How to Read the Evidence Behind the First FDA Cleared AI Skin Cancer Device

The first artificial-intelligence device the FDA authorized to help detect skin cancer in a primary care setting, granted through the De Novo pathway in January 2024, reports a sensitivity near 96 percent and a negative predictive value near 97 percent for the three common skin cancers. Those numbers are real and were earned in a blinded multicenter trial. But they answer only half of the question a careful reader should ask. Authorization certifies that a narrowly defined, adjunctive use is reasonably safe and effective; it does not certify that the device finds cancer on its own, replaces a biopsy, or performs equally well for every patient and every mole.

What the FDA actually authorized

The device (marketed as DermaSensor) uses elastic scattering spectroscopy: it directs light into a lesion, and an algorithm interprets how the tissue scatters that light, returning a binary output that tells the clinician to either investigate further or monitor. The FDA classified it as a Class II device under a newly created generic category, "software-aided adjunctive diagnostic devices for use by non-dermatology providers," with special controls (De Novo order DEN230008). Read the indication closely. It is authorized for patients aged 40 and above, for lesions a clinician has already judged suspicious, used by physicians who are not dermatologists, and only alongside the clinical history and exam. The order is explicit that the device is not a screening tool, not to be used as the sole diagnostic criterion, and not a way to confirm a diagnosis of skin cancer.

Sensitivity and NPV, and what they assume

In the pivotal DERM-SUCCESS study (NCT06690086), primary care physicians at 22 sites in the United States and Australia evaluated 1,579 lesions, of which 224 were cancers. The device's overall sensitivity was 95.5 percent, and about 96 percent in the age-40-and-older group the label targets, with a negative predictive value near 97 percent. Sensitivity is the share of true cancers the device flags, so a high value is exactly what you want from a rule-out tool. Negative predictive value, though, is not a fixed property of the hardware. It rises and falls with how common cancer is in the population being tested. In a primary care sample where roughly one in seven biopsied lesions was malignant, a 97 percent NPV is reassuring; apply the same device to a much lower-risk stream of moles and that figure would drift.

The number the marketing tends to skip

Sensitivity has a mirror image: specificity, the share of benign lesions correctly called benign. Here the picture is far less flattering. In the pivotal study the device's specificity was only about 21 percent, and specificity in this low range is the recurring weakness of optical skin-cancer devices. Low specificity means many benign lesions get flagged, which drives more referrals and more biopsies. The companion clinical-utility data show the tradeoff plainly: adding the device roughly halved missed cancers, cutting false-negative referrals from about 18 percent to 9 percent, and pushed clinicians toward more appropriate referral, while specificity fell at the same time. This is the same failure mode that troubled earlier optical devices; a peer-reviewed analysis in npj Digital Medicine notes that one such device was discontinued precisely because its specificity near 10 percent generated too many needless biopsies. A tool that rarely misses cancer but frequently cries wolf can still be worthwhile, provided the health system can absorb the added biopsies and patients understand that a positive reading is a prompt to look closer, not a diagnosis.

Who was studied, and who was not

Evidence is only as generalizable as the people who produced it. In DERM-SUCCESS, roughly 97 percent of patients were of white race and only about 13 percent had the most pigmented skin types (Fitzpatrick V and VI), as the npj analysis underscores. Skin cancer can present differently on darker skin, and algorithms trained largely on lighter skin have repeatedly underperformed on darker phenotypes. The FDA recognized this and imposed post-market requirements to test performance in underrepresented populations. Until that data matures, the reported accuracy is most honestly read as accuracy in the population that was actually enrolled.

Reading "FDA cleared" like a scientist

"FDA cleared" is a marketing-friendly phrase that compresses a specific regulatory finding into two words. De Novo authorization means the agency judged a novel, low-to-moderate-risk device safe and effective for one defined use, and it also creates a predicate that future competitors can clear against through the lighter 510(k) pathway. It is not a verdict that the device improves long-term outcomes, is cost-effective, or outperforms a well-trained clinician. Those are separate questions settled by different studies. The disciplined way to read any authorized AI diagnostic is to ask five things: what exactly is the indication, what is the sensitivity-and-specificity pair (never one figure alone), what population generated the numbers, what served as the reference standard, and what the device actually replaces versus merely informs. On every one of those, the primary sources here reward reading past the press release.

This article is educational and not medical advice; any decision about a specific lesion belongs with a qualified clinician.