Welcome back to the journal review. This is our September twenty twenty-six run-through of the Journal of the American Academy of Dermatology, and we've got four pieces on the docket today — one full ecological cohort study, a letter-and-reply exchange on AI dermoscopic diagnosis, an ethics-focused letter on visually impaired patients, and a brief report pulling from the Global Burden of Disease data. Let's get into it.

First up is an original research study, and it's the meaty one today: "Higher dermatologist density is associated with lower proportions of late-stage melanoma," an ecological study built on the State Cancer Profiles database, out of University of Miami.

The background here is one you already live and breathe: localized melanoma has a five-year survival north of ninety-nine percent, but that craters to the seventies for regional disease and down into the thirties for distant metastatic disease. And yet nationally, something like one in seven melanomas still presents late. We know dermatologists find thinner melanomas and diagnose earlier stage disease than primary care physicians do, but the specific question this group wanted to answer is a geographic-access one: does the raw density of dermatologists in a region actually track with how many melanomas in that region get caught late? Prior work looking at this used states or counties as the unit of analysis, which are administratively convenient but don't really map onto how patients actually access care. So the gap they're filling is both a substantive question and a methodologic upgrade.

For methods, they used what NCI calls Health Service Areas — HSAs — which are defined by counties that are self-contained with respect to hospital utilization patterns, so they function more like real "care markets" than arbitrary political boundaries. That's a smart choice if you're trying to model access to care, and it's the authors' own stated rationale, not my inference. They pulled five years of data, 2017 through 2021, merging the State Cancer Profiles registry data with the Area Health Resource File for physician counts, then stratified each HSA into five buckets of dermatologist density per hundred-thousand population, from zero up to more than four. Of nine hundred forty-eight HSAs nationally, five hundred fifty-seven made the final cut — the rest were dropped mostly because of small-cell censoring, since the registry suppresses areas reporting fewer than sixteen late-stage cases over the five-year window to protect confidentiality, plus a couple of states excluded for legislative suppression or missing workforce data.

This is a genuine stats-teaching moment worth pausing on. Their outcome — proportion of melanoma cases that are late-stage within an HSA — is a percentage bounded between zero and one, which violates the assumptions of ordinary linear regression if you tried to model it directly, since a regression line has no idea it shouldn't predict below zero or above a hundred percent. So they used a logit transformation on the proportions before running the linear regression. In plain terms, the logit takes a proportion and re-expresses it on an unbounded scale — essentially the log of the odds — which lets you fit a standard linear model whose coefficients, once you exponentiate them back, hand you clean, interpretable odds ratios at the population level. It's a fairly elegant way to get OR-style output out of aggregate, ecological — not individual-patient — data, which is otherwise an awkward format to run logistic regression on directly. They also used robust standard errors, which is a hedge against the heteroskedasticity you'd expect when your units — HSAs — vary enormously in population size and case volume. And covariate selection was governed by Akaike Information Criterion minimization, essentially an automated way of choosing the set of predictors that best balances model fit against model complexity, penalizing variables that don't earn their keep.

Onto results. The median HSA had about two dermatologists per hundred-thousand people, and the median proportion of melanoma diagnosed late was about sixteen percent — consistent with the national figures you'd expect. Unadjusted, there was a clean stepwise gradient: late-stage proportion ran from about nineteen percent in HSAs with zero dermatologists down to under fourteen percent in HSAs with more than four per hundred-thousand. After adjustment, that gradient held up and stayed statistically significant at every step: compared with zero-dermatologist HSAs, the odds of late-stage diagnosis were about eleven percent lower in the lowest density tier, another eleven percent lower in the next tier, sixteen percent lower in the >2-to-4 tier, and twenty percent lower in the highest tier, greater than four per hundred-thousand. That's a genuine dose-response pattern, which is one of the more persuasive features you can see in observational data — it's not just "some dermatologists versus none," it's a graded relationship that strengthens as density increases, which argues against the finding being purely due to some binary confounder. A few sociodemographic covariates were independently significant too — greater proportions of White or Hispanic residents nudged odds of late-stage disease up slightly, while higher educational attainment and insurance coverage were protective — though those effect sizes were quite small, on the order of half a percent to one percent change in odds per one percent change in population composition, which is statistically real but not clinically the headline here. Notably, primary care physician density fell out of the model entirely during AIC selection, meaning it didn't add explanatory value once dermatologist density and the other covariates were in there.

Discussion and limitations: the authors are appropriately careful to frame this as correlation, not causation — it's an ecological design, meaning all these associations exist at the level of the geographic area, not the individual patient, so you can't say any specific patient's outcome was caused by their local dermatologist supply. There's also real censoring bias baked in, since over three hundred HSAs were dropped for low case counts, and those tend to be lower-population, likely more rural areas — precisely the places you'd most want data on. The authors' proposed mechanism is plausible and consistent with prior literature: more dermatologists likely means shorter wait times, more routine skin exams catching thinner lesions incidentally, and more consistent dermoscopy use. They also flag a striking descriptive finding — a cluster of forty-seven HSAs with literally zero dermatologists and above-median late-stage burden, concentrated in the South and Midwest, which is a nice hypothesis-generating breadcrumb for targeted workforce interventions even though it's not itself statistically tested.

Practical takeaway: this isn't practice-changing for what you do chairside tomorrow — you're not going to alter your own diagnostic approach based on this. But it is a strong piece of policy-relevant evidence for anyone involved in workforce planning, teledermatology triage design, or advocacy for expanding specialty access in underserved HSAs, particularly the South and Midwest pockets they identify. If you're involved in residency expansion decisions, satellite clinic placement, or state-level scope-of-practice advocacy, this is a citation worth having in your pocket.

Next, a letter-and-reply pair rather than a standalone study — this is the reply from Fathy, Tadros, and Alhariri responding to a comment letter from Akhter and colleagues, both revisiting the group's earlier brief report on GPT-4 Vision in dermoscopic image analysis. There's no formal methods-results-limitations structure here since it's correspondence, but there's real new data packed into it, so it's worth walking through.

The original commentary letter had pointed out that large language model performance in dermoscopy is evolving fast, and that version-specific, clinically grounded testing matters — you can't judge a model's real-world utility off a stale software version. In this reply, the authors respond by actually rerunning the experiment at scale: seventeen different multimodal model configurations, seven from OpenAI and ten from Gemini, tested against one hundred biopsy-confirmed dermoscopic images spanning eight diagnoses, under six different prompting strategies — four that forced the model to pick from a fixed label list, and two free-form. That's over ten thousand total model-image-prompt trials.

The headline number: overall eight-class accuracy across everything was about seventy-one percent, which is a substantial jump from the roughly thirty-six percent melanoma-detection accuracy reported in earlier GPT-4 Vision evaluations. Prompt structure mattered enormously — label-constrained prompting hit about seventy-six percent accuracy versus under sixty percent for free-form responses, a meaningful gap that tells you prompt engineering isn't a cosmetic detail, it's load-bearing for clinical-grade performance. The single best configuration, a Gemini model run with low reasoning effort but high image resolution, hit about eighty-three percent overall and close to eighty-nine percent within the constrained-prompt arms — numerically edging out the roughly eighty-two percent aggregate dermatologist accuracy benchmark from their earlier study, though they're upfront that cross-study comparison here is soft, not a head-to-head trial. Interestingly, bumping image resolution alone moved one Gemini model's accuracy from sixty-nine up to eighty-two percent, suggesting a lot of the "diagnostic" gain is really about image processing fidelity rather than reasoning.

But — and this is the clinically important caveat they lead with — per-diagnosis performance was wildly uneven, from about fifty percent on melanocytic nevi up to eighty-five percent on verruca vulgaris, with melanoma sitting around seventy-five percent and squamous cell carcinoma as the weakest malignant category at under sixty percent. When they collapsed everything into malignant-versus-benign, sensitivity was about eighty-five percent and specificity about eighty-one percent, but the raw error counts were substantial — roughly nine hundred malignant lesions called benign and about five hundred benign lesions overcalled as malignant across the full trial set. And the confusion patterns mirror exactly the traps human learners fall into — melanoma read as seborrheic keratosis, basal cell carcinoma read as melanoma, benign nevi overcalled as melanoma.

Their bottom line, which they state explicitly, is that prompt-optimized configurations can now approach prior dermatologist benchmarks on paper, but clinically important errors persist, and current systems remain unsuitable for unsupervised dermoscopic diagnosis — useful only as adjunctive tools for education, triage, and decision support within a dermatologist-supervised workflow. For you, the practical read is: interesting, rapidly moving, not remotely practice-changing. Nothing here should touch how you use — or don't use — AI tools in your own diagnostic process yet, but it's worth tracking as a benchmark of how fast the error profile is shrinking.

Third is a letter to the editor, an ethics and practical-care piece on skin cancer management and postoperative considerations specifically in blind and visually impaired patients, from authors at Wayne State and Mayo Clinic. This is a response to an earlier report on skin cancer screening in this population, and it extends the conversation from screening into treatment and postoperative logistics — there's no data or methods here, it's a framework built around the core bioethical principles of autonomy, beneficence, and non-maleficence.

On autonomy, the point is straightforward but easy to overlook: don't assume a blind or visually impaired patient is indifferent to cosmetic outcome just because they can't see the result themselves — walk through Mohs versus other modalities, wound care logistics, and repair preferences the same way you would with any patient, and communicate directly with the patient rather than defaulting to a caregiver, with clear spatial language and real-time narration of each procedural step to avoid startle responses.

On beneficence, the practical suggestions are concrete and very actionable in a Mohs practice specifically: consider triaging these patients earlier in the clinic day to ease transportation burden, and preferentially use dissolvable epidermal sutures when appropriate so patients aren't forced back in purely for suture removal. If a patient's ability to manage postoperative wound care independently is in question, they raise intralesional chemotherapy as an option to consider for amenable tumors, paired with closer follow-up for repeat injections.

On non-maleficence, the ask is to simply build extra time into these encounters and narrate surgical steps continuously rather than assuming silence is neutral — for a sighted patient silence during a procedure is just quiet, for a patient who can't see what's happening it can read as alarming or disorienting.

Postoperatively, they emphasize multi-format instructions — verbal, Braille, large print, recorded audio — home health services when caregiver support isn't available, and specifically educating patients and caregivers on the non-visual symptoms of hematoma or infection, since the visual cues we typically rely on for catching these complications early aren't available to this population. And on an ongoing basis, they suggest a tactile sunscreen dosing cue — two finger-lengths per body area using pump bottles — as a practical photoprotection strategy.

None of this changes your oncologic decision-making, but it's a genuinely useful checklist for accommodating a patient population that's easy to underserve simply through workflow defaults nobody thought to question. Worth flagging to your own clinic staff even if you never write it down anywhere formal.

Last is a brief report, an epidemiologic analysis of melanoma burden across G20 countries using Global Burden of Disease Study data from 1990 through 2023, out of Peking University. This is descriptive and projective rather than a comparative clinical study, so there's no traditional methods-versus-results structure — it's a data synthesis piece — but there is a forecasting technique worth a quick stats note.

The topline trend: across G20 nations, raw incidence, mortality, and disability-adjusted life year counts all rose substantially from 1990 to 2023 — incidence numbers roughly doubled, and the disability-adjusted life year burden grew by more than fifty percent. But raw counts conflate population growth with true disease risk, so the more informative numbers are the age-standardized rates, which strip out the effect of population size and aging. Age-standardized incidence rose modestly from 1990 through about 2010, then actually declined from 2011 onward, while age-standardized mortality and disability-adjusted life year rates held flat until around 2008 before gradually declining too — with an odd little uptick in twenty twenty-two that the authors attribute to pandemic-era disruption in timely diagnosis catching up in the data. Sex patterns crossed over interestingly: women had higher incidence before age fifty, but after fifty men pulled ahead with a widening gap into older age.

Country-level variation was wide — Australia had the highest age-standardized rates in the entire G20 in 2023, unsurprisingly, while India had the lowest incidence and mortality rates and Saudi Arabia the lowest disability-adjusted life year rate. The United States stood out as the only G20 country with a declining age-standardized incidence rate, and it also posted the largest reductions in mortality and disability-adjusted life year rates — a genuinely notable outlier worth remembering next time someone asks whether photoprotection messaging and earlier detection efforts in the US are actually working at a population level; this dataset says yes, comparatively. Saudi Arabia showed the fastest rising incidence and mortality, and India the fastest-growing disability-adjusted life year burden.

Methodologically, they ran a decomposition analysis to tease apart how much of the rising disability-adjusted life year burden is attributable to population growth, aging, versus underlying epidemiological change — a technique that essentially partitions a total change into additive components, letting you say population growth added this much, aging added that much, and shifts in actual per-person disease risk subtracted from the total. Here, growth and aging both pushed the burden up, while the epidemiological component — the true underlying risk trend — actually worked in the opposite, protective direction. They also used an autoregressive integrated moving average model, a standard time-series forecasting method that extrapolates future values from the autocorrelation structure of past data — essentially fitting the pattern of how each value relates to its own recent history and trend, then projecting forward. Their forecast to 2040 predicts continued modest rises in age-standardized incidence but continued declines in mortality and disability-adjusted life year rates, which would mean more melanomas but proportionally better outcomes — consistent with earlier detection and better systemic therapy access outpacing any rise in true incidence.

The limitations are the ones inherent to any Global Burden of Disease analysis: data heterogeneity and estimation uncertainty across very differently resourced health systems, no granularity on tumor thickness, stage, or histologic subtype, and a forecast that by construction just extrapolates historical trend lines and can't anticipate new diagnostics or therapies. For you, this is squarely interesting-not-actionable — there's nothing here that changes a single decision at the patient level — but it's a useful global-context data point, especially the U.S.-as-outlier finding, if you're ever asked to comment on whether public health messaging around melanoma is actually moving the needle.

That wraps our four articles this month — a solid access-to-care study reinforcing the value of dermatologist density, a fast-moving but still-not-ready AI diagnostic benchmark, a practical ethics reminder for an easily overlooked patient population, and a global epidemiologic snapshot putting U.S. trends in international context. Thanks for listening, and we'll see you next month.