Welcome back to this July twenty twenty-six rundown of JAMA Dermatology. Four papers on deck this month, and honestly a nice thematic thread runs through all of them — how do we allocate limited resources, whether that's supplement dollars, AI triage, germline testing, or your own clinic slots, toward the patients who actually stand to benefit. Let's get into it.

First up is a brief report, an economic evaluation, looking at cost-effectiveness of oral nicotinamide for keratinocyte carcinoma prevention. You know the background here — nicotinamide, five hundred milligrams twice daily, has trial evidence behind it, including the original ONTRAC trial showing about a quarter reduction in new keratinocyte carcinomas, and more recent VA-based work from this same group showing a smaller effect overall that jumped substantially when nicotinamide was started after a patient's first keratinocyte carcinoma. It's cheap, it's safe, plenty of us are already recommending it to high-risk patients. What hasn't been done is a formal look at whether it's actually a good economic bet. This paper does that.

Methodologically, this is a modeling study built on real observational data rather than a fresh trial — they took the Veterans Health Administration cohort from that earlier Breglio paper, about thirty-four thousand veterans with a prior keratinocyte carcinoma history, split into those exposed to nicotinamide for a month or more and those not, and layered a one-year cost-effectiveness model on top. Why observational VA data instead of trial data? The authors are explicit about this: trial populations, like the Australian ONTRAC cohort, only enrolled patients with two or more prior keratinocyte carcinomas, reported aggregate outcomes vulnerable to outlier skew, and didn't give individual-level event data, so you can't tell if repeat cancers are piling up in the same few patients. Real-world VA data, despite the usual confounding and selection-bias caveats, better reflects everyday prescribing and adherence patterns, and lets them model cost-effectiveness in a genuinely high-risk population. They calculated incremental cost-effectiveness ratios — cost difference divided by quality-adjusted life-year difference — using VA procedural costs for destruction, excision, and Mohs surgery, inflation-adjusted to twenty twenty-four dollars, and assumed a small quality-of-life decrement, one-hundredth of a QALY, for each treated keratinocyte carcinoma. They stress-tested everything with probabilistic and one-way sensitivity analyses and also modeled non-VA civilian pricing.

The numbers: keratinocyte carcinoma incidence was meaningfully lower in the nicotinamide-exposed group, about two-tenths of an event per person-year versus roughly a quarter in the unexposed group. That absolute risk reduction translates to a number needed to treat of about twenty for one year to prevent one keratinocyte carcinoma, and across the exposed cohort of just over twelve thousand veterans, that works out to roughly six hundred twenty-five keratinocyte carcinomas prevented annually. Here's the economic punchline: total nicotinamide cost across that group was about a hundred sixty thousand dollars, but it offset over half a million dollars in avoided treatment costs, for net savings of about three hundred sixty thousand dollars a year — nearly a fifth reduction in cohort-wide keratinocyte carcinoma treatment spending. Translated into the cost-effectiveness metric everyone actually cares about, this came out as a negative cost per QALY gained, meaning nicotinamide didn't just meet the standard fifty-thousand-dollar-per-QALY willingness-to-pay threshold, it was actually cost-saving. Even under civilian, non-VA cost assumptions, the number came out around fourteen thousand dollars per QALY gained — still comfortably in the highly cost-effective range. And sensitivity analyses, both the one-way and probabilistic versions, kept the result cost-saving or highly cost-effective across nearly the entire range of plausible inputs. They also translated the benefit into patient-reported terms using the Skindex-16 instrument, estimating that preventing those keratinocyte carcinomas preserved a meaningful chunk of symptom burden per lesion avoided.

Limitations are the ones you'd expect from any modeling exercise — this is an approximation, not an empirical trial result. Long-term adherence, interactions with other preventive measures like field therapy or sunscreen, and facility-to-facility variation in procedure reimbursement could all shift the actual numbers. The VA population also skews heavily older, male, and white, which are all keratinocyte carcinoma risk factors, so generalizability to a broader civilian population needs some caution, though the sensitivity analyses using civilian pricing help there. The dataset also didn't capture superficial keratinocyte carcinomas managed with topical chemotherapy, which likely means the benefit is actually underestimated.

Practically, this is about as clean a "reinforces what you're already doing" paper as you'll see. It's not going to change your prescribing pattern — most of us already recommend nicotinamide to high-risk patients with a keratinocyte carcinoma history — but it gives you a genuinely useful health-economics argument to bring to payers, health systems, or skeptical patients: this is a rare preventive intervention that is both clinically beneficial and cost-saving, not just cost-effective. Worth keeping in your back pocket next time you're justifying supplement recommendations to an insurer or a hospital formulary committee.

Second article, and this one's a meatier original investigation: a diagnostic accuracy study asking how artificial intelligence models for skin cancer stack up against real physicians of varying experience levels, specifically under realistic, messy, everyday-practice conditions rather than the sanitized benchmark datasets most AI validation studies use.

The gap they're targeting is well-articulated: most AI-versus-human comparisons use curated dermoscopic image sets, narrow single-image classification tasks, and rarely include the rare, atypical presentations that actually trip up clinicians — amelanotic melanoma, Merkel cell carcinoma, that kind of thing. Real clinical diagnosis integrates history, demographics, physical exam, and image, and prior studies mostly ignore that.

Design-wise, this was a prospective diagnostic study built on a retrospectively collected but deliberately realistic dataset — the TODIV platform, eleven hundred seventeen cases from three French tertiary centers, each with clinical history, demographics, at least one macroscopic photo, and a dermoscopic image. Critically, they intentionally kept image quality variable — pen markings, rulers, hair artifacts, inconsistent lighting — specifically to simulate what you actually see in clinic, rather than the pristine, preprocessed images typical AI training sets use. That's the authors' own stated rationale, and it's a smart methodological choice: it directly tests whether AI performance holds up outside the curated-dataset bubble. They also intentionally overrepresented rare malignant diagnoses at about twelve percent of the malignant cases, again explicitly to probe a weak spot that low population prevalence would otherwise make impossible to evaluate.

They tested three AI systems — a first-generation convolutional neural network fine-tuned on dermoscopic images only, and two configurations of a newer transformer-based foundation model called PanDerm, one dermoscopy-only and one multimodal, incorporating clinical photos and metadata — against six hundred fifty-two physician readers spanning under one year to over ten years of dermoscopy experience, each reading a stratified random hundred-case subset.

Results: the older convolutional neural network was beaten by every human subgroup, even the least experienced readers — its mean accuracy was around fifty-seven percent, meaningfully and statistically below human performance. The newer unimodal foundation model did better, beating readers with under three years of experience, essentially matching the mid-experience group in the three-to-ten-year range. But experts with over ten years of experience still came out on top of everything, human or machine, with the highest multiclass accuracy of the whole study, around seventy-four percent, versus about seventy-two percent for the best AI configuration. Interestingly, the multimodal AI configuration — the one that added clinical photos and metadata on top of dermoscopy — actually underperformed the dermoscopy-only unimodal version, which is a bit counterintuitive and worth sitting with; adding more information didn't help the machine the way you might expect it to help a human clinician.

Discussion-wise, the honest read here is nuanced rather than a simple "AI wins" or "AI loses" headline. Modern foundation models have clearly closed most of the gap with junior and mid-level clinicians, but expert-level pattern recognition — the kind you and I build over a decade of dermoscopy — still has a ceiling AI hasn't reached, at least in this realistic, artifact-laden, rare-disease-enriched test set. Limitations include the predominantly European-ancestry patient population, limiting generalizability across skin phototypes, and the fact that most physician readers were French, recruited through residency and university dermoscopy programs, which may not mirror the experience distribution in other health systems.

The practical takeaway for you as a fellowship-trained Mohs surgeon and dermatologic oncologist: this isn't practice-changing in the sense of telling you to adopt or avoid any particular AI tool tomorrow, but it is a genuinely useful data point for how you think about AI triage in your own practice or in training pipelines. The clear message is that AI performs best as an augmentation tool for less experienced clinicians — physician extenders, general practitioners, junior residents — rather than as a replacement for expert judgment, and that multimodal integration of clinical data into AI models is not yet a solved problem, despite the intuitive appeal.

Third article, another substantial original investigation, this one genomic: prevalence of pathogenic variants in familial melanoma genes and their associated cancer risks among genomically ascertained individuals, rather than the traditional phenotype-first, family history-first cohorts most of our prior data comes from.

The gap being addressed is ascertainment bias. Almost everything we know about genes like CDKN2A, BAP1, POT1, and MITF E318K comes from melanoma-prone families identified because someone already had a strong personal or family cancer history — which inflates the apparent penetrance and risk estimates because those families also likely share environmental and lifestyle exposures on top of the genetics. This study flips that: it starts from genetic findings first, in unselected population biobanks, and looks outward to phenotype.

Methodologically this is a cohort study pooling two enormous population-scale genomic databases — UK Biobank, with genetic sequencing linked to national cancer registry data, and the US Geisinger MyCode cohort, linked to institutional registry data — together covering nearly seven hundred thousand individuals, spanning cancer records from nineteen seventy through twenty twenty-four. The authors chose these genome-first cohorts specifically because genomic ascertainment removes the selection step that biases traditional family studies; you're not enrolling people because they already had cancer or a striking family history, you're just looking at everyone who happened to get sequenced and then checking who has a pathogenic variant and what happened to them. They restricted their gene panel to eight well-established, clinically actionable familial melanoma genes — ACD, BAP1, CDKN2A, CDK4, MITF E318K, POT1, TERF2IP, and the TERT promoter — specifically because a positive result in any of these would actually change clinical management, prompting enhanced dermatologic surveillance or, for some genes, additional organ-specific cancer screening.

Key results: combined pathogenic variant prevalence across all eight genes was low in the general population, around one in two hundred in the Geisinger cohort and about one in one hundred ten in UK Biobank — roughly half a percent to just under one percent overall. Most of that signal was being driven by the MITF E318K variant alone. But — and this is the clinically load-bearing finding — when they restricted to individuals with multiple primary cutaneous melanomas, or a first melanoma diagnosed before age forty, prevalence climbed above two and a half percent, which is the established threshold that the PREMM-plus model uses to justify germline testing for high- and moderate-penetrance cancer genes. In other words, in an unselected population, these variants are rare, but in the specific clinical subgroup of early-onset or multiple melanoma, they cross the actionable testing threshold. On cancer risk, the study replicated known associations — CDKN2A with brain, cutaneous melanoma, head and neck, and pancreatic cancers; MITF E318K with cutaneous melanoma and kidney cancer; POT1 with cutaneous melanoma, hematologic malignancies, and thyroid cancer. But it also flagged several new or previously inconsistent associations worth watching: BAP1 with prostate cancer, CDKN2A with biliary tract, breast, nonmelanoma skin, and small intestine cancers, MITF E318K with cervical, nasal cavity and middle ear, and nonmelanoma skin cancers, and POT1 with myeloma. Age-of-onset patterns for cutaneous melanoma were consistently earlier in CDKN2A and MITF E318K carriers versus noncarriers, though that earlier-onset pattern was less consistent when they looked at internal malignancies.

The authors' own discussion is appropriately measured about the novel associations — with population biobank data you're doing many comparisons across many cancer types, so some of these new signals, particularly the rarer ones like BAP1-prostate or POT1-myeloma, need independent replication before they're treated as established. The strength of the paper is precisely that genome-first ascertainment minimizes the shared-exposure confound of family-based studies, but the tradeoff is that biobank cohorts skew heavily toward European genetic ancestry in both cohorts, over ninety percent in each, so applicability to other ancestral backgrounds is genuinely uncertain, and biobank participants are a self-selected, generally healthier and older population than the general public to begin with.

Practical takeaway: this is edging toward practice-relevant, though not immediately practice-changing on its own. The finding that supports lowering your testing threshold trigger — multiple primary melanomas or melanoma before age forty crossing that two-and-a-half-percent actionable prevalence line — reinforces existing genetic counseling referral practice rather than overturning it. What is worth filing away and watching for validation is the expanded cancer spectrum, particularly the BAP1-prostate and expanded CDKN2A associations, since if those replicate, they could eventually reshape screening recommendations for confirmed carriers beyond the cancers we currently monitor for.

Fourth and final article, a brief report, this one a cross-sectional study asking a very practical clinic-operations question: who actually is the patient population showing up for asymptomatic, worry-driven periodic skin checks, and how much yield are we really getting from examining them.

The clinical gap here is that periodic total body skin exams for asymptomatic patients are widely performed and widely expected by the public, yet no major body — not the American Academy of Dermatology, not the American Cancer Society, not the US Preventive Services Task Force — has a formal recommendation supporting this practice, and the Task Force has explicitly stated the evidence is inadequate to show visual skin screening reduces morbidity or mortality. This study wanted to characterize exactly who is filling those visit slots.

This is a single-site secondary analysis of a routine previsit survey at Emory's dermatology clinic, collected over about a year and a half, capturing demographics and skin cancer risk factors — phototype, sunburn tendency, hair and eye color, personal and family cancer history — for new patients presenting specifically because they wanted a skin check for skin cancer concern, without any specific lesion complaint. The rationale for this design is straightforward and the authors more or less say it directly: this is a pragmatic, real-world clinic population, capturing actual triage behavior rather than a research-selected cohort, and it lets them calculate a very concrete, actionable metric — number needed to examine to find one skin cancer — stratified by simple, chartable risk factors.

The results are pretty striking. Of over a thousand asymptomatic patients seeking a general skin check, about half were fifty or younger, and among that younger half only around one in eleven had a personal history of skin cancer. Only about twelve percent of the whole cohort ended up needing a biopsy at all, and only thirty-eight skin cancers were found across the entire group. Of those thirty-eight, thirty-five — over ninety percent — occurred in patients over fifty, and thirty-seven of thirty-eight occurred in patients with phototypes one through three. The number needed to examine to find a single skin cancer was a striking one hundred eighty-one in patients fifty and younger, compared to just seven in patients over seventy — a twenty-six-fold difference driven by age alone. Personal history of skin cancer mattered too: number needed to examine was twelve in those with a prior skin cancer history versus fifty-two in those without. Phototype and eye color showed similar stratification — phototypes one through three had comparable, much lower numbers needed to examine than the aggregate for phototypes four through six, and brown-eyed patients had a substantially higher number needed to examine than blue, green, or hazel-eyed patients. Sunburn tendency, hair color, and family history showed comparatively modest differences in yield by comparison.

The authors frame this plainly: a substantial proportion of the patients currently filling skin-check appointment slots are, by any reasonable epidemiologic measure, at very low near-term risk of an actual skin cancer diagnosis. As a single-center, cross-sectional study, the obvious limitations apply — this is one academic practice, roughly half of eligible patients completed the survey so there's some selection even within the sample, and it captures near-term yield only, not longitudinal cancer development, so it can't speak to whether periodic exams have downstream benefit through habit-building or earlier detection of future lesions.

Practically, this one is genuinely actionable at the operations level, even if it's not the kind of finding that changes what you do at the biopsy table. It supports building simple, low-friction triage criteria — age, phototype, and personal skin cancer history — into scheduling or intake workflows to preferentially direct comprehensive skin exam slots toward patients who are actually likely to yield a cancer diagnosis, freeing up capacity that's currently being spent on very-low-yield exams in young, darker-phototype, no-history patients. It won't change your exam technique or your differential, but it's a solid, data-backed argument for smarter triage in a landscape where dermatology access is chronically capacity-constrained.

That wraps our four papers for July. A cost-effectiveness case for nicotinamide that should make budget conversations easier, a sobering but nuanced reality check on AI's current ceiling relative to true expertise, a genome-first look at familial melanoma genes that mostly reinforces existing testing thresholds while flagging some new associations to watch, and a very practical argument for smarter triage of the worried-well seeking skin checks. Thanks for listening, and we'll see you next month.