Welcome back to the Journal Review, your monthly deep dive into the Journal of the American Academy of Dermatology. This is the January 2026 issue, and we've got four pieces on deck — a quick surgical pearl on sharps safety, part two of a CME series on artificial intelligence in dermatology, a visually driven images article covering the clinical spectrum of cutaneous melanoma, and an ethics column tackling AI scribes. Let's get into it.

First up is a surgical pearl — a short, practical technique piece out of the University of Oklahoma, addressing something every one of us handles dozens of times a day without thinking much about it: how you take the blade off the handle at the end of a case. The authors frame the problem with a stat worth knowing — in the sharps injury literature, scalpel blades are the second most common cause of surgical sharps injuries, behind only suture needles, accounting for something like one in six surgical sharps injuries, with suture needles responsible for closer to four in ten. Safety scalpels have been floated as a fix, but the evidence for mandating them is thin, and the strategies with the best track record right now remain double-gloving and blunt suture needles. Blade removal itself, they argue, just hasn't gotten the ergonomic attention it deserves.

Their fix is refreshingly simple and needs no results section or data to justify it — it's a technique observation, not a trial. Instead of grasping the blade at a single point at its base with forceps or a needle driver, the traditional method, they advocate a two-point grasp: cover the belly of the blade with the instrument, contacting it at two points rather than one, lift slightly, and slide it off at a downward angle toward the tray. The stated rationale is biomechanical — distributing pressure across two contact points instead of one reduces blade flexion, slippage, and the chance of sudden uncontrolled release, and it does this while actually being a smoother, less strenuous hand motion. There's no cohort here, no injury-rate comparison, just a proposed ergonomic modification with a supplemental video demonstrating it. The takeaway is low-cost and immediately actionable — this is the kind of thing you could adopt in your very next case with zero downside, worth teaching to residents and staff during blade handling instruction, even if it hasn't been formally validated against injury outcomes.

Next, part two of a two-part CME review on augmented intelligence in dermatology, picking up where part one left off on core AI concepts and now tackling the messier back half — bias, benchmarking, ethics, regulation, and where the field is headed. This is a review and educational piece, so there's no methods or results to walk through — it's a synthesis, and I'll hit the throughlines that matter for practice.

The first big theme is data quality as the rate-limiting step for everything downstream. The authors walk through how dermatology datasets are inconsistent in format, curation, and consent standards across sites, and how that variability — different imaging setups, different lighting, different labeling conventions — can bake in spurious correlations. Their example, which is worth remembering when you're evaluating a vendor's tool, is that models have been shown to key in on surgical marking ink or rulers as a proxy for malignancy, rather than the actual lesion morphology, simply because those markers correlated with malignant cases in the training data. They also flag the labor-intensive nature of expert annotation as a bottleneck, discuss crowdsourcing and federated learning as partial workarounds, each with its own tradeoffs, and note that synthetic image generation is being explored but doesn't perfectly capture real-world nuance.

The second theme is algorithm performance and the sensitivity-specificity balancing act, with a concrete callout that's directly relevant to your patients — smartphone melanoma detection apps have been shown to have highly variable and generally low sensitivity, and some suffer from low specificity too, which translates clinically into both missed melanomas and a lot of unnecessary anxiety and downstream visits. They also discuss overfitting, the risk that a model dazzles on its training set but fails on external data, and they're candid that datasets skew toward lighter skin types, which degrades accuracy and risks misdiagnosis in patients with darker skin — a point with obvious equity implications for anyone doing total body screening or triage in a diverse population.

Third, they move through transparency, trustworthiness, and the practical mechanics of clinical integration — electronic health record compatibility, workflow alignment, and the need for dermatologist buy-in and human-centered design so tools actually solve problems we care about rather than problems that are easy to build a model for. They close with a call for prospective, real-world validation and ongoing monitoring for model drift, meaning performance can degrade over time as the clinical landscape shifts even without the underlying model changing.

There's no practice-changing action item here in the sense of a technique to adopt tomorrow — this is foundational literacy. But the practical takeaway is real: before you or your practice adopts any AI-based diagnostic aid, you now have a framework for the right questions to ask a vendor — what population was the training data drawn from, was performance validated prospectively across skin types and external sites, and what's the plan for monitoring drift after deployment.

Third is a dermatology images article — essentially an atlas piece, not a study, so there's no methods or results scaffolding to walk through, just a curated visual tour of cutaneous melanoma's clinical diversity, contributed by authors in Chongqing, China. It's organized around the traditional four-category framework — superficial spreading, lentigo maligna, nodular, and acral melanoma, in descending order of prevalence — while also referencing the WHO's more granular nine-subtype schema from the 2018 classification, which splits things out by pathogenesis and sun exposure relationship: superficial spreading, lentigo maligna, desmoplastic, spitzoid, acral, mucosal, uveal, melanoma arising in congenital nevus, and melanoma arising in blue nevus, with nodular growth pattern able to occur across any of these.

The images themselves are the real content — two dozen cases spanning an impressive breadth of presentations you don't all get to see routinely in one sitting: facial lentigo maligna melanoma as flat, variably pigmented macules; a spitzoid melanoma presenting as an innocuous-looking red dome-shaped papule on an elderly patient's face, a reminder that spitzoid lesions can be deceptively bland and amelanotic regardless of patient age; a desmoplastic melanoma that was initially misdiagnosed as a cyst and excised without biopsy, only to recur as a large plaque at the surgical site — a cautionary tale about pathological confirmation before assuming a benign scar-like lesion is what it looks like; melanoma arising within a congenital nevus in a teenager, evolving over sixteen years before a rapid enlargement following cryotherapy; and a full run through acral and subungual disease — palm, sole, and multiple digit cases showing nail destruction and ulceration. They round out with mucosal and metastatic presentations — oral, glans, vulvar, and both nodal and in-transit metastatic disease from acral primaries. There's no discussion or limitations section to speak of; the value here is purely pattern recognition. For those of us doing Mohs and managing melanoma across diverse anatomic sites, it's a useful refresher gallery, particularly for the less common subtypes — desmoplastic, spitzoid, and congenital nevus-associated melanoma — where the index of suspicion has to stay high despite an often unimpressive clinical appearance.

Last is a short ethics letter, styled as an advice-column exchange with "Dr Dermatoethicist," addressing a question that's landing in a lot of our own practices right now — the ethics of using AI scribes in place of human scribes. It's not a study, so think of this as a structured ethical analysis rather than data.

The letter opens by acknowledging why AI scribes are attractive: traditional human scribes are expensive, need training, and have high turnover, and their physical presence in the room can inhibit patients from disclosing sensitive information. Ambient and voice-to-text AI scribes promise to cut charting time, let providers stay focused on the patient, reduce overhead, and sidestep the recruitment and retention headaches of human staff. But the piece is careful to lay out the ethical tension using a classic four-principle framework — autonomy, beneficence, non-maleficence, and justice.

On non-maleficence, the concerning data point is that AI transcription has been documented to produce on the order of twenty-some errors per case, meaning physicians still need to spend time correcting output, and subtle nuances that a human scribe would catch may get missed, with real potential for patient harm. On the beneficence side, the counterpoint is that AI systems can self-improve with use, potentially closing that gap over time and reducing the variability you get with human scribes of differing experience levels. The autonomy discussion is the most actionable piece for your own practice: the authors argue patients need explicit disclosure that an AI scribe is being used, in writing, with capabilities and data storage explained, formal consent documentation, and a genuine opt-out option, mirroring the right patients already have to decline a human scribe. Justice and beneficence get invoked again around standardization — the argument that AI scribes could make documentation quality more consistent across encounters and free up clinician time for direct patient care. And finally, on privacy, the letter flags that third-party AI platforms are collecting encounter data to improve their own algorithms, which raises real questions about whether protected health information is adequately safeguarded and whether it could be repurposed beyond the original clinical encounter.

There's no formal recommendation or vote here — it's a balanced ethical airing, closing simply with the principle that accuracy, informed consent, and data security all have to be assured before you deploy one of these tools. Practically, if you're considering an AI scribe in your own practice, the actionable piece is the informed consent infrastructure — written disclosure, signage, and a real opt-out — and vetting exactly what your vendor does with the recorded data on the back end.

That wraps our four articles for January. A blade-handling tweak worth adopting immediately, a foundational AI literacy piece worth sitting with even without an action item, a melanoma image gallery worth bookmarking for teaching, and an ethics framework worth applying the next time a scribe vendor calls your office. Thanks for listening, and we'll see you next month.