What if a simple algorithm, by analyzing your medical file, could identify your risk of developing melanoma long before any visible signs appear? That is now possible, according to a major study published in the journal Acta Dermato-Venereologica. Swedish researchers at the University of Gothenburg trained an artificial intelligence on data from more than 6 million adults to predict, with unprecedented accuracy, who is at risk of developing this particularly feared skin cancer.

What is melanoma and why is it so dangerous?

Melanoma is the most aggressive form of skin cancer. It originates in melanocytes, the cells responsible for skin pigmentation, and can spread rapidly to other organs if diagnosis comes too late. In France, approximately 17,000 new cases are recorded each year, a number that has been steadily rising for several decades.

The key to recovery lies in early diagnosis. Detected at an early stage, melanoma is treatable in the vast majority of cases. But identified too late, the prognosis becomes much grimmer. This is where artificial intelligence comes into play, with detection capabilities that are beginning to surpass those of traditional diagnostic tools.

An AI trained on 6 million Swedish medical records

The Swedish study stands out for its exceptional scale. The researchers did not simply analyze photographs of moles: they exploited the national health registers of the entire Swedish adult population, covering more than six million people. The data taken into account included:

• Patients' age and sex
• Their medical history and past diagnoses
• Previously prescribed medications
• Sociodemographic information (place of residence, socioeconomic status)

The aim was to identify combinations of factors that, even without a dermatological examination, can predict a high risk of melanoma in the coming years.

73% accuracy: a significant quantitative leap

The best-performing AI model in the study achieved an accuracy of 73% in identifying individuals who would actually develop melanoma, compared to only 64% for classical models based solely on age and sex. This gain may seem modest in absolute terms, but in practice it represents thousands of patients better targeted for early screening.

Even more impressive: by cross-referencing all available data, the AI was able to isolate small very high-risk groups, with a probability of developing melanoma reaching 33% over five years. In other words, one in three people identified by this risk group will develop melanoma within five years of the analysis — valuable information for guiding medical follow-up.

How does this AI system actually work?

The model relies on machine learning techniques, in particular gradient boosting algorithms, which are especially effective on medical tabular data. Contrary to a common misconception, the AI does not "look" at skin images here: it analyzes patterns in administrative and clinical data, searching for correlations invisible to the human eye.

For example, certain types of medications taken regularly, combined with specific medical histories and demographic data, may constitute a weak but statistically significant signal. It is this ability to detect composite signals that physicians cannot integrate manually that gives these algorithms their strength.

Toward personalized skin cancer screening?

If these results are confirmed in other populations and healthcare settings, they open the way to more personalized preventive medicine. In concrete terms, this could mean:

• Automatic referrals to a dermatologist based on the calculated risk level
• Better allocation of medical resources toward the highest-risk patients
• Valuable time savings in countries where dermatology consultation waiting times are long

In France, where dermatologists are few in number and often concentrated in major cities, such a tool could transform preventive melanoma monitoring. General practitioners could receive an automated risk report for their patients, helping them decide whom to refer to a specialist as a priority.

Limitations to bear in mind

Despite the enthusiasm this study generates, several caveats apply. First, the model was built on specifically Swedish data, a population with particular genetic and climatic characteristics (fair skin, seasonal sun exposure). The results are not directly transferable to Mediterranean or equatorial populations.

Furthermore, an accuracy of 73% also means that 27% of cases remain undetected. AI does not replace clinical examination: it should be seen as a decision-support tool, not an oracle. Finally, the question of medical data privacy is central: exploiting national health registers, even anonymized ones, raises major ethical issues that will need to be governed by robust regulations.

A quiet but profound revolution

This study is part of a broader movement to integrate AI into preventive medicine. Beyond melanoma, similar algorithms are already being tested to predict the risk of diabetes, cardiovascular disease, or even certain forms of depression. The medical record of tomorrow could include an AI risk score for several conditions, updated regularly throughout consultations.

This is not science fiction: the data exists, the models work, and the most advanced healthcare systems are beginning to integrate them. The challenge now is to do so in an ethical, transparent way, always placing the patient — and not the algorithm — at the center of the medical decision.

Key takeaway: A Swedish AI analyzes millions of medical records to predict melanoma risk with 73% accuracy. A promising tool for early skin cancer detection, pending validation at a larger scale.