Welcome back to the journal review. This is our look at the Journal of the American Academy of Dermatology, April twenty twenty-six issue, and we've got four pieces on the docket today — a letter to the editor on antidepressants and surgical wound healing, a brief report proposing a new predictive algorithm for sentinel lymph node biopsy in melanoma, a letter on the ethics of federated learning and AI model selection in dermatology, and a research letter running a pharmacovigilance disproportionality analysis on nirogacestat. Let's get into it.

First up is a letter to the editor, and it's a really nice piece of methodological pushback worth sitting with. It's responding to a propensity-matched cohort study, also published in JAAD, that looked at antidepressant use and postoperative outcomes after basal cell carcinoma surgery — that original study found antidepressant users had higher complication rates and called for more research into the biology of impaired healing on these drugs. The letter writer here is a psychiatrist, not a dermatologist, and his concern isn't with the statistics or the matching technique — he's explicit that he has no quarrel with the methodology itself. His concern is about causal inference and downstream clinical messaging. His core argument: the exposure variable in that original study was antidepressant prescription, not antidepressant adherence. And he cites adherence data suggesting real-world compliance with antidepressants can run as low as fifty percent. So what the original study can actually tell us is that patients who were prescribed antidepressants had more complications — not that patients who were actually taking them, at therapeutic levels, had more complications. That's a meaningful gap. He also raises confounding by indication — clinicians are going to preferentially prescribe antidepressants to patients with more severe depression, and severe untreated or undertreated depression is independently linked to worse postoperative outcomes. So flip the causal arrow: instead of the drug impairing healing, it could be that inadequately treated depression is the actual driver, with antidepressant prescription simply being a marker of that underlying severity rather than a mediator of harm. If that's right, the clinical implication inverts completely — instead of surgeons weighing whether to hold or adjust antidepressants perioperatively, the actual actionable move would be optimizing psychiatric treatment, not modifying it. There's no new data here, no methods section to walk through — this is a conceptual and epistemological corrective. The practical takeaway for you as a Mohs surgeon: this is a good reminder not to over-index on that original propensity-matched study when counseling patients or making perioperative medication decisions. Don't stop or taper antidepressants based on that data alone, and don't let this trickle into patient-facing conversations as settled biology. It's association, not mechanism, and the letter's psychiatrist author is basically asking the dermatology community to hold the causal claim loosely until adherence and depression-severity data are actually incorporated.

Next is a brief report, and this one is a genuine methods-and-results piece, even though it's short — it's proposing an alternative statistical framework for predicting sentinel lymph node biopsy positivity in melanoma. The clinical problem is a familiar one: selecting which melanoma patients should actually undergo sentinel node biopsy, and existing tools — most notably the Melanoma Institute Australia, or MIA, nomogram — already do this using six factors, but the authors think they can do better with a technique they call patient-centered methodology, or PCM, which they've previously published on in other cancer contexts.

Methodologically, they pulled a large dataset from the National Cancer Database, over thirty-six thousand melanoma patients diagnosed in twenty eighteen and twenty nineteen, and also worked with a complete-case subsample of about nineteen thousand patients who had no missing data across eight prognostic factors — things like Breslow depth, ulceration, mitotic rate, lymphatic invasion, histologic subtype, age, sex, and anatomic site. Why build both a full dataset algorithm and a complete-case algorithm? The authors don't spell out their full reasoning, but the logical inference is that real-world charts frequently have missing histologic data — mitotic rate or lymphatic invasion status isn't always recorded — so a tool that only works when every field is populated would be clinically limited. Building a parallel algorithm off the full dataset, missing data and all, lets them generate a usable estimate even when the input is incomplete, which is a meaningful practical design choice for a tool meant for bedside or clinic use.

The other interesting methodological wrinkle is how they handled thick tumors. When they ran the existing MIA six-factor model on the complete-case cohort, it produced an area under the curve — essentially a discrimination score where higher is better, one being perfect — of about zero point seven three. Excluding patients with tumors thicker than ten millimeters bumped that up slightly, to about zero point seven four, but the authors flag this as a bit of a statistical sleight of hand, because excluding those thick-tumor patients actually inflated the predicted probability of node positivity for the remaining thick-tumor patients by around thirty-nine percent — meaning the model's apparent improvement came partly at the cost of distorting risk estimates for exactly the patients where accuracy matters most.

On to the actual results: applying their PCM approach to the same six factors on that complete-case cohort improved discrimination to about zero point seven six, and adding two more factors — sex and anatomic location — nudged it further to around zero point seven six as well, essentially the same ceiling. In head-to-head, patient-level comparisons, the PCM algorithm's individual predictions were more accurate than the MIA model's about forty percent more often, and that difference was statistically significant. Calibration — meaning how well the predicted number of node-positive patients matched the actual observed number — was excellent, with an R-squared close to one. They also did the sensible validation step of splitting the complete-case cohort into training and validation halves and got essentially identical discrimination in both, which supports the model not being overfit to a single sample. And critically, running the same approach on the full dataset — including patients with missing fields — still produced comparable discrimination, which supports their central selling point: usability despite missing data.

One more piece worth flagging: their risk-stratification step, which grouped patients into low, intermediate, and high risk using just tumor thickness and age, showed sentinel node positivity rates of about eight percent, nineteen percent, and thirty-eight percent respectively across those tiers — a genuinely clinically meaningful spread. And interestingly, the dominant predictive factor differed by risk tier: mitotic rate mattered most in the low-risk group, while lymphatic invasion took over as the top factor in the intermediate- and high-risk groups, which is a subtle but clinically intriguing point about effect modification across risk strata.

Limitations here are mostly about scope rather than internal validity — this is registry data, so it inherits all of the National Cancer Database's usual issues: no central pathology review, variable data completeness, and a floor effect since NCDB itself restricts analysis to tumors at least half a millimeter thick, so this doesn't help with the thinnest melanoma decision-making. The authors have made this into a live interactive online calculator, which is a nice translational touch. Practically speaking, this is interesting and represents real incremental improvement in discrimination over existing nomograms, but I'd call it useful decision-support rather than practice-changing — it doesn't upend national sentinel node biopsy guidelines, it refines the probability estimate you'd already be discussing with a patient, particularly in that ambiguous intermediate-risk zone.

Third, a letter to the editor on artificial intelligence ethics in dermatology, responding to an earlier piece on AI and pediatric dermatology consent issues. This is a conceptual, forward-looking letter with no original data — think of it as a set of technical and policy recommendations layered onto an ethics discussion. The authors, one with an AI and public health background, argue for three main things. First, model selection should match the data type — classical algorithms like logistic regression or random forests for structured, tabular clinical data, deep neural networks for unstructured data like images or free text — which is really just a call for methodological literacy so dermatology researchers don't reach for a sledgehammer when a scalpel will do. Second, they push multi-institutional collaboration toward federated learning approaches, where the model moves between institutions rather than the raw patient data — citing real precedents like the MELLODDY drug-discovery consortium and the FeTS brain tumor segmentation challenge — plus newer decentralized variants like blockchain-based federated learning and swarm learning, which distribute trust across peer nodes rather than a single central server. Third, they advocate layering in differential privacy, which adds statistical noise to protect individual-level data, or homomorphic encryption, which allows computation on data while it stays encrypted. None of this is tested in a dermatology dataset here — it's a set of best-practice recommendations, endorsed by citing standards bodies like the US National Institute of Standards and Technology and the UK National Cyber Security Centre. The letter closes with a public-administration angle: multilevel education, from training researchers to embed privacy protections at the design stage, to giving clinicians structured communication tools for discussing AI with patients, to community-level outreach. For practicing purposes, there's nothing actionable at the patient-care level here — this is squarely an interesting-not-yet-actionable piece, but if you're involved in any multi-site AI or imaging-database research, it's a reasonable primer on the vocabulary and options you should know you have before your data-sharing agreement gets drafted.

Last, a research letter presenting a disproportionality analysis of dermatologic adverse events linked to nirogacestat — this is a pharmacovigilance study, structurally more substantial than a typical letter, building directly on a prior case series of twenty-eight patients with nirogacestat-associated skin toxicity. Nirogacestat is the first-in-class oral gamma secretase inhibitor, FDA-approved in late twenty twenty-three for progressing desmoid tumors, and because it's new and desmoid tumors are rare, real-world cutaneous safety data have been thin.

Methodologically, the authors mined the FDA Adverse Event Reporting System, or FAERS, from the drug's approval through mid-twenty twenty-five, standardized the terminology, deduplicated cases, and restricted their analysis to monotherapy reports with a single dermatologic adverse event — a sensible move to reduce confounding from concomitant drugs or overlapping toxicities. They then ran four separate disproportionality metrics in parallel — reporting odds ratio, proportional reporting ratio, empirical Bayes geometric mean, and information component — which is standard practice in pharmacovigilance precisely because any single metric can throw a false positive signal, so requiring agreement across all four is the field's way of building confidence into what is otherwise fairly noisy spontaneous-reporting data.

The results: rash, alopecia, and dry skin were the most frequently reported events, each with a positive signal across all four metrics and reporting odds ratios in the roughly three-and-a-half to five-fold range — statistically real signals, though not shockingly large effect sizes. The more clinically striking finding is the follicular pattern — alopecia, hidradenitis, acneiform dermatitis, and papules all showed disproportionately elevated signals, and the reporting odds ratio for hidradenitis specifically was strikingly high, well over forty-fold, which is a genuinely large signal even accounting for the play in these metrics with small case counts. This clusters mechanistically with the drug's known biology — gamma secretase inhibition affects NOTCH signaling, and NOTCH pathway disruption is already implicated in follicular biology, including genetic hidradenitis suppurativa. So this isn't just numbers matching statistical thresholds — there's a coherent, biologically plausible story tying the signal to mechanism, which is exactly the kind of corroboration that makes a disproportionality signal more trustworthy.

Timing matters clinically here too: about eighty percent of these positive-signal events occurred within the first month of treatment, but hidradenitis was the outlier, with a delayed onset — most cases arising after around day one hundred, with a mean onset closer to two hundred eighty days. That's a genuinely actionable clinical detail: early rash and dry skin surveillance in month one, but a much longer tail of vigilance specifically for hidradenitis-type follicular disease, potentially out to six to nine months or beyond. Age and sex subgroup analysis, using the information component metric because it's more stable with fewer data points, showed rash signals concentrated in adults, while alopecia, hidradenitis, and papules trended higher in pediatric and elderly subgroups despite smaller absolute case counts, and dry skin was more frequently reported in women.

Limitations are the ones inherent to any FAERS-based disproportionality study — this is spontaneous reporting data, so there's no denominator of total drug exposure, meaning you get relative signal strength, not true incidence; there's reporting bias, since a paper drawing clinical attention to a toxicity itself can drive more reporting of it; and the authors themselves note that because nirogacestat is newly approved with a still-limited exposed population, some of the lower confidence bounds around the information component were unstable or negative. This is hypothesis-generating pharmacovigilance, not a controlled causality study.

Practically, this is genuinely useful for anyone likely to see gamma secretase inhibitor patients referred in for skin findings, which increasingly means dermatologic oncology and general dermatology given how these drugs are expanding in solid tumor and desmoid indications. The actionable takeaway is straightforward: counsel patients up front about early rash, dry skin, and hair thinning within the first month, and — more importantly — maintain a long surveillance window for follicular occlusive disease, specifically hidradenitis-pattern lesions, which can present many months into therapy and might otherwise be misattributed rather than linked back to the drug.

That wraps our four pieces for this issue — a cautionary methodological letter on antidepressants and wound healing, a refined statistical tool for sentinel node prediction in melanoma, a forward-looking ethics letter on AI model architecture and privacy, and a mechanistically grounded pharmacovigilance signal on nirogacestat's follicular toxicity. Thanks for listening, and I'll see you next issue.