Reading the Evidence Behind the First FDA-Authorized AI Skin Cancer Device

A high accuracy number on a new medical device is a headline, not a conclusion. When the FDA authorized the first artificial-intelligence tool to help detect skin cancer in a primary care office (through the De Novo pathway in January 2024), the reported figures were genuinely strong: sensitivity near 96 percent and a negative predictive value near 97 percent for the three common skin cancers, earned in a blinded multicenter trial. The skill you need is knowing what those two numbers leave out, because the same authorization letter that certifies a narrow, helpful use also draws firm lines around what the device cannot claim to do.

Key points#

• FDA authorization certifies a specific, limited use, not that a device diagnoses cancer, replaces a biopsy, or beats a clinician.
• The device is authorized only as an aid for non-dermatology clinicians, for already-suspicious lesions, in patients 40 and older.
• Sensitivity (near 96 percent) and negative predictive value (near 97 percent) make it a reasonable rule-out prompt.
• Specificity is low (around 21 percent), so many benign lesions get flagged and referral volume rises.
• The trial population was overwhelmingly light-skinned, so the numbers may not transfer to more pigmented skin.

Start with the indication, not the accuracy#

The most informative sentence in any device authorization is the indication for use, and it is easy to skim past. This tool (marketed as DermaSensor) is authorized as an adjunct for physicians who are not dermatologists, to be used on a lesion the clinician has already judged suspicious, in patients aged 40 and older, and always together with the clinical history and examination. The FDA created a new Class II category for it, "software-aided adjunctive diagnostic devices for use by non-dermatology providers," with special controls under De Novo order DEN230008. The order states plainly what the device is not: not a screening tool for the general public, not a standalone diagnostic criterion, and not a way to confirm that a lesion is cancer. Read in that order (indication first, accuracy second), the numbers describe performance inside a fenced-off use case rather than a general promise.

How the device works#

The underlying method is elastic scattering spectroscopy. A handheld probe shines light into a lesion and measures how the tissue scatters it, and an algorithm converts that optical signal into a binary output: investigate further, or continue to monitor. There is no image the clinician interprets and no probability score to weigh. That simplicity is part of the design goal, since the tool is meant for generalists rather than dermatology specialists, but it also means the clinician is trusting a single yes-or-no signal layered on top of their own judgment.

What sensitivity and NPV actually promise#

Sensitivity is the fraction of true cancers the device correctly flags. In the pivotal DERM-SUCCESS study (NCT06690086), primary care physicians across 22 sites in the United States and Australia assessed 1,579 lesions, of which 224 turned out to be cancers. Overall sensitivity was 95.5 percent, and roughly 96 percent within the 40-and-older group the label targets. For a rule-out tool, that is the right property to prize: you want a device that rarely tells you a cancer is nothing.

Negative predictive value is the companion figure, and it is the one most often misunderstood. NPV is not a fixed property of the hardware. It depends on how common cancer is in the group being tested. In this study roughly one in seven biopsied lesions was malignant, and against that prevalence a 97 percent NPV is reassuring. Point the same device at a much lower-risk stream of ordinary moles and the NPV shifts, because the math that produces it changes with prevalence. A number earned in a high-suspicion primary care sample should not be borrowed as-is for a walk-in screening line.

The figure the marketing tends to skip#

Every sensitivity has a mirror image called specificity: the fraction of benign lesions correctly called benign. Here the story turns. In the pivotal study specificity was only about 21 percent. Low specificity means the device labels a large share of harmless lesions as worth investigating, which pushes up referrals and biopsies of things that were never cancer. The clinical-utility data make the tradeoff concrete. Adding the device roughly halved missed cancers among the physicians studied (false-negative referrals fell from about 18 percent to about 9 percent) and nudged clinicians toward more appropriate referral, while specificity dropped at the same time. This is a familiar failure mode for optical skin-cancer devices. A peer-reviewed analysis in npj Digital Medicine notes that an earlier device was pulled from use precisely because a specificity near 10 percent generated too many unnecessary biopsies. A tool that seldom misses cancer but often raises a false alarm can still be worthwhile, as long as the health system can absorb the extra biopsies and everyone understands that a positive reading is a prompt to look harder, not a diagnosis.

Who the numbers came from#

Any accuracy figure is only as portable as the people who produced it. The npj analysis puts the numbers on this: 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). Skin cancer can look different on darker skin, and algorithms trained mostly on lighter skin have repeatedly done worse on darker phenotypes. The FDA flagged this and attached post-market requirements to test performance in underrepresented groups. Until that data matures, the honest reading is that the reported accuracy describes the population actually enrolled, not every patient who might be tested.

A five-question checklist for any authorized AI diagnostic#

"FDA cleared" is a marketing-friendly phrase that compresses a precise 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 later competitors can clear against through the lighter 510(k) route. It is not a verdict that the device improves long-term outcomes, saves money, or outperforms a well-trained clinician, since those are separate questions answered by separate studies. When you meet the next authorized AI diagnostic, five questions cut through the press release:

1. What exactly is the indication, and who is excluded from it?
2. What are the sensitivity and specificity together, never one figure alone?
3. What population generated the numbers?
4. What served as the reference standard the device was measured against?
5. What does the device replace, versus merely inform?

On all five, the primary sources reward reading past the summary line.