Welcome back to the journal review. This is the September twenty twenty-six issue of the Journal of the American Academy of Dermatology, and we've got four pieces worth your time this episode — an AI primer, a big multicenter cohort study on squamous cell carcinoma recurrence timing that has real implications for how you schedule follow-up, a brief report on keratoacanthoma versus squamous cell carcinoma discordance, and a methods-focused letter that's essentially a call-out of sloppy TriNetX research. Let's get into it.

First up is a brief report, really more of a primer or perspective piece, titled "AI agents assemble: the new age of AI agentic workflows and their clinical applications," out of Mayo Clinic. There's no methodology or results section here — this is conceptual, so I'll walk you through the framework rather than forcing it into a study template.

The core idea is distinguishing agentic AI from the large language model tools you're already used to. In a traditional nonagentic workflow, you prompt a single model like ChatGPT, it gives you an output, and you iterate back and forth manually. Agentic AI changes the architecture: the large language model acts more like a coach or coordinator, breaking a complex goal into subtasks and delegating each one to other specialized models or tools — an electronic health record query tool, a computer vision model for image interpretation, a separate evidence-based diagnostic model — and then synthesizing their outputs into a final product, largely without you in the loop for each step. The authors give a concrete dermatology example: you ask an agent to review your clinic schedule and flag which patients are likely to need a biopsy. The agentic system pulls chart data through an electronic health record tool, runs uploaded photos through a computer vision model, and feeds it all through a diagnostic large language model to generate a ranked differential with biopsy-yield predictions — all chained together automatically.

The article catalogs real-world examples, both academic and commercial. On the academic side: an agentic system that combines vision transformers reading histopathology biomarkers, literature search, and radiologic segmentation to support oncology decision-making; a "virtual laboratory" where an large language model acts as principal investigator, spinning up other agent "employees" to independently design SARS-CoV-2 nanobodies; and a vision-language model that both generates reports and predicts outcomes from CT pulmonary angiograms. On the industry side, they mention Epic's Penny for healthcare financial workflows, Microsoft's AI copilot framework, One Medical's patient-facing scheduling and prescription assistant, and a research-oriented tool from Castor and Google for end-to-end clinical trial conduct.

The authors' bottom line, and I think this is the practically useful takeaway for you: agentic systems are still tool-limited and task-specific rather than generalizable, they are not infallible, and their role right now should be as a supervised teammate — not an autonomous decision-maker — meaning anything agentic AI produces in your workflow, whether it's triage, prior authorization drafting, or biopsy-yield prediction, still needs human review before it touches patient care. This is interesting-not-yet-actionable territory for your practice today, but worth watching, especially as agentic tools start appearing bundled into your electronic health record vendor's offerings.

Next, the piece I'd call the clinical meat of this issue: a multicenter retrospective cohort study, "Prognostic determinants of early and late relapse in cutaneous squamous cell carcinoma," spanning twelve centers and just over twenty-three thousand cases.

The background problem is one you already know intuitively but that guidelines haven't operationalized well: cutaneous squamous cell carcinoma recurrence risk factors are well established — lymphovascular invasion, large-caliber and deep perineural invasion, poor differentiation, invasion beyond subcutaneous fat, and so on — but nobody has systematically asked which factors predict when a recurrence happens, not just whether it happens. Current follow-up guidelines just bucket tumors into high- versus low-risk and don't meaningfully risk-adjust the follow-up schedule over time, even though roughly half of all recurrences and disease-specific deaths happen within the first year.

Methodologically, this is a retrospective cohort with a clever prespecified design choice: the authors defined early recurrence as within four months, late recurrence as beyond twenty-four months, and intermediate in between — and critically, they set these cutoffs before running any analysis, specifically to avoid data-driven, post hoc threshold-shopping. That's a good practice to flag for your own reading of these papers generally: prespecified windows protect against a well-known bias where you'd otherwise tune your cutoffs to whatever produces the prettiest hazard ratio. They excluded nonsurgical management, adjuvant therapy recipients, and anyone lost to follow-up, and ended up with six hundred sixty-nine recurrent cases against a stratified random sample of non-recurrent controls at a three-to-one ratio.

The statistical workhorse here is a competing-risk regression generating subdistribution hazard ratios, and it's worth spending a moment on why that's the right tool rather than a standard Cox model. In this population, patients can experience local recurrence, regional metastasis, or die from something unrelated to their skin cancer — and death from another cause functionally removes them from being able to ever recur. A standard Cox model would just censor those competing deaths as if the patients were still at risk of recurrence, which inflates the apparent recurrence risk. A competing-risks framework, using subdistribution hazard ratios via the Fine-Gray approach here, keeps those competing-event patients in the risk set in a way that gives you a cumulative incidence estimate that actually reflects real-world probability of recurrence, accounting for the fact that some people die of other things first. On top of that, because this is a twelve-center pooled cohort, they used cluster-robust sandwich variance estimation with center as the clustering variable — essentially acknowledging that patients within the same center are likely to be more similar to each other than to patients elsewhere, so treating every patient as fully independent would understate the true uncertainty. That's a smart, underused adjustment in multicenter dermatology cohorts.

Now the results, and this is where it gets clinically actionable. For early locoregional metastasis — within four months — lymphovascular invasion carried roughly a three-fold increased hazard, and large-caliber deep perineural invasion carried a bit over a two-fold increased hazard, both statistically significant. But here's the key finding: this elevated risk from these two features was front-loaded and lost its statistical punch over time — meaning these are features that predict metastasis that shows up fast, not necessarily disease that recurs late. Poor differentiation behaved completely differently: it carried more than a three-fold increased hazard for early metastasis, dipped to a smaller but still significant effect in the intermediate window, and then came roaring back to over a four-fold increased hazard for late recurrence beyond two years. So poor differentiation is a durable, long-tail risk factor, while lymphovascular and large-caliber perineural invasion are early-warning signals that don't persist. Tumor diameter of two centimeters or more and head-and-neck location showed up as independent predictors across both early and late windows fairly consistently, suggesting these are just generally sticky risk factors regardless of timing.

For discussion and limitations, the authors are appropriately measured — this is a retrospective design, so residual confounding and variable ascertainment across twelve different centers over what's presumably a long accrual period are real concerns, and they ran sensitivity analyses shifting the time windows, which the manuscript references as available in supplementary material, though the given text doesn't detail whether those held up. The obvious generalizability caveat: this is a pooled international cohort, so practice patterns, biopsy techniques, and depth of pathologic reporting for things like perineural invasion caliber may vary center to center.

Practically, here's what I'd call genuinely practice-relevant: if you're staging a squamous cell carcinoma with lymphovascular invasion or large-caliber deep perineural invasion, this data supports considering baseline nodal staging at diagnosis rather than waiting, because that risk is front-loaded and detectable early — waiting to see if it shows up late doesn't fit the biology this paper describes. Conversely, if you have a poorly differentiated tumor, this is your argument for extending surveillance well beyond the typical two-year drop-off point that most current schedules default to, since that risk factor's hazard actually increases again after two years rather than fading. This won't rewrite NCCN guidelines tomorrow, but it's a solid evidence base for individualizing your own follow-up intervals by histologic feature rather than a flat high-risk-versus-low-risk binary.

Third, a brief report: "Risk factors associated with clinical-histopathological discordance in keratoacanthoma diagnosis," a single-center retrospective study out of Dartmouth spanning four years and four hundred seventy-seven biopsies where keratoacanthoma was in the clinical differential.

The clinical gap is one you deal with constantly — keratoacanthoma is managed as well-differentiated squamous cell carcinoma because of the risk of underdiagnosis, and some centers do same-day biopsy with electrodessication and curettage for clinically suspected keratoacanthoma for efficiency, which becomes a problem if the pathology comes back showing a higher-grade squamous cell carcinoma. This study asks which clinical features predict that discordance.

Methodologically it's straightforward: retrospective chart review, concordant defined as keratoacanthoma or well-differentiated squamous cell carcinoma on pathology, discordant defined as anything worse — moderately or poorly differentiated. They used univariate logistic regression to generate odds ratios for discordance, which is the appropriate first-pass tool here given a binary outcome and multiple candidate categorical predictors, though it's worth noting this is univariate, not multivariable — so these risk factors haven't been adjusted against each other, and some of that apparent independent effect could be confounded, say, by head-and-neck location correlating with male sex or with prior skin cancer history.

The numbers: of tumors clinically called keratoacanthoma, only about one in four were actually confirmed keratoacanthoma on pathology, just under half were well-differentiated squamous cell carcinoma, and just over one in four — the discordant group — turned out to be moderately or poorly differentiated disease. Risk factors for discordance were head-and-neck location, carrying roughly a four-fold increased odds; male sex, roughly doubling the odds; a history of basal cell carcinoma, also roughly doubling the odds; and a history of two or more prior squamous cell carcinomas, again roughly doubling the odds. Lower extremity location was protective, more likely to be concordant. Age and tumor diameter differences reached statistical significance but the authors themselves note these weren't clinically meaningful — a good example of exactly the kind of significant-but-trivial finding worth flagging rather than over-interpreting. Immunosuppression status and whether a dermatology-trained provider made the initial call were not significant risk factors, which is somewhat reassuring — this isn't simply about generalist misdiagnosis.

Clinically, recurrence was about three-fold higher in the discordant group compared to concordant, and there was one case of metastasis, occurring in the discordant group. Among tumors initially managed with electrodessication and curettage, about one in four had discordant pathology, and of those, roughly a third needed additional treatment after the fact.

Limitations are honestly stated: single-site, retrospective, with inherent interobserver variability in dermatopathology grading, and the authors can't fully capture clinicians' unmeasured gestalt for why they called something keratoacanthoma in the first place.

The practical takeaway is directly actionable at the biopsy-decision level: if you have a clinically suspected keratoacanthoma on the head or neck, in a male patient, with a history of basal cell carcinoma or multiple prior squamous cell carcinomas, this is a lower-threshold-for-full-biopsy-before-empiric-treatment scenario. Don't let a same-day electrodessication and curettage approach lull you into skipping adequate sampling in these higher-risk demographic and anatomic contexts, because a meaningful chunk of your presumed keratoacanthomas in that subgroup will actually be higher-grade disease requiring a different treatment plan.

Last, a letter to the editor, "Toward higher methodologic standards for TriNetX studies in dermatology." This isn't a study, it's a methodologic critique, so I'll treat it as the argument piece it is rather than force results and limitations onto it.

The author's concern is the explosion of TriNetX-based observational research in dermatology — TriNetX being the federated electronic health record network that lets you rapidly build matched cohorts across health systems. His point is that dermatology-specific data in these platforms is often missing exactly the things that matter clinically: severity, morphology, body-site distribution, adherence, concomitant topical or over-the-counter therapy, and the reasoning behind a given treatment choice. Few of these code-based phenotype definitions have ever been validated against actual clinical records, so the cohort itself may be poorly defined before any statistics even happen.

His central methodologic argument, and it's a genuinely useful teaching point regardless of which paper you're reading next: propensity score matching is often treated as a stamp of rigor, but it can only balance confounders that were actually measured and entered into the model. It cannot fix an index-event definition that's flawed, it cannot fix a badly chosen comparator group, and it cannot rescue you from immortal time bias or prevalent-user bias. He gives a concrete example — comparing two drugs by first prescription date without a new-user design or washout period. If one drug is preferentially started in sicker patients, that's channeling bias, and no amount of covariate matching afterward fixes it, because the bias is baked into how the cohort was assembled, not into which covariates you balanced.

His practical recommendation is that dermatology should hold TriNetX studies to the same standard as any other observational study: a defensible index event, a justified comparator, transparent baseline and follow-up windows, transparent exposure and outcome definitions, and causal language reserved only for designs that can actually support it — he points to the RECORD reporting guideline for routinely collected health data as a useful framework.

This is squarely a "read critically" takeaway rather than a practice-changing one for your own patient care — but it's directly relevant to how you should be evaluating the flood of TriNetX-based papers you're going to keep seeing cross your desk. Next time you see a slick, large-n TriNetX comparison with propensity matching and a confident causal conclusion, the questions to ask are: was this a new-user design, is the comparator appropriate, and does the phenotype definition actually reflect the dermatologic entity being studied — because sample size and matching alone don't make it valid.

That wraps our four articles this month — an AI framework worth watching but not yet acting on, a squamous cell carcinoma recurrence-timing study with genuinely actionable implications for staging and follow-up scheduling, a keratoacanthoma discordance study that should lower your threshold for full biopsy in higher-risk demographics, and a methodologic gut-check on how to read big-data dermatology studies going forward. Thanks for listening, and I'll see you next issue.