ORIGINAL RESEARCH
Dietrich von Kuenssberg Jehle, MD; Micah K. Browne, BS; Hailey P. Hoffmann, MD; Krishna K. Paul, BS; Matthew M. Talbott, DO; Kendall M. Campbell, MD; Anthony W. Linfante, MD; Bjorn A. Sorensen, MD; Michael G. Wilkerson, MD
Corresponding Author: Dietrich von Kuenssberg Jehle, MD; Department of Emergency Medicine, The University of Texas Medical Branch.
Email: dijehle@utmb.edu
DOI: 10.3122/jabfm.2025.250260R3
Keywords: Artificial Intelligence, Deep Learning, Dermatology, Emergency Medicine, Machine Learning, Medical Errors, Primary Health Care, Prospective Studies, Skin Cancer
Dates: Submitted: 07-08-2025; Revised: 08-31-2025; 09-26-2025; 10-08-2025; Accepted: 10-27-2025
Status: In production.
INTRODUCTION: Artificial intelligence (AI) and machine learning are increasingly used in dermatology, primarily for diagnosing skin cancer and classifying disease severity. Primary care providers (PCPs) and emergency physicians have higher error rates in diagnosing dermatologic conditions compared to dermatologists. This study evaluates the performance of an AI-driven dermatologic image analytics tool in identifying common skin disorders from classical images used in medical education.
METHODS: This prospective study analyzed 42 classical images of common skin disorders using the bellePro™ application (version 2.1.0), trained on over 400,000 dermatologic images. A set of dermatologic images used in a national board review course, underwent blinded and independent validation by an academic dermatology department. This module provides AI-powered predictions from cell phone photos, ranking them by "image match scores" (higher values indicate better predictions). The primary outcome was the accuracy of three image predictions in identifying known diagnoses, along with the corresponding "image match scores."
RESULTS: The AI tool correctly predicted 38 out of 42 skin disorders (positive predictive value was 90.5%) as the top differential on the first attempt for all three attempts. Four conditions (angioedema, squamous cell carcinoma, hand foot and mouth disease, and chickenpox) were included in the prediction differential list each attempt but never ranked first, bringing the total to 42/42 (100%). None of the correctly identified disorders had an average "image match score" below 0.64. The average "image match score" for correct diagnoses was 0.896 (SD - 0.098).
CONCLUSIONS: The findings suggest that this AI-based dermatologic image analytics tool performs effectively on classical images of common skin disorders typically seen in clinical practice. This has potential to serve as an adjunct for providers to improve patient outcomes in environments lacking timely access to dermatology resources.