Welcome back to the journal review. This is the March 2026 issue of the Journal of the American Academy of Dermatology, and we've got four pieces worth your time this month — a full gene expression profiling study in high-risk squamous cell carcinoma, a brief report stress-testing ChatGPT against dermoscopic images, a letter to the editor pushing back on an obesity-and-melanoma paper, and a database study on where Mohs surgery actually fits for acral lentiginous melanoma. Let's get into it.

First up is an original cohort study out of Mayo Clinic and a Spanish collaborating center, developing a gene expression panel to predict local recurrence, metastasis, and overall survival in intermediate- to high-risk cutaneous squamous cell carcinoma. You know the clinical problem already: cutaneous SCC metastasizes in up to about five percent of cases, with mortality approaching melanoma-level numbers in that subset, and our staging systems are mediocre at telling us who that five percent actually is. The authors point out that even Brigham and Women's staging, which beats AJCC-8 on specificity, still has a positive predictive value of only around thirty percent for metastasis or death. And there's a real upstaging problem — up to almost half of intermediate-to-high-risk tumors get reclassified between diagnosis and surgery. So the gap here is a better molecular tool layered on top of clinical staging, not replacing it.

Methodologically, this builds directly on the group's prior multi-omic work, where whole exome and transcriptome sequencing on stage-matched, outcome-differentiated tumors identified a candidate list of a hundred eighty-three functionally relevant genes. Here they took that gene list, built a custom NanoString probeset — so a targeted, clinically translatable panel rather than another whole-transcriptome experiment — and ran it on a hundred eighty-six archived tumors from a hundred eighty-three patients across four sites, all starting as Brigham and Women's stage T2a or T2b before pathology re-review. That stage-matching, outcome-differentiated design is deliberate: by intentionally enriching for tumors with known bad outcomes alongside matched tumors that behaved well, they maximize statistical power to find gene signals without needing a much larger unselected cohort — a sensible design choice when metastatic events in cutaneous SCC are still relatively rare. Gene selection for the final risk models used LASSO regression for the metastasis and recurrence endpoints and a regularized Cox model for survival, both cross-validated a hundred times to keep only genes that kept showing up as predictive — a reasonable safeguard against overfitting when you're staring down a panel of over a hundred candidate genes in a few hundred samples.

On to results. Of the original candidate genes, they distilled a fourteen-gene risk score for metastasis, and it performed well — area under the curve around eighty-two percent, with roughly two-thirds sensitivity and about eighty percent specificity. A separate three-gene signature for local recurrence was more modest, an area under the curve around seventy-five percent, with sensitivity and specificity both hovering in the high-sixties to low-seventies. And an eight-gene signature was independently associated with overall survival even after adjusting for age, immunosuppression, and metastasis status — a statistically significant and, importantly, an independent association, meaning the gene signal wasn't just a proxy for those known clinical risk factors. Worth noting for context: their prior discovery-phase work using full transcriptomic data had even stronger numbers, ninety percent accuracy and ninety-six percent sensitivity for metastasis — so this NanoString-translated panel is, unsurprisingly, somewhat less discriminating than the original hypothesis-generating dataset, which is the expected trade-off when you compress a genome-wide signal down into a clinically deployable targeted panel.

The provided text cuts off before the authors' own discussion and limitations section, so I won't put words in their mouth there. But a few things are worth flagging as my own read on the methodology: this is a retrospective, case-selected cohort enriched for outcomes, which inflates apparent performance relative to what you'd see screening an unselected real-world population — so these accuracy numbers likely represent a ceiling, not a floor, and prospective validation in an unselected T2a/T2b population will be the real test. It's also worth remembering this panel was tested in a population that was already Brigham and Women's stage T2a or T2b — this is not a tool for thin, low-risk tumors, and roughly nine percent of the cohort ended up reclassified to T1 or T3 on rereview, underscoring exactly the staging instability the panel is trying to compensate for.

Practical takeaway: this is genuinely interesting translational work from a group with a track record here, and the metastasis signature in particular has real discriminative power. But this is not yet practice-changing — there's no commercial assay on the table from this specific panel, and you should treat this as hypothesis-confirming groundwork alongside the already-commercialized SCC gene expression profiling tests, not as something to reach for tomorrow. The concept to bank, though, is that combining molecular data with Brigham and Women's staging is where this field is heading, and it may eventually help resolve exactly the T2a/T2b ambiguity that drives so much of your postoperative radiation and nodal surveillance decision-making.

Next, a brief report: pigmented lesion accuracy by ChatGPT in diagnosis of dermoscopic images, out of NYU. This one's straightforward and doesn't pretend to be more than it is — a diagnostic accuracy study of a publicly available large language model against a well-characterized image set. The authors pulled dermoscopic images from the Harvard HAM10000 dataset — melanoma, benign nevi, seborrheic keratosis, and vascular lesions, all with solid ground truth from pathology or expert consensus — and fed them into ChatGPT-4o with two straightforward prompts: what's the most likely diagnosis, and is this benign or malignant. Two raters did this independently, which lets them report both accuracy and interrater reliability.

The results are not close calls. ChatGPT's diagnostic accuracy was under fifty percent for melanoma, about fifty-three percent for benign nevi, only around thirty percent for seborrheic keratosis, and just sixteen percent for vascular lesions — compare that to a quoted dermatologist benchmark using dermoscopy of over seventy percent, and the gap is large and statistically significant. When simply asked benign versus malignant, melanoma sensitivity and specificity were both roughly in the high-forties, essentially coin-flip performance. The number that should stick with you: the model called an actual melanoma benign in about half of cases. Interrater agreement between the two human raters interpreting ChatGPT's outputs was mostly poor to moderate as well, except for vascular lesions, which the model at least classified consistently, even if not always correctly.

There's no methods complexity to unpack here beyond straightforward prompting of a general-purpose chatbot — this isn't a purpose-built dermatologic AI model, and the authors' point is precisely that the public doesn't know that distinction. Their discussion is appropriately blunt: this isn't a training-data artifact, since prior smaller studies with different image sets show the same pattern, and keratinocyte carcinomas apparently fare somewhat better for these general models than melanocytic lesions do. The clinical concern is false reassurance — a patient with a real melanoma being told it's benign, particularly if that delays them from seeking care they may already have difficulty accessing.

Practical takeaway: not remotely practice-changing for your own diagnostic workflow, but highly relevant for patient counseling. This is good ammunition for the conversation you're probably already having with patients who show you a phone photo and say "the app said it's fine." It's a citable data point that general-purpose chatbot dermoscopic interpretation is currently unsafe as a stand-alone triage tool.

Third is a letter to the editor, "Reassessing the role of obesity in melanoma stage and recurrence," responding to a prospective study published earlier in the Journal that found obese patients with high-risk primary melanoma were paradoxically more likely to be diagnosed at an earlier T-stage, and that body mass index showed no association with recurrence over seven years of follow-up. The letter writers are pushing on both findings. On the T-stage finding, they note it directly contradicts a fair amount of existing literature showing obese patients typically present with thicker melanomas and more nodal metastases, and they're skeptical of the original authors' proposed explanation — that obese patients have more healthcare encounters and thus more opportunities for incidental skin cancer detection. Their pushback is that more healthcare contact doesn't reliably translate into more skin cancer detection specifically, since detection depends heavily on examiner skill, patient comfort, and other variables that don't necessarily track with overall healthcare utilization.

On the recurrence finding, the letter offers a biologically plausible reconciling explanation rather than an outright rebuttal: since recurrence risk is driven heavily by stage at diagnosis and completeness of excision, and early-stage melanoma treated with adequate margins has a low recurrence rate regardless of other factors, the null body-mass-index association may simply reflect that this cohort was mostly early-stage and surgically well-controlled — meaning adipose-driven inflammatory and adipokine effects on tumor biology might only become clinically visible in more advanced or systemic disease, not in patients whose disease was already cured by surgery.

There's no new data here — this is commentary, not a study — so there's nothing to grade for methodology or bias. The value is entirely in the critical framework it offers.

Practical takeaway: not actionable on its own, but worth holding in mind as a caution against over-interpreting a single prospective study's paradoxical body-mass-index finding. If you're counseling patients on obesity as a melanoma risk modifier, the weight of evidence outside this one paper still supports obesity as associated with thicker tumors and worse nodal status at presentation — this letter is a reminder not to let one counterintuitive dataset override that broader picture.

Last, a retrospective cohort study using the National Cancer Database from 2004 through 2022, looking at utilization of Mohs micrographic surgery for acral lentiginous melanoma. The clinical setup here is one you know well — acral lentiginous melanoma sits on functionally sensitive sites where wide local excision or amputation carries real morbidity, and while Mohs has growing evidence for comparable disease-specific survival and better local control, national utilization patterns hadn't been well characterized.

The design is a straightforward large-database retrospective analysis, which is really the only feasible way to answer a utilization question like this — acral lentiginous melanoma is uncommon enough, and Mohs use within it uncommon enough, that you need a national registry to get any meaningful numbers, even though it costs you granular data on margins, recurrence, and long-term survival. They stratified by surgical modality — wide local excision, Mohs, or amputation — and ran multivariable logistic regression to find independent predictors of Mohs selection.

Of about sixty-seven hundred cases, the vast majority — nearly five thousand — got wide local excision, about fourteen hundred were amputated, and only three hundred thirty-five underwent Mohs. But the trend line matters more than the raw split: Mohs use climbed from under one percent in 2004 to about seven percent by 2022, while wide local excision's share declined over the same period. Patients selected for Mohs were older on average, disproportionately had in-situ or stage-one disease, had lower comorbidity burdens, and were overwhelmingly treated at academic or integrated cancer centers. On multivariable analysis, independent predictors of Mohs selection were older age, and treatment at an academic or integrated network cancer center — both statistically significant. Conversely, higher AJCC stage sharply reduced the odds of Mohs selection, which tracks with current practice patterns and guideline framing, since NCCN guidance specifically flags Mohs as an option mainly for minimally invasive, T1a acral melanomas rather than more advanced disease. No differences emerged by sex, race, ethnicity, rurality, radiation use, or short-term mortality.

The authors' own interpretation centers on access rather than biology: since tumor stage alone doesn't fully explain the pattern, and facility type and, with borderline significance, private insurance also predict Mohs use, they argue this reflects structural access barriers — geographic proximity to a center that offers Mohs for melanoma, and the insurance coverage to get there — layered on top of appropriate clinical selection for early-stage disease. That's a reasonable read of an association-only, registry-level dataset, though it's worth remembering the inherent limitations of NCDB analyses generally: no data on recurrence or margins, no central pathology review, and residual confounding from anything not captured in the database, like individual surgeon comfort with melanoma Mohs or local referral patterns.

Practical takeaway: not practice-changing in the sense of telling you to do anything differently at the bench, but a useful, data-grounded confirmation that the field is moving toward accepting Mohs for early-stage, particularly in-situ and T1a, acral lentiginous melanoma, largely concentrated at academic and integrated centers. If you're at such a center, this supports continuing to offer Mohs for appropriately selected early acral lesions; if you're not, it's a reminder that referral-pattern and insurance-driven access gaps are real and worth addressing at a systems level.

That wraps this episode. To recap: a promising but still investigational gene expression panel for high-risk squamous cell carcinoma risk-stratification, a sobering reality check on chatbot-based melanoma triage, a thoughtful critical letter on obesity and melanoma staging that keeps us appropriately skeptical of a single paradoxical finding, and national data confirming Mohs is quietly gaining ground for early acral lentiginous melanoma at specialized centers. Thanks for listening, and we'll see you next month.