Experts caution against overreliance on AI in dermatology at Korea Derma symposium

As artificial intelligence (AI) advances in healthcare, experts warn that the technology remains in its early stages, especially in dermatology, and is not yet equipped to fully transform medical practices. 

At the Korea Derma International Aesthetic Dermatology Symposium last weekend, hosted by the Association of Korean Dermatologists, an international panel of experts highlighted that although AI shows potential for aiding in diagnostics and treatment, it lacks the essential human judgment needed for managing complex skin conditions.

“While AI can help trainees in structuring patient notes, it falls short in delivering the nuanced care required by experienced clinicians,” said Professor Jane Yoo of New York’s Icahn School of Medicine at Mount Sinai during a presentation in Seoul last Friday. She emphasized the ongoing challenges in balancing technological innovation with patient care before AI can be reliably integrated into clinical practice.

Yoo presented a review of 35 AI-generated transcripts published in the NEJM Catalyst Journal, which found that AI failed to meet 48 out of 50 key criteria needed to replace clinicians. The study showed frequent inconsistencies in the AI-generated notes, requiring physician review and edits—casting doubt on AI’s reliability as it is often portrayed.

For AI-generated notes to be truly effective, Yoo said, they must be organized, clear, concise, internally consistent, free from "hallucinations" (where AI generates false information) and impartial. 

Yoo recently piloted an app called Suki in her practice, which integrates with the Athenahealth electronic medical record system. With patient consent, she recorded visits using the app, which generated SOAP notes (structured documentation of patient encounters) for her review within minutes.

“At first, I thought this would be a great time-saver, handling documentation in real-time,” Yoo said. But the experience fell short. She soon found herself making adjustments after each visit. “What seemed like a helpful app became an additional burden,” she said. “Medication names were often misspelled, and, alarmingly, the AI generated diagnoses that were never discussed, creating a misleading record.”

Yoo explained that dermatologists need AI tools that are straightforward and genuinely improve clinic efficiency. For simple tasks, like suture removal, AI tools might be useful, but for more complex procedures, such as biopsies or prescribing biologic medications, they often don’t perform as well as anticipated, she said. 

“AI can be a valuable learning tool for trainees—like physician assistants or medical students—showing them how to structure patient notes,” Yoo emphasized. “However, for experienced providers, it doesn’t necessarily expedite patient visits.”

In Indonesia, where a severe shortage of dermatologists affects rural areas especially, AI could play a critical role in dermatological care, said Anesia Tania Icksan, a dermatologist at Jakarta’s ZAP Premiere. 

She noted that AI can assist with diagnosis, treatment, and monitoring, particularly for pigmentary disorders. However, Icksan highlighted significant implementation challenges, ethical concerns, and limitations that need addressing before AI can be fully integrated into clinical practice.

On Friday, she presented a study published in The Lancet Digital Health by researchers at the University of Sydney which assessed an AI algorithm for mobile app-based teledermatology using a dataset of over 80,000 clinical, dermoscopic, and histopathologic images. The findings revealed that the AI system performed better than novice dermatologists in diagnosing skin conditions but was less accurate than experienced clinicians.

“A major issue is the lack of diversity in many datasets, which predominantly feature Fitzpatrick skin types 1 to 3,” Icksan said. The study also evaluated AI’s accuracy with skin types 4 to 6, finding an overall accuracy of 86.5 percent. While the model predicted benign lesions with 90 percent accuracy, its accuracy for malignant lesions was only about 70 percent.

“AI could outperform some clinicians in these assessments, but dermatologists in a clinical setting can examine lesions from multiple angles, feel the skin, and gather information through direct interaction—factors that AI cannot replicate,” Icksan explained. She added that in terms of treatment, AI is still in an “experimental phase,” with no validated applications yet available for clinical use. 

In the future, AI might help evaluate treatment effectiveness, predict responses, and personalize treatment plans, though Icksan warned there’s “a long way to go before AI can reliably perform in clinical settings,” especially in treatment and management.

Icksan emphasized that AI’s accuracy depends heavily on the quality and diversity of the data it’s trained on, including representative skin types. "Dermatologists should view AI as a supportive tool to enhance our practice, not as a replacement, particularly for managing pigmentary disorders,” Icksan said.