Welcome back to the journal review. This is our February twenty twenty-six installment, and we've got four pieces from the Journal of the American Academy of Dermatology on the docket — an ethics column on AI in low-resource settings, a two-part correspondence exchange on amelanotic melanoma survival that's genuinely worth sitting with methodologically, and a brief report on adjuvant radiation for pleomorphic dermal sarcoma that has direct bearing on how we counsel patients after resection. Let's get into it.

First up is an ethics case discussion — styled as a "Dear Doctor" column — on the ethical deployment of artificial intelligence in low-resource dermatologic settings. This isn't a study, so there's no methods or results scaffolding here; it's a structured ethical reasoning piece built around a vignette: an elderly woman with a suspicious pigmented lesion, sixty miles from the nearest dermatologist, whose primary care physician is considering an AI-powered triage device. The authors use this to walk through the ethical tension of deploying imperfect AI tools where the alternative is often no dermatologic evaluation at all.

The beneficence argument gets built around Dermasensor, the FDA-cleared, five-ten-k-pathway device intended to help non-dermatologist physicians triage lesions. The cited numbers are worth knowing if you get asked about this clinically — it improved management sensitivity from about eighty-two percent up to roughly ninety-two percent, improved diagnostic sensitivity from about seventy-one percent up to about eighty-two percent, and cut false-negative referral rates from eighteen percent down to under nine percent. Meaningful gains, but the authors are careful to flag that this tool is not validated as a stand-alone screening device and is not cleared for use by non-physician extenders — which actually narrows its real-world reach in exactly the underserved areas it's meant to help, since a majority of rural regions face primary care shortages themselves, and demand for dermatologic care in these areas is projected to outstrip supply by well over a hundred and fifty percent.

Then comes the equity problem you'd expect: these models are trained predominantly on lighter skin types, and cited data show significantly worse diagnostic performance for melanoma and basal cell carcinoma in Fitzpatrick types four through six. Layer onto that the practical infrastructure barriers — reliable electricity, connectivity, per-scan manufacturer fees, trained personnel to maintain the hardware — and you can see how an intervention meant to reduce disparity risks entrenching it instead if deployed carelessly. The authors also raise the softer but real issues of over-reliance eroding clinical judgment, unclear liability when the algorithm errs, and the risk of community trust collapsing if misdiagnoses become visible without transparent communication about the tool's limitations.

Their proposed path forward is basically a four-part bundle: build context-specific models trained on locally representative skin types and disease patterns, keep AI in a supportive rather than autonomous role, establish clear regulatory and data-transparency frameworks with mandated ongoing performance audits, and invest in local capacity-building so community health workers and patients themselves are stakeholders in how the tool gets used. The practical takeaway for those of us in academic and referral practice is less about changing what we do chairside and more about anticipating what's coming through the door — patients increasingly arriving after an AI-flagged lesion, sometimes from tools we've never heard of, with variable and unaudited real-world performance. Worth knowing the landscape, not something that changes Mohs practice today.

Now to a pair of pieces that I want to walk through together, because they're a correspondence exchange and they only make sense as a dialogue — a letter to the editor and the original authors' response, both revisiting a SEER-based analysis of disease-specific survival in amelanotic versus melanotic melanoma.

Some background: the original study by Nguyen and colleagues, published in this same issue, used SEER data from two thousand to twenty twenty-one and found that amelanotic melanoma carried worse disease-specific survival than melanotic melanoma, and this held up even after multivariate adjustment for stage. That finding was notable because it contradicted an earlier National Cancer Database analysis by Hopkins and colleagues, who'd found the survival gap essentially disappeared once you adjusted for Breslow depth and ulceration — implying the worse outcomes in amelanotic melanoma were more about delayed diagnosis at a deeper stage than about the tumor behaving more aggressively in its own right.

The letter from Joshi and colleagues zeroes in on exactly this discrepancy. Their critique is a methodology lesson in itself: Nguyen's group had adjusted for "stage," but they derived that variable by combining two different SEER staging schemes — Summary Stage 2000 and Combined Summary Stage — an approach that prior work has shown is prone to misclassification. Critically, they hadn't adjusted directly for Breslow depth and ulceration, the two variables that actually drove the attenuation in the earlier Hopkins analysis. So Joshi's group reran the comparison, restricting to SEER data from twenty ten onward — the only years Breslow depth and ulceration are actually captured — and adjusting the Cox model directly for those two variables along with age, race, sex, and site. Why restrict the years like that rather than working with the whole dataset? Because you simply can't adjust for a variable that wasn't recorded before twenty ten; that's a hard data-availability constraint, not a stylistic choice.

Their result: in univariate analysis, amelanotic melanoma still showed markedly worse disease-specific survival, more than a two-and-a-half-fold increased hazard, clearly significant. But once Breslow depth and ulceration were added to the model, that difference disappeared — the adjusted hazard ratio came in right around parity, and it was not statistically significant. They then stratified by AJCC eighth edition T-stage and found the same story at every level, including stage T1a, the earliest tumors — amelanotic and melanotic melanomas matched for depth and ulceration had comparable survival, even though survival did numerically worsen across increasing T-stage for both groups in parallel. Their conclusion, stated plainly: amelanotic melanoma's worse real-world survival is very likely explained by delayed diagnosis leading to deeper, more advanced tumors at detection — not by the tumor itself behaving more aggressively once matched for depth. They do flag their own limitation — restricting to twenty ten onward excludes the pre-immunotherapy era — but argue this is unlikely to distort the take-home message.

The original authors, Nguyen and colleagues, then respond, and this is where it gets pedagogically rich, because they push back on methodology rather than conceding the point outright. They note that Breslow depth and ulceration have non-negligible missingness in SEER, over ten percent for both groups, and ask pointedly how Joshi's group handled that — if it was a complete-case analysis, exclusion of missing cases could introduce selection bias that isn't accounted for. They also point out that adjusting only for local tumor characteristics like depth and ulceration, without incorporating regional or distant disease status, risks lumping together patients with very different nodal or metastatic burdens, which could violate the proportional hazards assumption underlying the Cox model itself — a subtle but important statistical caveat. They further note that AJCC eighth edition staging was only introduced in twenty eighteen, so a stage-stratified analysis spanning twenty ten to twenty twenty-two likely required conflating staging criteria across the sixth, seventh, and eighth editions — and T1a and T1b definitions actually changed between the seventh and eighth editions, raising real misclassification risk.

On their own methodological choices, Nguyen's group explains — this is stated rationale, not our inference — that they deliberately excluded Breslow depth and ulceration from their multivariable model to avoid multicollinearity with the SEER combined stage variable they'd already included, and that they intentionally kept their study window at two thousand through twenty twenty-one specifically to capture long-term trends spanning the pre- and post-immune-checkpoint-inhibitor eras, which the shorter twenty ten-onward window would sacrifice. They acknowledge the SEER staging inconsistency Joshi's group raised but note the previously reported misclassification rate is only around two and a half percent in a huge dataset, and that such errors tend to run in both directions and roughly cancel out in aggregate. Both sides ultimately converge on the same clinical message despite their statistical disagreement: amelanotic melanoma's worse outcomes are most plausibly a story of delayed detection rather than intrinsically worse tumor biology, and both explicitly caution that SEER-based work like this is hypothesis-generating, not definitive — true biologic differences would require prospective, molecularly annotated studies neither database can offer.

The practical takeaway from this whole exchange, for those of us doing skin exams and counseling patients on amelanotic lesions: treat amelanotic melanoma with the same index of suspicion and same urgency as pigmented melanoma once matched for depth — there's converging, though not definitive, evidence that its worse population-level survival reflects that these lesions are simply harder to catch early clinically, not that they're biologically more lethant stage-for-stage. That reinforces exactly what most of us already practice — a low threshold for biopsying atypical pink or flesh-colored lesions — but it's genuinely useful to have the data trend now pointing consistently in that direction across two independent reanalyses.

Last article is a brief report, a retrospective National Cancer Database study asking a very practical question for anyone managing pleomorphic dermal sarcoma: does adjuvant radiation therapy after resection actually improve survival?

Some framing first — pleomorphic dermal sarcoma is an aggressive mesenchymal tumor, typically on sun-damaged skin of elderly men, and it's rare enough that management guidelines, particularly around adjuvant radiation, have never been well defined. Reported local recurrence rates run as high as twenty-eight percent and distant metastasis up to twenty percent in prior retrospective series, so this is not a trivial diagnosis to under-treat. Surgical resection with margin control is unquestionably standard of care; the open question is what to do afterward, especially in high-risk resections.

The authors used the National Cancer Database — a large administrative registry — which is the sensible design choice here given the tumor's rarity; you simply can't power a randomized trial on a disease this uncommon, so a large national registry is really the only way to get numbers big enough to look at survival differences with any statistical power, even though it forces you to give up granularity elsewhere, which the authors are candid about later on.

They identified just over fourteen hundred patients with resected pleomorphic dermal sarcoma across a sixteen-year span, and only about fifteen percent received adjuvant radiation. Here's the key confounding pattern to hold onto: the radiation group wasn't a random slice of the population — they had meaningfully higher-risk disease across the board. Larger tumors over two centimeters were nearly twice as common in the radiation group, lymphovascular invasion was about five times more frequent, nodal involvement was roughly ten-fold higher, positive margins were nearly three times more common, and chemotherapy use was substantially higher too. In other words, clinicians were already selectively directing radiation toward the sicker, higher-risk patients — exactly the scenario that makes a simple radiation-versus-no-radiation comparison prone to confounding by indication.

On survival analysis, negative margins mattered enormously — median overall survival of about eighty-two months with clear margins versus about forty-four months with positive margins, a clinically substantial and statistically significant difference. But adjuvant radiation itself showed no survival benefit — median survival was essentially the same whether or not patients received it, and that held true even in the subgroup with microscopic residual disease after resection, the exact population you'd most expect radiation to help. On multivariable Cox regression, the independent predictors of worse survival were older age, higher Charlson comorbidity burden, tumor size over two centimeters, and positive margins — radiation didn't make the cut as a significant predictor either way.

The authors' own limitations are worth relaying directly: this is retrospective registry data, so there's no recurrence data captured at all — survival, not local control, is the only outcome available — and the registry doesn't capture radiation-specific details like dose or field design, so it is possible some undetected subgroup or radiation regimen could still matter. They're also honest that residual confounding by indication is likely, given radiation was so clearly preferentially used in higher-risk cases, and multivariable adjustment can only correct for the variables you actually have, not unmeasured differences in clinical gestalt that drove the original treatment decision.

Practically, this is a genuinely useful, if not fully practice-changing, data point for how we counsel patients post-resection. The clearest actionable message is the one that was already true — margin control is what drives survival, reinforcing the rationale for margin-controlled approaches like Mohs in appropriate cases — and this study doesn't resolve the standing question of optimal margin size or wide local excision versus Mohs for this tumor, which the authors explicitly flag as unresolved. What it does support is resisting a reflexive move toward adjuvant radiation in high-risk resected pleomorphic dermal sarcoma, including R1 resections, purely on a population-level survival basis — it should be considered selectively, case by case, for things like palliation or function preservation, rather than routinely recommended. This is more "reassuring negative data to individualize discussions" than "here's a new treatment algorithm," and definitive guidance will still need prospective data neither this study nor any registry can fully deliver.

That wraps our four articles for this issue. Taken together, there's a nice thread running through today's episode — real skepticism about how much we should trust a tool, a database, or a treatment reflex without interrogating the confounding underneath it, whether that's AI training data, SEER staging variables, or selection bias in who gets radiation. Thanks for listening, and we'll see you next month.