Design: External validation in a population-based prospective cohort
Population: Up to 43,235 adults aged 45 and over in New South Wales, Australia
Primary outcome: Discrimination and calibration for 10-year risk of first primary melanoma
Effect: AUC 0.58 (0.56–0.60) to 0.70 (0.68–0.72); most models underestimated melanoma numbers

Researchers tested six published melanoma risk prediction models in about 43,000 adults aged 45 and over from the New South Wales 45 and Up cohort, linked to the cancer registry. Over follow-up there were 946 invasive and 1,406 in situ melanomas.

Discrimination, the ability to rank people by risk, ranged from an AUC of 0.58 to 0.70: modest at best. Most models predicted fewer melanomas than actually occurred, so they would understate absolute risk.

The better models performed about as well in this independent cohort as in their original studies, which is encouraging for their use in sorting people into broad risk bands for prevention or surveillance. They are not precise enough to decide on an individual's need for skin checks alone.

This cohort is Australian, with very high melanoma incidence and mostly fair-skinned people. These models have not been validated in Indian populations, where melanoma is rarer and more often acral, and should not be applied there.

Use melanoma risk models to sort people into broad risk groups, not to decide individual care alone.
Expect most models to underestimate absolute risk in high-incidence populations.
Keep naevus count, skin type, sunburn history and family history central to risk assessment.
Do not apply these models to Indian patients; they were not built or tested there.

Why it matters

It tempers expectations of risk-stratified melanoma screening built on today's models.

The statistics, in plain English

An AUC of 0.5 is a coin toss and 1.0 is perfect; 0.70 means the model ranks a future melanoma patient above someone who stays well about 70% of the time. Calibration measures whether predicted numbers match observed ones; most models here predicted too few.