Welcome back to the journal review. This is our walkthrough of JAMA Dermatology for October 2026, and we've got two original investigations on deck — one a feasibility study applying AI-assisted optical imaging to hunt for invisible basal cell carcinomas, the other a large prognostic modeling study building risk calculators for skin cancer after blood or marrow transplant. Let's get into it.

First up is a prospective cross-sectional feasibility study out of Erlangen, Germany, looking at AI-assisted line-field confocal optical coherence tomography — LC-OCT — as a tool to detect subclinical basal cell carcinoma, meaning BCC with no visible lesion at all. You know LC-OCT already as a noninvasive alternative to biopsy for a single suspicious spot, with reported sensitivity and specificity in the high nineties and the classic mille-feuille lobular pattern on imaging. What nobody had done is turn that single-lesion tool into a screening instrument that sweeps the entire face of a high-risk patient looking for BCC that isn't clinically apparent yet. The gap here is conceptual as much as technical: we know subclinical extension exists at the margins of a known tumor, but a BCC with zero visible correlate, found purely by imaging surveillance, hadn't really been described.

Methodologically, this is explicitly labeled a feasibility study, and that framing matters — the authors aren't trying to prove a screening program reduces morbidity or saves money, they're asking whether the basic pipeline even works. They enrolled 150 consecutive inpatients, mostly admitted for skin cancer surgery, who had at least two established BCC risk factors — age over 65, Fitzpatrick type 1 or 2, personal BCC history, chronic UV exposure, genetic syndromes, or prior ionizing radiation to the head and neck. After excluding any lesion that was macroscopically suspicious — which in practice mostly meant areas already covered by postoperative dressings — a single trained operator ran the LC-OCT probe systematically across the remaining, visually unremarkable facial skin. An integrated AI pattern-recognition layer gave a real-time color-coded probability score, essentially a heat map nudging the examiner to slow down wherever the yellow signal appeared. Critically, the actual diagnosis of BCC was still made by a physician applying the established LC-OCT morphologic criteria — lobules, clefting, bright rim — independent of the AI score, so the AI here functions as an attention-directing triage layer, not an autonomous diagnostician. Think of it as a faster way to find the needle, not a replacement for recognizing the needle once found.

On the statistics: because only lesions the AI-plus-physician combination called positive went on to biopsy, there's a built-in verification bias — nobody biopsied the areas that were called negative, so you can't calculate sensitivity or specificity, full stop. The only clean metric available is positive predictive value, the probability that a positive call actually reflects disease, and they correctly limited themselves to that. For confidence intervals around proportions with pretty small numerators, they used the Clopper-Pearson, or exact, method — this is the right choice here rather than a normal approximation because standard Wald intervals get unstable and can even produce nonsensical bounds when you're working with small samples or proportions near the edges, whereas Clopper-Pearson guarantees proper coverage no matter how sparse the data.

Results: among the 150 patients, subclinical BCC was histologically confirmed in 14 patients, just over 9%, with a couple having more than one lesion for 17 subclinical BCCs total. Of 18 lesions the AI-assisted workflow flagged as BCC, 15 were confirmed on biopsy, giving a PPV of about 83%, with a reasonably wide confidence interval given the small numbers. One flagged lesion turned out to be an actinic keratosis, and two patients simply declined biopsy; if you credit those two as confirmed based on expert image review, the PPV climbs to roughly 94%. Subtype-wise, the overwhelming majority — about three-quarters — were superficial BCCs, the subtype most amenable to nonsurgical treatment, with a handful of nodular cases and a single infiltrative one. The LC-OCT subtype call matched histology in 13 of 15 cases, so subtype accuracy was also quite good, right around 87%.

A few practical findings worth flagging for your own workflow intuition: over half the patients had some portion of the face excluded from screening because of wound dressings or facial hair, and a full, unobstructed facial screen took about 40 minutes — that's a real time cost, not a quick add-on to a clinic visit. The authors are appropriately honest about limitations: this is a single-center, single-operator study, there's no sensitivity or specificity data because of the verification bias baked into the design, there's no genetic-syndrome patients despite that being an eligibility criterion, and — most importantly — there's no outcome data showing that catching these subclinical lesions earlier actually changes morbidity, cosmesis, or survival compared with waiting for clinical apparency. Finding a 9% subclinical BCC rate in a high-risk population is striking, but "detectable earlier" and "clinically beneficial to detect earlier" are not the same claim, and the authors say so explicitly.

So where does this land practically for you. This is interesting, not yet practice-changing. It's a genuine proof-of-concept that systematic AI-guided LC-OCT screening can find real, previously invisible BCCs in a high-risk population, and the subtype discrimination performance is good enough to be clinically relevant if validated, since superficial disease triages so differently than nodular or infiltrative disease. But a 40-minute scan time, single-operator reproducibility, the absence of true sensitivity data, and zero evidence of downstream benefit mean this is not something to build into your surveillance protocol yet. Keep it on your radar as a technology to watch, particularly for your syndromic and transplant populations, but don't expect reimbursement or guideline incorporation anytime soon.

Our second article is a prognostic study, and a much bigger analytic lift: risk stratification models for subsequent cutaneous malignant neoplasms after blood or marrow transplant, using the multi-institutional BMT Survivor Study cohort. The clinical problem is one you've probably sensed anecdotally — BMT recipients, especially those with chronic graft-versus-host disease or prior total body irradiation, run a markedly elevated lifetime risk of BCC, squamous cell carcinoma, and melanoma, with prior work from this same group putting 30-year cumulative incidence above one in four. What's been missing is a way to take an individual patient in front of you and say with some precision where they fall on that risk spectrum, so that dermatologic surveillance intensity and risk-mitigation counseling can be targeted rather than one-size-fits-all.

The cohort was substantial: 3,448 BMT Survivor Study participants who underwent a single transplant, survived at least two years post-transplant free of skin cancer, and had complete covariate data — drawn from sites in California, Minnesota, and Alabama. The design is retrospective and longitudinal, which is really the only feasible approach for a question like this — you need decades of follow-up to capture cumulative incidence out to 15, 20, 25 years, and no one is going to run a randomized trial of transplant characteristics.

Now, the statistical engine here is worth slowing down on, because it's a nice teaching example of a competing-risks problem. The authors built separate models for BCC, SCC, and melanoma using Fine-Gray subdistribution hazard models, treating death from other causes as a competing risk. Why not just use an ordinary Cox model? Because in a transplant survivorship cohort, a meaningful fraction of patients will die of something else — relapse, infection, graft complications — before they ever get the chance to develop a skin cancer, and a standard Cox model, or a naive Kaplan-Meier curve, treats those deaths as if the patient were simply "censored," which implicitly assumes they were still just as likely to go on and develop skin cancer, which obviously isn't true. The Fine-Gray approach instead directly models the cumulative incidence of the specific event of interest while properly accounting for the fact that competing events remove people from the risk set permanently. In plain terms, it's answering "what's the actual probability this patient develops skin cancer by year fifteen, given that some patients will die of other things first," rather than the more abstract and less clinically intuitive hazard-ratio-if-nothing-else-happened that a standard Cox model gives you.

From those Fine-Gray models, they built a simple point-based risk score — essentially summing the hazard ratios for each patient's risk factors, a practical approach borrowed from other published survivorship risk tools — and then validated discrimination using time-dependent area under the receiver operating characteristic curve, repeated across 200 random splits into training and test sets. That repeated-partitioning step is their internal validation strategy: rather than reporting performance on the same data used to build the model, which always looks falsely good, they rebuilt the model many times on random 60% subsets and tested it on the held-out 40%, then averaged the results — a form of cross-validation that gives you an honest read on how the model would perform on patients it hasn't seen. They judged an AUC of 0.7 or better as good, and 0.8 or better as excellent.

The results were genuinely strong. Across the full cohort, 282 patients developed BCC, 183 SCC, and 73 melanoma. At 15 years post-transplant, the test-set AUC was 0.81 for BCC, 0.90 for SCC, and 0.83 for melanoma — that SCC number in particular is excellent discrimination by any standard, or really by any field's standard. And the clinical separation this produced was dramatic, not just statistically significant. For BCC, low-risk patients had a 15-year cumulative incidence of about 2%, versus 13% in the high-risk group — roughly a five-fold difference. For SCC, it was about 3% versus 14%. For melanoma, under 1% versus around 4%. These are not marginal differences you'd need a p-value to appreciate — they're the kind of spread that would plausibly change how often you bring a patient back to clinic.

The models going into each subtype differed sensibly based on prior biology: all three included age at BMT, race and ethnicity, pre-BMT cutaneous malignancy, and post-transplant immunosuppression, while the BCC model additionally incorporated male sex, pretransplant radiation and monoclonal antibody exposure, total body irradiation, donor source, and chronic graft-versus-host disease — including a specific interaction term capturing that younger age at total body irradiation exposure carries extra BCC risk. The SCC model added male sex, donor source, and chronic GVHD, but not the irradiation variables, consistent with the authors' prior finding that TBI and monoclonal antibody exposure track with BCC risk but not SCC or melanoma risk. They also built an online risk calculator that spits out five-year interval cumulative incidence estimates and surveillance recommendations by risk strata, which is a nice translational touch.

On limitations, the authors are transparent that outcomes were largely patient-reported, with pathology confirmation only when records were available — though they cite their own prior validation work showing reasonable concordance, around three-quarters confirmed, and no difference in risk factor associations between confirmed and unconfirmed cases, which is a reasonable justification for including all reported cases rather than discarding unconfirmed ones and shrinking an already outcome-limited cohort. The cohort also spans transplants performed from 1974 through 2014, so conditioning regimens, immunosuppressive strategies, and surveillance practices have shifted substantially over that window, which could limit direct applicability to a patient transplanted under contemporary protocols. And of course this is a risk-prediction model, not an interventional trial — it tells you who's at higher risk, not yet whether intensified surveillance or specific mitigation strategies in that high-risk group actually changes outcomes.

Practically, this one leans much closer to practice-relevant than article one, even if not immediately guideline-mandated. If you see post-BMT patients in your practice — and many Mohs surgeons and dermatologic oncologists do, often referred for exactly this reason — a validated, freely available online calculator stratifying patients into low versus high risk for BCC, SCC, and melanoma with this level of discrimination is a genuinely usable tool today for tailoring surveillance intervals and counseling intensity, particularly for identifying the subset of patients carrying chronic GVHD, prior TBI, or monoclonal antibody exposure who clearly warrant closer follow-up. What's not yet established is whether acting on that risk stratum — seeing high-risk patients more often, starting field therapy earlier, escalating sun protection counseling — actually moves the needle on morbidity, so treat the score as a triage aid for your surveillance intensity rather than a license to change management of any individual lesion.

That's our two for this episode — a promising but early-stage imaging technology looking for invisible BCC, and a well-validated, genuinely usable risk calculator for your transplant survivors. Thanks for listening, and we'll catch you in the next issue.