An AI-powered dermatology program led to more confident, accurate diagnoses, cohort study shows

When viewing photos of dermatological problems, people who used an artificial intelligence (AI) prototype were more confident in naming a possible diagnosis and were more accurate when doing so, according to research published recently in JAMA Dermatology, called ‘Consumer Understanding of Skin Concerns With an AI-Powered Informational Tool.’ The findings highlight the growing role of AI in supporting patient understanding of common health concerns, particularly in areas where access to specialists may be limited or delayed.

A group of researchers, including lead author Rory Sayres, Ph.D., an AI health researcher at Google, enrolled 2,345 individuals and placed them randomly into one of three arms. These individuals were shown photos of dermatology cases. All groups were asked to imagine the presented skin issues as their own dermatological concerns. Photos were from 127 deidentified case patients from a teledermatology service.

The control group could use existing tools like web searches to identify conditions, reflecting how many people currently seek health information online.

The AI group had access to predictions from a prototype AI app designed to assist with image-based diagnosis.

The third group (Wizard of Oz method) had access to the same AI interface as Group 2, but it used dermatologist panels to give advice instead of the AI-generated predictions, allowing researchers to compare human expertise directly with AI outputs.

The results showed that the control group accuracy was 7.9%, the AI group accuracy was 22.8% and participants in the “Wizard of Oz” group demonstrated a 36.2% accuracy rate.

When it came to confidence in identifying a condition, 41.2% of participants in the control group showed confidence, compared with 62.3% in the AI-powered arm.

While accuracy and confidence increased with AI use, this was not true when it came to the participants’ understanding of the next steps needed to be taken after identifying a condition. This suggests that although AI can support recognition, it may not yet fully guide users toward appropriate care decisions.

“These findings suggest that AI applications may be able to help consumers understand the condition depicted in a case, but further progress remains in improving user understanding of possible next steps,” Sayres and his colleagues write in the study.

Skin conditions affect approximately 2 billion individuals worldwide, making them among the most common health issues globally. Yet only 28% of reported skin conditions are evaluated by dermatologists, leaving a significant gap in diagnosis and treatment, the report says.

Access to dermatological care is sometimes limited by factors such as long wait times, geographic barriers, and cost, particularly for underserved populations such as Medicaid recipients.

“AI–based tools present an opportunity to bridge the access gap by empowering clinicians, consumers, or both,” Sayres and his colleagues continue. “AI algorithms have demonstrated performance comparable to that of panels of trained specialists, and AI assistance has been associated with increased diagnostic agreement between dermatologists and nonspecialist clinicians.”

The authors do warn that AI-based interventions come with several risks, including inaccurate predictions and increased anxiety. They also include that this study did not account for skin tone. Other studies have shown that AI-assisted tools can lead to lower diagnostic accuracy in darker skin tones, which could further drive gaps in healthcare equality.