Background: Artificial intelligence (AI) has shown significant promise in augmenting diagnostic capabilities across medical specialties. Recent advancements in generative AI allow for synthesis and interpretation of complex clinical data including imaging and patient history to assess disease risk. Objective: To evaluate the diagnostic performance of a dermatology-trained multimodal large language model (DermFlow, Delaware, USA) in assessing malignancy risk of pigmented skin lesions. Methods: This retrospective study utilized data from 59 patients with 68 biopsy-proven pigmented skin lesions seen at Indiana University clinics from February 2023 to May 2025. De-identified patient histories and clinical images were input into DermFlow, and clinical images only were input into Claude Sonnet 4 (Claude) to generate differential diagnoses. Clinician pre-operative diagnoses were extracted from the clinical note. Assessments were compared to histopathologic diagnoses (gold standard). Results: Among 68 clinically concerning pigmented lesions, DermFlow achieved 47.1% top diagnosis accuracy and 92.6% any-diagnosis accuracy, with F1 = 0.948, sensitivity 93.9%, and specificity 89.5% (balanced accuracy 91.7%). Claude had 8.8% top diagnosis and 73.5% any-diagnosis accuracy, F1 = 0.816, sensitivity 81.6%, specificity 52.6% (balanced accuracy 67.1%). Clinicians achieved 38.2% top diagnosis and 72.1% any-diagnosis accuracy, F1 = 0.776, sensitivity 67.3%, specificity 84.2% (balanced accuracy 75.8%). DermFlow recommended biopsy in 95.6% of cases vs. 82.4% for Claude, with multiple pairwise differences favoring DermFlow (p < 0.05). Conclusions: DermFlow demonstrated comparable or superior diagnostic performance to clinicians and superior performance to Claude in evaluating pigmented skin lesions. Although additional data must be gathered to further validate the model in real clinical settings, these initial findings suggest potential utility for dermatology-trained AI models in clinical practice, particularly in settings with limited dermatologist availability.

Background/Objectives: Demand for dermatologic services exceeds specialist capacity, with average wait times of 26–50 days in the United States. Current triage methods rely on subjective judgment or disease-specific indices that do not generalize across diagnoses or translate to operational decisions. We developed and validated the Dermatologic Complexity Score (DCS), a standardized instrument to guide case prioritization across dermatology care settings and evaluate DCS as a workload-reduction filter, enabling safe delegation of approximately half of routine teledermatology cases (DCS ≤ 40) away from specialist review. Methods: We conducted a prospective validation study of the DCS using 100 consecutive teledermatology cases spanning 30 common conditions. The DCS decomposes complexity into five domains (Diagnostic, Treatment, Risk, Patient Complexity, Monitoring) summed to a 0–100 total with prespecified bands: ≤40 (low) (41–70), (moderate) (71–89), (high), ≥90 (extreme). Five board-certified dermatologists and an automated module independently scored all cases. Two primary care physicians completed all ≤40 cases to assess feasibility. Primary outcomes were interrater reliability using ICC (2,1) and agreement with automation. Secondary outcomes included time-to-decision, referral rates, and primary care feasibility. Results: Mean patient age was 46.2 years; 47% of cases scored ≤40, 33% scored 41–70, 18% scored 71–89, and 2% scored ≥90. Interrater reliability was excellent (ICC (1,2)) = 0.979; 95% CI 0.974–0.983), with near-perfect agreement between automated and mean dermatologist scores (r = 0.998). Time-to-decision increased monotonically across DCS bands from 2.11 min (≤40) to 5 (90) min (≥90) (p = 1.36 × 10⁻¹⁴). Referral rates were 0% for ≤40, 3% for 41–70, 27.8% for 71–89, and 100% for ≥90 cases. DCS strongly predicted referral decisions (AUC = 0.919). Primary care physicians successfully managed all ≤40 cases but required 6–8 additional minutes per case compared to dermatologists. Conclusions: The DCS demonstrates excellent reliability and strong construct validity, mapping systematically to clinically relevant outcomes, including decision time and referral patterns. The instrument enables standardized, reproducible triage decisions that can optimize resource allocation across teledermatology, clinic, procedural, and inpatient settings. Implementation could improve access to dermatologic care by supporting appropriate delegation of low-complexity cases to primary care while ensuring timely specialist evaluation for high-complexity conditions.

Background and Objectives: Shortages of dermatologists create significant barriers to care, particularly for inflammatory and history-dependent conditions where image-only artificial intelligence (AI) classifiers have limited applicability. Current teledermatology solutions largely focus on single-task, morphology-based neoplasm classifiers, leaving the vast majority of dermatologic presentations underserved. This study evaluated the diagnostic accuracy and management plan quality of Dermflow (Prava Medical, Delaware, USA), a proprietary dermatology-focused Multimodal Image-and-Data Assistant (MIDA) that autonomously gathers dermatology-specific history, integrates data with patient-submitted images, and outputs structured differential diagnoses and management summaries. Materials and Methods: Two AI systems, Dermflow and Claude Sonnet 4 (Claude, a leading vision–language model), analyzed 87 clinical images from the Skin Condition Image Network and Diverse Dermatology Images databases, representing 10 inflammatory dermatoses and 9 neoplastic conditions stratified across Fitzpatrick Skin Tone (FST) categories (I–II, III–IV, V–VI). For the diagnostic comparison, Dermflow received images and autonomously gathered clinical history, while Claude received identical images without history. For the management plan comparison, both systems received the correct diagnosis and the clinical histories gathered by Dermflow. The primary outcome was diagnostic accuracy. The secondary outcome was management plan quality, assessed by two blinded dermatologists across eight clinical dimensions using 5-point Likert scales. Chi-square tests compared diagnostic accuracy between models; t-tests and ANOVA compared management quality scores. Results: Dermflow achieved markedly superior diagnostic accuracy compared to Claude (86.2% vs. 24.1%, p < 0.001). Both models maintained consistent diagnostic performance across FST categories without significant within-model differences (Dermflow p = 0.924; Claude p = 0.828). Management plan quality showed no significant overall differences between models. However, composite management quality scores declined significantly for darker skin tones across both systems: Dermflow scored 4.20 (FST I–II), 3.99 (FST III–IV), and 3.47 (FST V–VI); Claude scored 4.35, 3.97, and 3.44, respectively (p < 0.001 for most pairwise FST comparisons within each model). Conclusions: Multimodal AI integrating targeted history with image analysis achieves substantially higher diagnostic accuracy than image-only approaches across both inflammatory and neoplastic dermatologic conditions. Autonomous history gathering addresses fundamental limitations of morphology-only classifiers and enables scalable, patient-facing triage across the full spectrum of dermatologic disease. However, both models demonstrated reduced management plan quality for darker skin tones despite receiving the correct diagnosis, suggesting persistent training data limitations that require targeted bias-mitigation strategies beyond domain-specific instruction.

Background: Artificial intelligence (AI) has shown significant promise in augmenting di-agnostic capabilities across medical specialties. Recent advancements in generative AI allow for synthesis and interpretation of complex clinical data including imaging and patient history to assess disease risk. Objective: To evaluate the diagnostic performance of a dermatology-trained multimodal large language model (DermFlow, Delaware, USA) in assessing malignancy risk of pig-mented skin lesions. Methods: This retrospective study utilized data from 59 patients with 68 biopsy-proven pigmented skin lesions seen at Indiana University clinics from February 2023 to May 2025. De-identified patient histories and clinical images were input into DermFlow, and clinical images only were input into Claude Sonnet 4 (Claude) to generate differential diagnoses. Clinician pre-operative diagnoses were extracted from the clinical note. Assessments were compared to histopathologic diagnoses (gold standard). Results: Among 68 clinically concerning pigmented lesions, DermFlow achieved diagnostic performance of 47.1% accuracy for top diagnosis and 92.6% accuracy when the correct diagnosis was included within the limited differential (any-diagnosis accuracy). Claude performed significantly worse with 8.8% top diagnosis accuracy and 73.5% any-diagnosis accuracy. Clinicians achieved 38.2% top diagnosis accuracy and 72.1% any-diagnosis accuracy. DermFlow recommended biopsy in 95.6% of cases compared to 82.4% for Claude. Statistical analysis revealed several significant differences between DermFlow and both comparators (p < 0.05). Conclusions: DermFlow demonstrated comparable or superior diagnostic performance to clinicians and superior performance to Claude in evaluating pigmented skin lesions. Although additional data must be gathered to further validate the model in real clinical settings, these initial findings suggest potential utility for dermatology-trained AI models in clinical practice, particularly in settings with limited dermatologist availability.

Background: Histopathologic grading is an essential step in guiding treatment recommendations and follow-up for cutaneous lesions. However, inter-observer variability in determining differentiation of cutaneous squamous cell carcinoma (cSCC) and dysplasia of melanocytic nevi remain a challenge. Retrieval-augmented generation (RAG)-assisted artificial intelligence may offer educational and diagnostic support in this process. Methods: Two isolated RAG pathways using Claude 4.5 Opus, each grounded via ChromaDB vector retrieval of task-specific literature, graded 60 cSCC cases by differentiation and 67 melanocytic nevus cases by dysplasia. Each case was analyzed three times, with the majority result used as the final grade. Concordance with reference dermatopathologist grading was calculated with 95% Wilson score confidence intervals; inter-rater reliability was assessed using Cohen’s kappa. Results: For cSCC, the LLM achieved 90.0% concordance (95% CI, 79.9–95.3%) with excellent agreement (κ = 0.85, 95% CI, 0.74–0.96). For nevi, concordance was 44.8% (95% CI, 33.5–56.6%) with agreement not statistically distinguishable from chance (κ = 0.072, 95% CI, −0.10–0.24). Discordant nevus classifications skewed toward the moderately dysplastic category (65.7% of LLM classifications vs. 41.8% of reference classifications), and no severely dysplastic case was concordantly classified (0/6). Conclusions: Despite an identical RAG architecture, concordance with reference dermatopathologist grading was markedly task-dependent: high for cSCC differentiation, low for nevus dysplasia grading. This pattern parallels reported human inter-rater reliability for these tasks, suggesting that task-specific validation is needed before clinical or educational use.

Multimodal large language models are increasingly used for dermatologic queries, but whether Fitzpatrick skin type (FST) labels affect diagnostic accuracy is unknown. We evaluated 656 biopsy-confirmed photographs from the Diverse Dermatology Images (DDI) dataset under no-FST, DDI-concordant FST, and two DDI-discordant FST conditions using ChatGPT 5.2 Edu and Gemini 3.1 Pro browser configurations (5248 evaluations). Confirmed malignant diagnosis omission from the top three differential diagnoses (“lethal miss”) was a prespecified exploratory outcome. Gemini had higher ordinal accuracy than ChatGPT (odds ratio 2.32; 95% confidence interval 2.01–2.67; false discovery rate-adjusted p < 0.001). No FST prompt condition significantly improved accuracy over no-FST prompting; an equal-image-weighted sensitivity analysis yielded similar results. Among malignant images eligible for extreme discordance, extreme DDI-discordant prompting was associated with more ChatGPT lethal misses than no-FST prompting (paired odds ratio 5.0; p = 0.043); Gemini showed no significant paired shift (p = 1.00). Post hoc binary malignancy detection showed low no-FST specificity for ChatGPT (42.5%) and Gemini (22.9%), indicating frequent overcalling. Gemini showed lower no-FST accuracy for malignant FST V–VI images than for FST I–II and III–IV images. FST prompting provided no measurable diagnostic benefit. Neither configuration supports autonomous dermatologic use.