Welcome to this June twenty twenty-six review of JAMA Dermatology. Four pieces on the docket this time: a survey study on an AI-powered consumer dermatology tool, a brief report on cutaneous eruptions tied to lifileucel and interleukin-2 in metastatic melanoma, a prospective diagnostic accuracy study comparing physical exam, ultrasound, and CT for nodal staging in high-risk cutaneous squamous cell carcinoma, and finally a patient page on sunscreen that's worth knowing about for your own patient handouts. Let's get into it.

First up is an original investigation out of Google Research and Stanford, titled "Consumer Understanding of Skin Concerns With an AI-Powered Informational Tool." This is a between-participants randomized survey study, not a clinical trial, so keep that framing in mind throughout.

The background here is the familiar access gap — only about a quarter to a third of skin conditions actually get seen by a dermatologist, and direct-to-consumer AI tools are being pitched as a bridge. The specific gap this paper addresses is subtler than "does the AI work" — it's asking whether giving consumers AI-generated differential diagnoses actually changes their comprehension and their next-step decision-making, compared to what they'd do on their own with a web search.

Methodologically, this is a nicely designed three-arm study, and the middle arm is the clever part. Participants were randomized to a control group using their usual resources like web search, an AI arm where they saw a prototype model's top three to seven predicted conditions with textbook images and descriptions, and a Wizard-of-Oz arm with the identical interface but where the "predictions" were actually the ground-truth differential from a dermatologist panel — essentially simulating a hypothetically perfect AI. The rationale for that third arm is explicit and it's methodologically elegant: it lets you separate two things that would otherwise be conflated — the effect of the interface and framework itself, versus the effect of prediction accuracy. If Wizard-of-Oz outperforms real AI, that tells you the model still has headroom; if AI and Wizard-of-Oz perform similarly, the interface itself is doing the work. Participants looked at retrospective, deidentified teledermatology cases with images and structured history and were asked to imagine the case as their own skin issue, then report whether they could name the condition, what they thought the next step should be, and their confidence and satisfaction.

On results: compared to control, where about four in ten participants were willing to venture a condition name, willingness to name a condition jumped to roughly six in ten in both the AI and Wizard-of-Oz arms — a clear, significant increase, but notice it's about willingness, not necessarily correctness. Accuracy is where it gets more interesting. Control-arm accuracy was low, under one in ten. The AI arm roughly tripled that to about one in five correct. But the Wizard-of-Oz arm — the perfect-prediction condition — did meaningfully better still, at roughly one in three correct. So there's a real dose-response relationship between prediction quality and consumer comprehension, and even with perfect ground-truth differentials shown to them, two-thirds of consumers still couldn't correctly name their own condition. That's a sobering ceiling. For next-step accuracy — whether participants correctly identified what they should actually do, like seek urgent care versus watchful waiting — only the Wizard-of-Oz arm showed a significant improvement over control, and numerically it was a modest bump, from about six in ten to about six in ten as well, just barely significant. The real AI arm did not significantly improve next-step accuracy at all.

The authors' discussion is appropriately measured: AI assistance helped people engage and correctly name conditions, but the gain was bounded by how accurate the presented differential was, and even under best-case conditions, next-step guidance didn't reliably improve. They interpret this as pointing to a design problem — the condition information architecture itself may need rework, separate from model accuracy. Limitations worth flagging for your own read: this used retrospective vignettes, not real patients making real decisions about their own skin, so real-world anxiety, stakes, and image-capture quality aren't captured; the panel was a commercial survey population, not necessarily representative of who would actually download a consumer skin app; and the model tested is a research prototype, not a shipping product.

Practical takeaway: this is interesting, not practice-changing. It doesn't change what you do in clinic tomorrow, but it's a useful data point if patients start showing up having used consumer AI tools — expect them to arrive more willing to name a diagnosis and more confident, but not necessarily more correct about next steps, meaning your triage conversation is still essential and shouldn't be short-circuited by what the patient's app told them.

Second article, a brief report: "Cutaneous Eruptions and Lifileucel/Interleukin 2 in Individuals With Metastatic Melanoma," from the Mass General Brigham and Dana-Farber group. This is a retrospective single-network cohort study, and it's a nice example of dermatology finding a signal buried inside an oncology treatment pathway.

Background: lifileucel is the first FDA-approved autologous tumor-infiltrating lymphocyte therapy for melanoma progressing after PD-1 blockade or BRAF-targeted therapy. In the pivotal trial, "rash" was noted in roughly a third of patients, but never clinically or histopathologically characterized — dermatology simply hadn't looked closely at it yet. That's the gap.

Methods: they retrospectively pulled every patient treated with lifileucel outside of clinical trials at their institutions, abstracting eruption features, photographs, dermatopathology, and correlating eruption occurrence with RECIST-defined radiographic response at three time points — around one month, around six weeks, which they designated as the primary endpoint, and around three months. The choice of the six-week mark as primary, rather than the more commonly performed one-month restaging, is explicitly justified by the authors: median time to initial response to lifileucel has been reported at about six weeks, and that timepoint was also a biomarker assessment point in the pivotal trial — so they're aligning with when a treatment effect would actually be expected to show up radiographically, not with local practice patterns for convenience. They ran an unadjusted logistic regression of response against eruption occurrence, then progressively adjusted for interleukin-2 dose, demographics, and melanoma-specific prognostic factors like M stage, LDH, and prior lines of therapy — a reasonable approach on forty-four patients where you can't do much more sophisticated modeling, and layering adjustments this way lets the reader see whether the association survives each added confounder rather than hiding behind one fully-adjusted black box.

Results: half of the forty-four patients developed a cutaneous eruption during hospitalization, appearing at a median of four days post-infusion. Where photographs were available, the picture was fairly consistent — central-predominant, often purpuric, morbilliform eruptions, generally mild, and notably none were severe enough to be classified as a severe cutaneous adverse reaction, and none limited interleukin-2 dosing. All had resolved or were improving by discharge, with most fully resolved within about five days after that. The clinically important finding is the association with response: at the six-week mark, patients who developed the eruption had an objective response rate of roughly two-thirds, versus about one in five among those without an eruption — a substantial and statistically significant difference. That association held up as an odds ratio in the range of seven-to-one to nearly twelve-to-one depending on which covariates were included, and importantly it survived adjustment for interleukin-2 dose and for melanoma-specific prognostic factors, meaning the signal isn't simply being driven by sicker or more heavily pretreated patients responding differently. Patients who developed eruptions did receive more interleukin-2 doses on average, but higher interleukin-2 dosing alone didn't track with response, so the eruption itself — not just the dose — appears to be the meaningful signal. By the three-month mark, the association weakened and was no longer statistically significant, though the confidence intervals were wide given the small numbers.

The authors' interpretation is that this represents a peritreatment efficacy biomarker — something visible on the ward days into hospitalization, well before the conventional six-week restaging scan. Limitations they and you should keep front of mine: this is forty-four patients at essentially one network, photographic documentation was only available for fourteen of the twenty-two eruption cases so the morphologic characterization is based on a subset, there's no central dermatopathology review panel described beyond a representative case, and with this sample size the confidence intervals on those odds ratios are wide enough that the point estimates shouldn't be over-read.

Practical takeaway: this is genuinely interesting and plausibly practice-relevant for those of you doing inpatient oncodermatology consults — a purpuric morbilliform eruption appearing about four days after TIL infusion in a lifileucel patient should probably be read as a reassuring sign rather than purely a toxicity to treat and move past, and it's reasonable to counsel the oncology team that this may be an early favorable indicator, manageable with emollients and mild topical steroids alone. It's not yet practice-changing in the sense of altering restaging protocols, but it's the kind of finding worth watching for in your consult notes and worth replication in a larger, multi-institutional cohort.

Third article, and this is a substantial original investigation: "Diagnostic Modalities and Nodal Staging in High-Risk Cutaneous Squamous Cell Carcinoma," the LACUNAS study out of thirteen tertiary dermato-oncology centers in Spain.

The clinical problem is one you all live with — high-risk cutaneous squamous cell carcinoma, meaning Brigham and Women's T2b or T3, or T2a with additional high-risk features, carries meaningful occult nodal metastasis risk, and once nodes are involved, five-year survival drops sharply, into roughly the twenty-to-sixty-five percent range depending on the series. Physical exam alone is known to be an unreliable staging tool, with reported false-negative rates as high as seventeen to fifty percent. Imaging has been proposed as better, but — and this is the actual gap — no prior prospective study had directly compared ultrasound and CT performed simultaneously in the same patients, benchmarked against physical exam, in this specific high-risk population.

Methods: this was a prospective, multicenter, paired diagnostic design — meaning every patient underwent all three modalities, physical exam, high-resolution ultrasound, and contrast-enhanced CT, around the time of surgery, each interpreted blinded to the others' results. The paired design is the key methodological choice here, and it's the right one: because each patient serves as their own comparator across all three tests, you eliminate the confounding you'd get from comparing separate cohorts who happened to get different imaging, and it allows direct concordance statistics like Cohen's kappa between modalities. Reference standard was histopathologic confirmation via fine-needle aspiration or excision for anything suspicious, and for negative baseline studies, a three-month clinical follow-up served as the negative reference. That's a sensible pragmatic reference standard, since you obviously can't ethically excise every clinically and radiographically negative node just to confirm true-negative status.

Results, and here's where the exact numbers matter because they'd change how you counsel and stage patients. Out of a hundred fifty-five patients, twelve, or about eight percent, developed nodal metastases within three months. Sensitivity was highest for ultrasound at roughly sixty-four percent, CT close behind at about fifty-five percent, and physical exam trailing badly at just eight percent. Specificity was high across the board, in the mid-to-high nineties for all three. Ultrasound and CT agreed with each other almost perfectly — a kappa around point-eight-seven — while physical exam's agreement with either imaging modality was poor. Then comes the finding that really matters clinically: when they stratified by immune status, the picture split dramatically. In immunocompetent patients, both ultrasound and CT caught essentially everything — one hundred percent sensitivity, excellent area-under-the-curve around point-nine-eight for both. But in immunosuppressed patients, sensitivity collapsed to roughly twenty percent for ultrasound and under seventeen percent for CT, with area-under-the-curve dropping to essentially coin-flip territory, around point five-five to point five-seven. In other words, in your immunosuppressed patients, a negative baseline ultrasound or CT tells you almost nothing reassuring — metastases in that subgroup often appeared abruptly on follow-up despite clean baseline imaging.

The authors conclude that ultrasound and CT are essentially interchangeable and both clearly superior to physical exam for baseline nodal staging in high-risk cutaneous squamous cell carcinoma overall, but that the profound underperformance in immunosuppressed patients means guidelines need a tailored approach for that subgroup, with heavier reliance on close clinical follow-up rather than false reassurance from a single negative scan. Limitations to weigh: only twelve total metastatic events drove these sensitivity calculations, so the confidence intervals, especially in the immunosuppressed subgroup, are wide — that twenty percent sensitivity figure is built on a small handful of events, and a couple of cases going the other way would move the estimate substantially. This was also a prospective but still observational design without a true independent blinded central adjudication panel described beyond blinding between modalities, and follow-up was capped at three months, so later-emerging metastases beyond that window aren't captured here.

Practical takeaway, and I'd call this the closest thing to practice-changing in today's lineup: for your immunocompetent high-risk cutaneous squamous cell carcinoma patients, either ultrasound or CT is a reliable baseline staging tool, and which one you pick can reasonably come down to availability, cost, and radiation considerations, with ultrasound having practical advantages for older, frailer patients. But for immunosuppressed patients — your transplant recipients and chronic lymphocytic leukemia patients especially — do not let a negative baseline ultrasound or CT lower your guard. This data argues for genuinely close interval clinical follow-up in that subgroup regardless of imaging results, since occult nodal disease seems to emerge unpredictably and imaging just isn't catching it early in that population.

Last, a quick one — the JAMA Dermatology Patient Page, "Sun Protection 101, Your Guide to Sunscreen." This isn't a study, it's a patient education handout, so there's no methods or results to walk through — just a clean rundown you already know cold, but worth having the framing fresh for patient conversations. It covers chemical versus mineral filters and their respective tradeoffs around irritation and white cast, tinted formulations and their added benefit of blocking visible light for patients with dyspigmentation, the various vehicle options and the reminder that the best sunscreen is the one patients will actually use, the SPF explainer — SPF fifteen blocking about ninety-three percent of UVB, SPF thirty about ninety-seven percent, SPF fifty about ninety-eight percent, with the reminder that higher SPF doesn't mean longer-lasting protection — the broad-spectrum labeling point, water-resistance ratings capping at forty or eighty minutes before reapplication is required, the teaspoon-rule application guide, and a brief safety note addressing oxybenzone and reef-safety concerns, concluding there's no conclusive evidence of harm. Useful as a handout to print for your patients, nothing here changes your own clinical practice.

That wraps this episode. To summarize the through-line: consumer AI tools can boost patients' confidence and naming accuracy but not yet their next-step judgment, so don't let that substitute for your triage; a purpuric morbilliform eruption after lifileucel may be a reassuring early efficacy signal worth flagging to your oncology colleagues; ultrasound and CT are solid and interchangeable for nodal staging in high-risk cutaneous squamous cell carcinoma but far less trustworthy in immunosuppressed patients, who need closer clinical follow-up regardless of imaging; and keep that sunscreen patient page handy for your next appointment. Thanks for listening, and I'll see you next month.