Welcome back to the Journal Review, your monthly deep dive for the practicing Mohs surgeon and dermatologic oncologist. This is the May 2026 issue of JAMA Dermatology, and we've got four articles on deck — an HPV genotyping study out of Singapore, a case series on intralesional interleukin-2 for cutaneous squamous cell carcinoma, a systematic review and meta-analysis on AI-assisted melanoma diagnostics, and a large Danish registry study on skin cancer risk in chronic lymphocytic leukemia. Let's get into it.

First up is a brief report — a case-control study titled "Human Papillomavirus Diversity in Asian Individuals With Cutaneous Squamous Cell Carcinoma," out of Singapore General Hospital.

The background here is straightforward: we know HPV, particularly beta-HPV, has been linked to cutaneous squamous cell carcinoma risk, especially in organ transplant recipients. But essentially all of that literature comes from White populations, and it's focused narrowly on beta-HPV, largely ignoring alpha- and gamma-HPV genotypes. Whether that Western paradigm even applies to Asian patients has been an open question, and that's the gap this group set out to fill.

Methodologically, this is a multicenter case-control study running from 2018 to 2024, comparing organ transplant recipients with cutaneous squamous cell carcinoma against immunocompetent patients with cutaneous squamous cell carcinoma. They took three tissue samples per patient where possible — eyebrow hair as a proxy for cutaneous HPV reservoir, the squamous cell carcinoma lesion itself, and prior cutaneous wart biopsies — and ran next-generation sequencing using a custom Illumina panel targeting 193 known HPV types across all three genera. The rationale for sampling three sites rather than just the tumor is smart and worth flagging: eyebrow hair has been used in prior HPV literature as a surrogate for a patient's baseline cutaneous HPV "virome," so by sampling hair, wart, and tumor in the same patient, they could actually ask the concordance question — does the HPV type sitting in your hair follicles show up in your cancer — rather than just cataloging what's present in the tumor alone.

On to results. Eighty-five patients total, 37 transplant recipients and 48 immunocompetent controls, and this population was overwhelmingly Chinese, which is itself the point — this is genuinely new genotype data from an East Asian cohort. As expected, transplant recipients developed their squamous cell carcinomas roughly a decade earlier than immunocompetent patients, medians of 68 versus 80, and they had substantially more viral warts and larger tumors. Across all samples, they identified 65 distinct HPV genotypes — the majority, well over half, were beta-HPV types, with the rest split between gamma and alpha. Beta-HPV was the genus most consistently found across all three tissue sites and carried the highest viral loads overall. Interestingly, viral load patterns diverged by immune status: transplant recipients trended toward higher alpha- and gamma-HPV loads, while immunocompetent patients trended toward higher beta-HPV loads — though none of these differences reached statistical significance.

The most clinically provocative finding is the concordance data. If eyebrow hair genuinely reflected what seeds the tumor, you'd expect a meaningful overlap between hair and paired tumor genotypes. Instead, among the patients with both samples, only about one in eleven — roughly 9 percent — showed even a single concordant genotype between hair and tumor. So the causal story of "your baseline cutaneous HPV reservoir directly seeds your squamous cell carcinoma" is not well supported by this concordance data, at least not in this cohort.

The authors are appropriately measured in their discussion. This is a modest sample, heavily skewed toward one ethnicity, and the genotype spectrum they found — different predominant beta types, low hair-tumor concordance — differs from what's published in White cohorts, suggesting real population-level variation in cutaneous HPV ecology rather than a universal HPV-cSCC genotype signature. They call for larger Asian-specific mapping studies before this can meaningfully inform prevention strategy.

Practical takeaway: this is interesting, hypothesis-generating biology, not practice-changing. It doesn't change how you evaluate or manage a transplant patient's squamous cell carcinoma today. But it's a useful reminder that the beta-HPV paradigm from Western literature may not generalize globally, and it tempers any enthusiasm for eyebrow hair or similar surrogate sampling as a clinical risk-stratification tool — at least until concordance data look a lot stronger than 9 percent.

Next, a brief report case series: "Intralesional Interleukin-2 Therapy for Treatment of Cutaneous Squamous Cell Carcinoma," from a single referral center in Halifax with deep experience in intralesional immunotherapy for melanoma.

The clinical problem is one you know well — recurrent, locally advanced, or anatomically awkward cutaneous squamous cell carcinoma in elderly, comorbid patients where another excision or flap isn't necessarily the best answer. Systemic options like cemiplimab exist but come with real toxicity burden. This group's angle is biologically clever: they note the well-established inverse relationship between IL-2-based immunosuppression and squamous cell carcinoma incidence — transplant recipients on calcineurin inhibitors have up to a 250-fold increased incidence — and reasoned that if suppressing IL-2 signaling raises cSCC risk, then locally restoring IL-2 signaling might have direct antitumor effect. Combined with their existing melanoma experience using intralesional IL-2, with response rates over 90 percent and minimal severe toxicity, they extended the protocol to a selected group of squamous cell carcinoma patients who'd failed surgery or weren't surgical candidates, or who wanted to try nonsurgical treatment first in a cosmetically sensitive site.

Methodologically, this is a retrospective case series, not a trial — the authors don't explicitly justify the design, but it's the obvious choice here: this is a small, heterogeneous, off-label compassionate-use style cohort accumulated over seven years at one center, and a case series is really the only feasible way to first report on a novel indication before anyone commits to a randomized design. Treatment itself was biweekly intralesional injections, dosed by lesion size up to a ceiling, with response assessed clinically and photographically using RECIST-based criteria, and biopsy confirmation when response was ambiguous.

Results: sixteen patients, mean age 79, mostly head and neck and a couple of perianal lesions — genuinely high-risk, sensitive-site disease. Patients received a mean of ten treatments over about thirteen months. The complete response rate was 81 percent — thirteen of sixteen patients — with zero partial responses, meaning responders essentially cleared completely. Progression-free survival across the cohort was around 19 months. Toxicity was limited entirely to grade 1 and 2 events — a self-limited flu-like syndrome of fever, chills, and malaise lasting under a day, occasionally nausea. No grade 3 or higher toxicity in this elderly, comorbid population at all. The three nonresponders all went on to salvage surgery or, in one case, palliative chemotherapy after failed chemoradiation and abdominoperineal resection.

In their discussion, the authors explicitly benchmark against the recent cemiplimab pilot data, where complete response was around 53 percent with an additional 13 percent partial responders, but with grade 3 or higher toxicity in almost one in five patients. Their point, reasonably made, is that in early-stage, non-metastatic disease, intralesional IL-2 achieved a numerically higher complete response rate with essentially no serious toxicity and at a fraction of the cost.

Limitations are honestly stated and important: this is a small, nonrandomized, single-center, retrospective case series with no control arm, so response rates are almost certainly optimistic, there's obvious selection bias in who gets referred and offered this therapy, and it's not directly comparable to the cemiplimab trial population, which included more advanced, potentially unresectable disease.

Practical takeaway — this is genuinely interesting and worth knowing about, but it is not yet practice-changing given the total absence of comparative or randomized data. That said, for your specific patient population — an elderly, comorbid patient with recurrent or anatomically difficult squamous cell carcinoma in a cosmetically sensitive area who isn't a great surgical candidate and where systemic immunotherapy carries too much toxicity risk — this is a reasonable option to know exists and to consider referring for, particularly given the favorable toxicity profile and low cost. Just counsel patients that the evidence base is a sixteen-patient single-center series.

Third, a systematic review and meta-analysis: "Prospective Evidence on Artificial Intelligence-Assisted Melanoma Diagnostics," from the German Cancer Research Center group.

The gap they're addressing is an important one for anyone following the AI-in-dermatology literature: the vast majority of AI melanoma classifier studies are retrospective, run on curated, often enriched image datasets that don't reflect real clinical case mix or real-world uncertainty. That inflates performance estimates and tells you little about how these tools would behave prospectively, in vivo, against an actual histopathologic reference standard. So this group set out to systematically pool only prospective studies comparing dermatologists, AI alone, and AI-assisted dermatologists in dermoscopic melanoma detection.

Methodologically, this followed PRISMA reporting guidelines with a registered protocol, searched four databases through mid-2025, and used pretty strict eligibility criteria — prospective design only, dermoscopic images only, a histopathologic reference standard, and a minimum of twenty histopathologically confirmed melanomas per study to avoid small-sample noise skewing pooled estimates. Risk of bias was assessed with QUADAS-2 and the comparative QUADAS-C tool. From 308 initial records, they landed on eleven eligible studies. Worth teaching here: the choice to restrict to prospective-only studies is the methodological backbone of this whole paper — it's specifically designed to filter out the retrospective, curated-dataset literature that has been criticized for overstating AI performance, so this represents a deliberately higher evidentiary bar than most prior AI dermatology meta-analyses.

Results: across the eleven studies, more than 2500 patients and 50 participating dermatologists were pooled. Dermatologists alone achieved pooled sensitivity around 79 percent and specificity around 75 percent. AI alone performed comparably — sensitivity around 81 percent, specificity around 76 percent — essentially statistically indistinguishable from dermatologist performance, though the confidence interval around the AI sensitivity estimate was noticeably wider, reflecting fewer AI studies and more heterogeneity. Only a single study reported on AI-assisted dermatologists, and there the numbers looked more favorable — sensitivity around 92 percent and specificity around 84 percent — but that's one study, so it's a signal, not a conclusion. Across the head-to-head comparisons, AI tended to show somewhat higher specificity with similar sensitivity compared to dermatologists.

The critical appraisal finding is really the headline here, and it should temper any enthusiasm: most included studies were rated high risk of bias, primarily because lesions were preselected as melanoma-suspicious before being fed to the AI or the dermatologist, and outcomes were often forced into binary malignant-versus-benign classification — neither of which reflects the messy, low-pretest-probability reality of a general dermatology or primary care clinic.

The authors' own conclusion is appropriately cautious: AI performs at a level comparable to dermatologists in these prospective but still enriched and imperfect study settings, and the one data point on AI-assisted dermatologists is intriguing for a decision-support role, but the evidence base remains small, heterogeneous, and at high risk of bias, and broader validation in genuinely unselected clinical populations is still needed.

Practical takeaway: not practice-changing. This is reassuring background evidence that AI classifiers are approaching dermatologist-level performance under prospective conditions, and it strengthens the case for AI as a decision-support adjunct rather than a replacement — but with only one study on human-plus-AI performance and pervasive selection bias in patient recruitment across the literature, there's nothing here that should change how you use or don't use AI tools in your own practice today.

Last, an original investigation: "Risk of Skin Cancer Among Patients With Chronic Lymphocytic Leukemia," a nationwide Danish registry study.

The clinical problem is familiar in concept — chronic lymphocytic leukemia causes both disease-related and treatment-related immunosuppression, and it's long been suspected to raise skin cancer risk — but the gap is that prior studies lacked properly matched control groups and, critically, nobody had quantified the absolute risk in a way useful for actual patient counseling and surveillance planning, nor followed the causal chain all the way through to skin cancer-specific metastasis and death.

Methodologically, this is about as rigorous as registry epidemiology gets. Leveraging Denmark's national registries linked by unique personal identifiers, they identified all CLL patients from 1990 to 2020, excluded anyone with prior skin cancer or other immunosuppression, and matched each CLL patient to five controls using exposure-density matching on birth year, sex, region, education, income, marital status, and comorbidity burden. The rationale for this design is essentially forced by the question: you cannot randomize people to develop leukemia, so a large matched cohort is the only ethical and feasible way to isolate the CLL-specific contribution to skin cancer risk from the general age-related and comorbidity-related background risk. They then used cause-specific Cox models treating death from other causes as a competing risk — which matters a great deal in an elderly, comorbid population like this, because ignoring competing mortality would otherwise distort the cancer risk estimates.

The results are compelling. Among just over 8,300 CLL patients matched to nearly 42,000 controls, median age about 71 in both groups, the ten-year absolute risk of any skin cancer was 13.5 percent in CLL patients versus 6.9 percent in controls — essentially double the risk, an absolute difference of about 6.6 percentage points, and this was highly statistically significant. Breaking it down by subtype, basal cell carcinoma was the most common contributor — about 8.6 percent versus 5.4 percent — and squamous cell carcinoma showed a similar magnitude of increase, roughly 4.7 percent versus 1.4 percent, more than a three-fold relative increase. Critically, and this is the number that should actually inform your counseling: skin cancer-specific metastasis was significantly higher in CLL patients, about 0.7 percent versus 0.1 percent, and skin cancer-specific death was also higher, about 0.3 percent versus 0.1 percent. Both differences were statistically significant, but the authors are careful to frame these absolute numbers as still low in isolation — contextualized against an all-cause mortality rate of 56 percent in CLL patients versus 39 percent in controls at ten years, reflecting just how much competing mortality risk dominates the overall picture in this population.

In discussion, the authors interpret this as clear confirmation that CLL carries a real, quantifiable, and clinically meaningful excess skin cancer risk — driven predominantly by keratinocyte carcinomas — that does translate into a small but statistically real increase in metastasis and disease-specific death, not just more biopsies and more diagnoses. Limitations worth flagging: Rai stage, the marker of CLL severity, was only available for about 45 percent of patients due to incomplete lab registry coverage, so they couldn't fully stratify risk by disease severity or clearly separate disease-related from treatment-related immunosuppression effects. It's also a Danish population, so predominantly Fitzpatrick I to II, and generalizability to more richly pigmented populations is genuinely unknown. And as with all registry studies, granular data on treatment regimens, cumulative sun exposure, and comorbidities beyond what's captured in the Charlson Index isn't available.

Practical takeaway — this one edges toward practice-changing, at least for surveillance policy. A properly matched, nationwide cohort showing a doubled ten-year skin cancer risk, with real if small increases in metastasis and disease-specific death, is solid justification for incorporating routine, probably annual, skin surveillance into standard CLL care pathways, and for having a lower threshold to biopsy in these patients given their squamous cell carcinoma risk in particular is more than tripled. It's also useful ammunition for pushing hematology-oncology colleagues toward earlier dermatology referral rather than waiting for a visibly concerning lesion.

That wraps our four articles for the May 2026 issue. To summarize the through-line: new population-specific virology data from Asia that complicates rather than confirms the Western HPV paradigm, encouraging but still early toxicity and efficacy signal for intralesional IL-2 in squamous cell carcinoma, a sobering reality check on AI melanoma diagnostics that keeps it firmly in the decision-support lane for now, and solid registry evidence that should tighten your surveillance threshold for CLL patients. Thanks for listening, and we'll see you next month.