant false-positive overdiagnosis -- the model suggests it may be possible to identify the specific one or two percent of the population that carries the large majority of the risk, and to concentrate dermatology resources there.

There are important caveats. An AUC of 0.73, while substantially better than chance or simple demographic stratification, still means that a significant fraction of individuals who will develop melanoma are not identified -- and conversely, that a substantial number of people flagged as high-risk will not develop the disease. The 33 percent five-year risk in the highest-risk group means that roughly two in three people in that group will be screened without ultimately developing melanoma. Whether that rate of false positives is acceptable depends on the cost, invasiveness, and anxiety burden of the screening procedure being proposed.

None of these caveats change the underlying fact: a machine reading routine administrative records can meaningfully stratify melanoma risk five years in advance. That was not previously known to be possible. It now has been demonstrated in one of the largest and most rigorously constructed cancer prediction studies ever published. The next questions -- validation in non-Swedish populations, healthcare system integration, clinical pathway design -- are the ones that will determine whether this finding becomes a changed standard of care or remains a proof of concept.