Welcome to this November 2026 run-through of the Journal of the American Academy of Dermatology. We've got four pieces to get through today — a consensus guideline on PTEN hamartoma tumor syndrome, a research letter out of Mass General Brigham looking at keratinocyte carcinoma risk in lupus, a JAAD International highlights column on artificial intelligence and Mohs frozen sections, and an ethics column on disclosing AI use to patients. Let's get into it. First up is a brief report, really a consensus statement, on dermatologic manifestations of PTEN hamartoma tumor syndrome, or PHTS — what many of you still think of by its older name, Cowden syndrome. This is the first formal set of dermatology-specific management guidelines for this condition, produced by a dedicated PHTS Dermatology Working Group. As background, PHTS is an autosomal dominant condition driven by germline PTEN mutations, and it's loaded with cutaneous findings — trichilemmomas, acral keratoses, oral papillomas — that show up in nearly all affected adults by the time they're diagnosed, with the literature cited here putting the prevalence of cutaneous findings as high as 98%. Despite how common these findings are, there's never been a standardized approach to managing them, and that's the gap this group set out to fill. Methodologically, this isn't a clinical trial — it's a modified Delphi consensus process, which is really the only appropriate tool when you're dealing with a rare disease where randomized data will never exist. A core panel of four experts plus a patient representative did a systematic literature review, drafted recommendation statements, and then ran those statements past an independent external panel of four additional reviewers in two rounds, with consensus defined as at least eighty percent agreement. They graded the evidence behind each statement as strong, moderate, or weak, borrowing a grading framework from a prior PTEN cancer surveillance guideline, and when evidence was essentially absent, statements were simply labeled expert opinion. This is worth pausing on as a methods lesson: Delphi consensus is a formal way of aggregating expert judgment when you can't generate empirical data, and the iterative, anonymized voting process is specifically designed to avoid the trap of one dominant voice in the room steering the outcome — it's a structured alternative to a simple committee opinion. The output was thirteen recommendation statements, and every single one reached consensus, though the evidence grading behind them varied enormously — much of it landed in the expert opinion or weak evidence category, which tells you something about how thin the underlying literature is. The headline clinical points: full skin exams by a dermatologist every one to two years starting at diagnosis, regardless of age, because melanoma risk in PHTS is meaningfully elevated — the panel cites a lifetime risk by age eighty in the range of five to six percent, which is well above general population risk. Importantly, the melanoma risk itself is evidence-based, but the one-to-two-year screening interval is not — that's expert opinion, not validated screening data. For benign lesions like trichilemmomas and lipomas, the recommendation is conservative management — reserve surgical intervention for lesions that are symptomatic, functionally impairing, or suspicious for malignancy, because excision in this population carries high rates of keloid formation and local recurrence. Systemic mTOR inhibitors like sirolimus are floated as an option for severe, symptomatic cutaneous disease, but only under specialist guidance, and that recommendation carries weak evidence. On the oral side, annual dental evaluation is recommended, along with counseling patients about medications known to cause gingival hyperplasia, like phenytoin and nifedipine, and fractional carbon dioxide laser gets a mention as having anecdotal efficacy for symptomatic oral papillomas. The honest limitation here, which the authors state plainly, is that this entire document rests on a rare disease evidence base — small cohorts and expert opinion, not randomized trials — and they explicitly call for future research on mTOR inhibitors and laser therapy combinations. For your practice, this is genuinely useful as a reference framework rather than a practice-changer in the sense of altering what most of you are already doing — if you're already managing these patients conservatively and screening regularly, this validates your approach and gives you something citable. The practical new information worth flagging to patients and referring physicians is the specific melanoma risk quantification and the reminder that pediatric melanoma in this population tends to present atypically — amelanotic, nodular, spitzoid — which is a teaching point worth keeping in your back pocket for genetics-driven surveillance conversations. Next, a highlights column from JAAD International summarizing a study on artificial intelligence and Mohs frozen section interpretation — this is a commentary, not an original study, so I'll walk it through as the author intended, with an eye toward what it tells us about where AI-assisted frozen section review actually stands. The underlying study, from a Swedish group, developed a weakly supervised deep learning model to detect basal cell carcinoma directly on Mohs frozen section whole-slide images. Weakly supervised here means the model was trained without pixel-level tumor annotation — in other words, nobody had to painstakingly outline every tumor focus on every training slide, which is a huge practical advantage given how labor-intensive full annotation is for frozen sections full of artifact and inflammation. The model trained on nearly a thousand whole-slide images from one center and was then externally validated on almost four hundred slides from two separate centers using different scanners — that external, multi-site validation is the methodologically important part, since internal validation numbers are almost always optimistic and don't tell you how a model behaves on unfamiliar slide preparation and staining. And that's exactly what happened here: internal performance was excellent, with accuracy and area under the ROC curve both above ninety percent, but sensitivity dropped from ninety-nine percent internally down to eighty-eight and then eighty-one percent at the two external sites. That's a meaningful, clinically relevant drop — missing roughly one in five tumors at the weaker external site is not something you could deploy unsupervised. Interestingly, when the authors reviewed five cases flagged as model errors, some of those turned out to be mistakes in the original pathology ground truth rather than true model failures, which is a nice reminder that reference standards in these validation studies aren't infallible either. The author frames this fairly as an impressive proof of concept rather than a deployment-ready tool, and highlights slide-quality heterogeneity across institutions as the central challenge for any AI system built on rapidly prepared frozen sections, which notoriously vary more in quality than permanent sections. He also situates this within a broader wave of AI-in-dermatology work — patient preference studies showing people actually prefer AI-generated explanations in some contexts, simulations of how AI tumor-mapping tools might affect Mohs lab staffing, and open ethical questions about who bears responsibility when an AI system gets it wrong. For your practice, the takeaway is interesting but clearly not yet actionable — this is not a tool to bring into your lab tomorrow, but it's a signal that frozen-section AI assistance is maturing quickly, and the open questions about multicenter training and how such tools might change operative decision-making under uncertainty are worth watching over the next couple of years. Now to the original research piece I think deserves the most attention today: a retrospective cohort study out of the Mass General Brigham Research Patient Data Registry looking at keratinocyte carcinoma and precursor lesion risk in systemic lupus erythematosus. The clinical question is one you've probably wondered about yourself — lupus patients carry classic keratinocyte carcinoma risk factors, chronic photosensitivity and long-term immunosuppression, but prior studies on whether SLE itself actually elevates basal cell and squamous cell carcinoma risk have been inconsistent, and almost nobody had looked carefully at the precursor lesions — actinic keratosis and squamous cell carcinoma in situ — or at how quickly those precursors progress to invasive disease. Methodologically, this is a large electronic health record-based retrospective cohort, pulling just over seventeen thousand adults with incident SLE from nearly three decades of records and matching them to controls on age, sex, race, and estimated healthcare utilization — that utilization matching is a smart design choice because it helps control for surveillance bias, meaning you want to make sure any difference in skin cancer detection isn't just because lupus patients see dermatologists more often and therefore get diagnosed more. They excluded anyone with prior actinic keratosis, in situ, or invasive keratinocyte carcinoma at baseline, so this is a true incident-disease design. The core analytic tool here is the Cox proportional hazards model, and it's worth spending a moment on why that's the right choice for this question. You've got patients followed for wildly different lengths of time, some for a couple of years, some for two or three decades, and you've got the usual right-censoring problem — most patients never develop the outcome during observation, but you don't want to just throw out that follow-up time. A Cox model handles this by comparing, at every moment someone is still at risk and event-free, the instantaneous rate of the event between the exposed and unexposed groups, without requiring you to assume what the underlying baseline risk curve actually looks like over time — that's the "proportional hazards" part, it only assumes the relative risk between groups stays constant, not that the absolute risk itself follows any particular shape. That makes it far better suited here than something like a simple logistic regression comparing fixed-interval outcome rates, which would throw away all that timing information and couldn't handle variable follow-up gracefully. Onto results. Overall, SLE was associated with a statistically significant and clinically meaningful increase in actinic keratosis risk, about a fifty percent increase, and an even larger increase in squamous cell carcinoma in situ risk, roughly double. But — and this is the key finding — there was no significant increase in invasive squamous cell carcinoma or basal cell carcinoma risk for the SLE cohort as a whole. Even more counterintuitively, among patients who did develop an actinic keratosis, having SLE was associated with a longer interval before that lesion progressed to invasive squamous cell carcinoma, not a shorter one — roughly a one-third reduction in the hazard of progression compared to controls. The story changes substantially once you stratify by lupus subtype. Patients with SLE alone showed that same pattern — elevated precursor risk, unchanged invasive carcinoma risk, slower progression. But the smaller subgroup with comorbid discoid or subacute cutaneous lupus, roughly ten percent of the cohort, showed elevated risk across the board — precursors and invasive basal and squamous cell carcinoma alike — and critically, without that protective slower-progression effect. Medication stratification added another layer: hydroxychloroquine and methotrexate exposure were not associated with increased invasive carcinoma risk, systemic steroids were linked to increased squamous cell carcinoma risk specifically, mycophenolate to increased basal cell carcinoma risk, and azathioprine along with calcineurin inhibitors to increased risk of both. Hydroxychloroquine was also the only exposure associated with that favorable, prolonged actinic keratosis-to-squamous-cell-carcinoma interval. The authors' interpretation is that precursor dysplasia may be a fairly intrinsic consequence of SLE itself, driven by photosensitivity and immune dysregulation, while progression to actual invasive carcinoma depends more on cutaneous phenotype and specific immunosuppressive exposure. The limitations are the ones you'd expect from an EHR-based registry study — no data on lupus disease severity, no medication dosing information, no quantified ultraviolet exposure, and actinic keratosis treatment intensity couldn't be fully captured, though they note field therapy use looked similar between groups, which somewhat reassures against that particular confounder. From a practical standpoint, I'd call this genuinely practice-informing, if not fully practice-changing. It doesn't mean you suddenly treat every SLE patient as high invasive-carcinoma risk — the data argue against that blanket assumption. But it does sharpen your risk stratification: patients with discoid or subacute cutaneous lupus, or those on azathioprine, calcineurin inhibitors, mycophenolate, or chronic steroids, deserve closer surveillance and a lower threshold for treating actinic keratoses aggressively, while patients on hydroxychloroquine alone appear to carry a more favorable risk profile. It's also a nice reminder that lupus patients accumulating actinic keratoses aren't being over-surveilled for nothing — the precursor burden is real — but that precursor burden doesn't automatically translate to invasive risk for the whole population, only for specific subgroups. Last on the list is an ethics column — a Dear Dermatoethicist letter, not a study, so no methods or results here, just a case-based ethical discussion worth knowing about since you're all going to run into this personally. The scenario is a dermatologist who consults an AI language model before a visit with a patient with refractory hidradenitis suppurativa, gets a biologic suggestion they hadn't considered, incorporates it into the plan after reviewing the evidence themselves, and then doesn't mention the AI consultation when the patient later asks how the decision was made. The ethicist's reasoning centers on patient autonomy — disclosure is warranted when the information would be meaningful to a reasonable patient, and here the AI suggestion directly changed the treatment offered, which clears that bar. The column draws a distinction that's worth internalizing: AI tools are different in kind from reference sources like UpToDate or a textbook, because they generate novel content rather than retrieve vetted, peer-reviewed information, they can fabricate details, and they may carry data-sharing concerns if you're not using a HIPAA-compliant platform with deidentified information. At the same time, the piece argues against reflexive disclosure of every tool used in every encounter, since over-explaining can erode trust or distract from the visit when the AI's role was minimal or purely confirmatory. The practical framework offered is to scale disclosure to influence — minimal, confirmatory AI use probably doesn't need mention, but when a tool meaningfully expands your differential or surfaces a treatment you hadn't considered, as happened here, a brief acknowledgment is appropriate, something as simple as noting you consulted an AI tool to make sure no options were being missed, while being clear it supports rather than replaces your clinical judgment. There's no data or methodology to critique here — it's a reasoned ethical position — but it's a genuinely practical framework given how quickly these tools are showing up in real clinical decision-making, including in dermatologic surgery contexts where off-label and complex treatment decisions are common. That wraps our four articles for November. To summarize the signal versus noise: the PHTS guideline is a solid new reference document rather than a change to what most of you are already doing; the Mohs AI frozen-section work is a promising but clearly pre-deployment proof of concept; the lupus keratinocyte carcinoma cohort study offers real, actionable risk stratification based on cutaneous phenotype and specific immunosuppressants; and the AI disclosure ethics piece gives you a practical framework worth keeping in mind as these tools inevitably creep further into your own decision-making. Thanks for listening, and I'll see you next month.