Welcome to this journal review podcast, covering the Journal of the American Academy of Dermatology, July 2026 issue. We've got four articles to get through today, spanning workflow technology, artificial intelligence bias, imaging in high-risk squamous cell carcinoma, and a large population-based survival analysis of cutaneous sarcomas. Let's get into it.

Our first article is titled "Enhancing Dermatology Workflow through a Replicable Epic-Based Module: Impact on Documentation and After-Hours Electronic Medical Record Use," by Alomary, Baker, Taylor, and Mollanazar, from the University of Pennsylvania Perelman School of Medicine.

The clinical question here is whether a customized, specialty-specific documentation module built within the Epic electronic medical record system can improve documentation efficiency and reduce after-hours charting burden for dermatologists, compared with the default, generic dermatology application.

This was a pre-post observational study. The authors developed what they call DermModule, an enhancement to Epic's existing dermatology application that adds disease-specific treatment plans, embedded patient education materials, and structured procedure note templates with dropdown menus. They extracted Epic signal data for 74 active dermatology providers, comparing a seven-month preimplementation period, March through September 2024, against the same seven-month window one year later, March through September 2025, after DermModule rollout. Metrics were compared using paired t-tests.

The key findings: time in notes per day decreased from 46.64 minutes to 39.32 minutes, a highly significant difference with a P value of point zero zero zero zero two six. Time in notes per appointment dropped from 8.91 minutes to 7.65 minutes, P equals point zero zero two nine. Pajama time — after-hours electronic medical record use between five-thirty p.m. and seven a.m. and on weekends — fell from 34.50 minutes to 28.29 minutes per day, P equals point zero three six zero. Appointments per day trended upward from 11.74 to 12.30, which did not reach significance at P equals point zero five one, but importantly showed no evidence that efficiency gains came at the cost of productivity. Cumulatively, this corresponds to roughly 2,880 minutes of documentation time saved per clinician over the analysis period.

The evidence-based takeaway for surgical practice: specialty-specific customization of the electronic medical record beyond the vendor's default dermatology build can yield measurable reductions in both daytime documentation burden and after-hours charting, without sacrificing visit volume — a proof of concept that's particularly relevant for procedurally heavy practices like Mohs surgery, where structured procedure note templates could meaningfully cut down operative note time.

Limitations are important here. This is a pre-post design without a concurrent control group, so temporal confounding is a real concern — concurrent Epic platform updates, staffing changes, or shifts in documentation policy over that one-year interval could all contribute to the observed effect. There's also no adjustment for changes in patient case mix or note complexity between the two periods.

Moving to our second article, "Framing Bias in a Large Language Model: Prompt Framing Influences ChatGPT's Accuracy in Melanoma Classification, A Diagnostic Accuracy Study," by Traini, Palmisano, Di Stefani, and Peris, from Fondazione Policlinico Universitario Agostino Gemelli in Rome.

The clinical question: does the narrative framing of a prompt — independent of the underlying image — systematically bias ChatGPT's diagnostic output when classifying pigmented lesions on dermoscopy?

Study design: a diagnostic accuracy study using ChatGPT-5, the August 2025 release from OpenAI, tested as a black box with no fine-tuning. The dataset comprised 100 dermoscopic images from 100 patients at Policlinico Gemelli, all Fitzpatrick skin types one through three, with histopathology-confirmed outcomes — 11 melanomas and 89 dysplastic nevi. Each image was presented six times under six different prompt conditions: a neutral baseline asking for a score from 1, nevus, to 5, melanoma; and five framed variants — a concerned patient wanting a same-day appointment, a patient minimizing concern based on prior similar lesions, an irrelevant social detail about a tattoo near the lesion, an anxious spouse convinced it's melanoma, and a strongly reassuring framing suggesting it's nothing.

Key findings: relative to the neutral baseline, mean lesion scores rose significantly under the concerned-patient framing, an increase of 0.29, and under the minimizing-concern framing, an increase of 0.26, both P less than point zero five. Scores fell under the strong-reassurance framing, a decrease of 0.21. Diagnostic discrimination varied substantially by framing: area under the ROC curve ranged from 0.56 under strong reassurance, with a 95% confidence interval of 0.407 to 0.711, up to 0.75 under the irrelevant tattoo detail framing, with a 95% confidence interval of 0.593 to 0.875. For reference, the neutral baseline prompt yielded an area under the ROC curve of 0.722. Sensitivity ranged from 0.818 at baseline up to 0.909 under most framed conditions, while specificity was highest at baseline, 0.528, and fell as low as 0.36 under the anxious-spouse and strong-reassurance framings. Positive predictive values were low throughout, ranging from 0.149 to 0.176, reflecting the low melanoma prevalence in this sample. Outputs across prompts were only moderately correlated, with Spearman coefficients ranging from 0.54 to 0.79, indicating that while relative lesion ranking was partly preserved, absolute scores shifted enough to flip classification at plausible clinical thresholds.

The takeaway for practice: if large language model tools are used adjunctively for lesion triage or patient-facing decision support, output is not a stable, image-only function — it is highly sensitive to incidental narrative context embedded in the prompt, mirroring the framing biases well documented in human clinician decision-making. The authors argue that standardized, context-minimal prompting and explicit threshold calibration are prerequisites before such tools could be considered for safe clinical deployment, and that published evaluations of medical large language models should report full prompt wording to allow reproducibility.

Limitations: this is a modest sample of 100 images with only 11 melanomas, meaning confidence intervals around area under the ROC curve are wide and subgroup precision is limited. The dataset was restricted to Fitzpatrick skin types one through three, so generalizability to darker skin types is unknown. And this evaluated a single, black-box, closed-weight model at one point in time, so findings may not generalize to other large language models or future ChatGPT versions.

Our third article is a letter to the editor titled "Radiologic Imaging in High-Risk Cutaneous Squamous Cell Carcinoma: Clarifying the Role Beyond Selection Bias," by Xie, Wu, and Xu, from the Affiliated Hospital of Qingdao University.

This is a critical commentary responding to a prior retrospective cohort study by Wei and colleagues, which had reported that radiologic imaging revealed unexpected findings in nearly half of high-risk cutaneous squamous cell carcinoma cases and altered management in 47% of cases.

The clinical question raised by this letter is whether the original study's implied endorsement of routine imaging in high-risk cutaneous squamous cell carcinoma is actually supported by its data, or whether the findings are more parsimoniously explained by selection bias.

The authors raise three specific methodological critiques. First, they note a strong association in the original cohort between imaging use and more aggressive tumor features — Brigham and Women's Hospital T2b or T3 staging, perineural invasion, and poor differentiation — arguing this pattern suggests imaging was clinically driven by preexisting suspicion of aggressive disease, rather than imaging itself uncovering unsuspected pathology. In other words, the higher rate of nodal or distant metastases in the imaged group likely reflects which patients were selected for imaging, not an independent diagnostic contribution of imaging.

Second, they point out that the original study did not perform adjusted prognostic analyses — no multivariable regression or propensity-score adjustment for confounders such as tumor size, depth, or anatomic location. They highlight that the original study's own five-year disease-related outcome comparison was nonsignificant, 36% in the imaged group versus 25% in the non-imaged group, with a P value of point zero six, and overall survival was comparable between groups. Without adjustment for confounding, the letter argues it's not possible to conclude that imaging independently improves outcomes as opposed to simply identifying patients who already had more severe disease.

Third, the letter critiques the original study's recommendation to image "the local site and nodal basins" as lacking modality-specific guidance. Computed tomography and PET-CT were the dominant modalities used, while ultrasonography — increasingly validated for nodal assessment — was rarely used. The letter acknowledges this may partly reflect the study period, 2010 to 2012, which preceded broader adoption of ultrasound for staging and surveillance in cutaneous squamous cell carcinoma.

The evidence-based takeaway for surgical and Mohs practice: this letter is a call for caution against over-interpreting retrospective, clinically-driven imaging data as evidence supporting routine imaging in all high-risk cutaneous squamous cell carcinoma. Imaging decisions should continue to be individualized based on clinical risk factors rather than applied reflexively across the board, pending prospective studies with standardized imaging protocols and confounder-adjusted outcome analyses.

Now to our fourth and final article, "Relative Survival Analysis of Dermatofibrosarcoma Protuberans, Kaposi Sarcoma, and Pleomorphic Sarcoma across Intersectional Demographics: A Surveillance, Epidemiology, and End Results Study," by Hu, Zhu, Lambert, Block, Rabinowitz, Verma, Jónasson, Sigurgeirsson, Meehan, Ungar, Lewin, Gulati, and Adalsteinsson, from the Icahn School of Medicine at Mount Sinai and the University of Iceland.

The clinical question: how do five-year relative survival rates for three rare cutaneous sarcomas — dermatofibrosarcoma protuberans, Kaposi sarcoma, and pleomorphic sarcoma, the latter encompassing undifferentiated pleomorphic sarcoma, pleomorphic dermal sarcoma, atypical fibroxanthoma, and malignant fibrous histiocytoma — vary across intersecting demographic categories of age, race and ethnicity, and sex?

Study design: a retrospective population-based analysis using the Surveillance, Epidemiology, and End Results database, identifying 7,145 patients with dermatofibrosarcoma protuberans, 8,602 with Kaposi sarcoma, and 5,700 with pleomorphic sarcoma, diagnosed between 2000 and 2021. Age-standardized five-year relative survival rates and 95% log-log confidence intervals were calculated across demographic intersections.

Key findings, starting with overall rates: five-year relative survival differed significantly from the expected value of 100% for pleomorphic sarcoma, at 71.3%, with a 95% confidence interval of 69.6% to 72.9%, and for Kaposi sarcoma, at 76.6%, confidence interval 75.1% to 78.0%. Dermatofibrosarcoma protuberans did not differ significantly from expected survival overall.

Within dermatofibrosarcoma protuberans, the youngest and oldest age groups were most affected, though all age groups maintained absolute relative survival rates above 97%. Men showed a significant reduction from expected survival, 98.2%, confidence interval 95.2% to 99.3%, while women did not. The most pronounced reduction was in men aged 75 and older, at 88.4%, with a wide confidence interval of 58.9% to 97.2%. Patients of color had statistically significant reductions from expected survival, though absolute rates remained above 97%. Asian men had the lowest relative survival rate in this cancer, at 95.9%, confidence interval 87.9% to 98.7%.

For Kaposi sarcoma, the lowest relative survival by age was seen in patients 15 to 44 years old, at 65.5%, confidence interval 64.0% to 67.0%, while patients 75 and older fared best, at 92.5%, confidence interval 84.8% to 96.4%. Women had worse relative survival overall than men, 67.5% versus 76.9%, with the disparity most pronounced in younger patients — 47.2% in women aged 15 to 44 versus 66.1% in men of the same age range — a gap that narrowed substantially by age 75 and older, 91.0% versus 93.3%. By race, Native American and Asian patients had the lowest relative survival, at 63.5% and 68.7% respectively, while white and Hispanic patients fared best, at 78.3% and 78.6%. Notably, white women and men had similar survival, 77.6% versus 78.3%, but Asian, Black, and Hispanic women had relative survival rates 20.2%, 11.0%, and 7.0 percentage points lower, respectively, than their male counterparts.

For pleomorphic sarcoma, overall survival was similar between men and women, 70.6% versus 71.3%, with no clear trend by age alone. However, when age and sex were combined, women's relative survival declined progressively with age, from 81.7%, confidence interval 75.8% to 86.3%, in those 15 to 44 years old, down to 65.4%, confidence interval 59.4% to 70.7%, in those 75 and older — meaning younger women fared better than younger men, but older women fared worse than older men. By race and sex combined, Black men had the lowest relative survival rate of any subgroup identified in the study, 48.3%, confidence interval 39.8% to 56.3%.

The evidence-based takeaway: these data identify substantial, previously underrecognized survival disparities across intersecting demographic strata for these rare cutaneous sarcomas — particularly for older men and Asian men with dermatofibrosarcoma protuberans, young women and several racial minority-sex combinations with Kaposi sarcoma, and Black men and older women with pleomorphic sarcoma. For dermatologic oncology practice, this reinforces the need for heightened clinical vigilance and consideration of closer surveillance in these identified higher-risk demographic subgroups, while acknowledging the descriptive nature of the findings.

Limitations, as the authors note: subtype-level data within Kaposi sarcoma and pleomorphic sarcoma were inadequate for granular subgroup analysis, despite likely prognostic heterogeneity within these categories. Relative survival rates in the Kaposi sarcoma cohort cannot disentangle Kaposi-attributable deaths from other HIV-related mortality, since the Surveillance, Epidemiology, and End Results database does not allow this distinction. Some race-sex subgroups could not be age-standardized due to insufficient sample size. And critically, these subgroup patterns are descriptive and observational — the authors explicitly state that further research is needed to determine whether these disparities reflect diagnostic bias, unequal access to care, or true biological differences.

That concludes our four articles for this July 2026 episode. In summary: a workflow intervention study suggesting that specialty-tailored electronic medical record customization can meaningfully cut documentation time and after-hours charting; a cautionary diagnostic accuracy study showing that ChatGPT's melanoma risk scoring is meaningfully destabilized by incidental prompt framing; a critical letter urging caution against over-generalizing retrospective imaging data into routine imaging recommendations for high-risk cutaneous squamous cell carcinoma; and a large population-based analysis revealing significant intersectional demographic disparities in relative survival across three rare cutaneous sarcomas. Thanks for listening, and we'll see you next month.