2026/07/08 by Yunpeng Wei, Hong Xu, Hu Yuan · 1 voice
Health Professions · Medicine · #Artificial Intelligence in Healthcare and Education #Cutaneous Melanoma Detection and Management #Patient-Provider Communication in Healthcare
paper · pdf · doi:10.3389/fmed.2026.1875075
openalex publication_date 2026/07/08 · openalex created_date 2026/07/09 · openalex updated_date 2026/07/26
Background Generative artificial intelligence (GenAI), particularly large language models (LLMs), is rapidly integrating into clinical settings. However, its net effect on dermatological practice remains poorly defined. This study investigates the dual impact of GenAI on clinician–patient communication using a multi-component data approach. Methods We conducted an exploratory multi-component cross-sectional study from February 2025 to January 2026 in the dermatology departments of two tertiary hospitals in China. This study included physician surveys ( n = 25), outpatient questionnaires ( n = 60), and standardized case assessments. For the standardized component, 20 physicians completed two cases under randomized no-AI and AI-assisted pre-consultation preparation sequences, yielding 80 assessment records. The AI-assisted condition used DeepSeek-R1 for 5 min of pre-consultation preparation before a 10-min standardized patient encounter. Results Physician GenAI use frequency was associated with lower emotional exhaustion (rs = −0.692, p = 0.002) and higher communication self-efficacy (rs = 0.848, p < 0.001). In the adjusted physician model, GenAI use frequency was associated with higher PCSES scores ( B = 3.204, 95% CI 2.263 to 4.145), and this association was maintained in leave-one-out analyses ( B range 3.010–3.374), a parsimonious model ( B = 3.151), and a model excluding influential observations ( B = 3.055). Patient GenAI users reported higher communication satisfaction than non-users (18.46 ± 2.48 vs. 15.17 ± 2.43, p < 0.001). In standardized cases, the AI-assisted condition was associated with higher estimated marginal scores for information gathering (difference 3.175, 95% CI 0.605 to 5.745), information giving (5.675, 95% CI 3.432 to 7.918), structural efficiency (5.575, 95% CI 3.751 to 7.399), and total score (2.490, 95% CI 1.374 to 3.606). Humanistic care showed a reduction in the AI-assisted condition in the primary mixed-effects model; however, this difference did not reach statistical significance. Sensitivity analyses using fully adjusted mixed-effects models accounting for period, condition order, and case order indicated a consistent negative effect of AI assistance on humanistic care ( B = −3.414, 95% CI −5.723 to −1.105), indicating the sensitivity of this outcome to model specification, with a consistent negative effect observed in fully adjusted models. Conclusion These exploratory findings suggest that GenAI-assisted preparation was associated with stronger information organization and communicative efficiency in dermatology, while not automatically improving empathic or humanistic communication. GenAI should therefore be positioned as a supervised supportive tool rather than as a replacement for clinical judgment or relational care.