Efficacy of Artificial Intelligence‐Powered Software in Enhancing Diagnostic Accuracy in Medical Residents

Abstract

Artificial intelligence (AI) tools are increasingly used for diagnostic support [1]. VisualDx is a commercially available AI-powered clinical decision support system used in both dermatology and general medicine [2]. Despite growing interest in AI applications within dermatology, the evaluation of the impact of such tools in improving diagnostic performance remains limited. This study aimed to evaluate the effectiveness of VisualDX as a diagnostic support tool for improving the performance of medical residents in Tunisia. We conducted a cross-sectional study from January to June 2025. Fifty clinical scenarios (each including clinical images) were hand-picked by two board-certified dermatologists to ensure representation of a clinically relevant, diverse spectrum of dermatological conditions. The selection was deliberately non-random to ensure diverse coverage, including infectious dermatoses, inflammatory conditions, pigmentary disorders, adnexal diseases, and cutaneous tumors. All Tunisian residents who completed their dermatology rotation during their training were invited to complete an online study form. First, participants selected the most likely diagnosis for each clinical scenario. Then, the same images were presented, accompanied by the top five VisualDx suggestions, and participants were able to reassess their diagnoses. The diagnostic accuracy score was calculated for each participant (one correct answer: One point, maximum score: 50). Scores were compared using the Wilcoxon test, with the effect size determined for every analysis. A p-value < 0.05 was considered statistically significant. A total of 113 residents completed the study, including 43 (38%) dermatology residents, 27 (24%) family medicine residents, and 43 (38%) residents in other specialties. The overall diagnostic accuracy increased from 46.3% to 62.3% with VisualDx assistance (Table 1). Subgroup analysis showed significant improvement among first- and second-year dermatology residents (p = 0.001), while no significant change was observed among third- and fourth-year residents (p = 0.771) (Table 1). Diagnostic scores among other specialty residents showed substantial improvement, with meaningful gains observed across all training levels (effect size ranging between 0.8 and 0.9) (Table 1). Our findings showed that AI assistance can improve diagnostic accuracy among residents. The most notable improvements were observed among non-dermatology residents and junior trainees. By offering diagnostic suggestions and curated image examples, AI systems can help residents refine their pattern recognition and strengthen their differential diagnoses [2]. Overall, these findings align with prior research showing that AI can serve as a complementary aid to human expertise in dermatology [3, 4]. While many studies focus on AI as an independent diagnostic tool, our work emphasizes its role in enhancing human performance, consistent with the intended use of clinical decision support systems. This study has limitations. Residents evaluated static images rather than interacting with patients, and VisualDx interactivity was mediated rather than fully explored by participants. In conclusion, AI-powered systems hold promise as supportive tools in dermatology, particularly for early-career clinicians and non-specialists. Issues of data privacy, medicolegal responsibility, and ethical deployment must be carefully navigated, particularly in dermatology, where images may include identifiable features [5]. Over-reliance on diagnostic decision-support tools, particularly among junior residents, poses an educational risk by potentially hindering the development of independent clinical reasoning; this can be mitigated by integrating these tools into a structured curriculum that emphasizes their use for validation and knowledge expansion rather than primary diagnosis generation. The authors have nothing to report. Informed consent was obtained from the patients (or their legal guardians) when the clinical photographs were taken. No author has financial ties to VisualDx or provides any kind of paid or unpaid assistance to VisualDx. One of the authors, Dr. Noureddine Litaiem, holds a one-year license for VisualDX, which was provided to him in his capacity as a member of the Editorial Board of the International Journal of Dermatology. The VisualDx team was formally notified by two co-authors, Dr. Noureddine Litaiem and Dr. Yosr Daoud, on August 26, 2024, that their software would be utilized for this study. It is important to state that they had no involvement in the study's protocol development, its design, or the selection of the clinical scenarios used. The data that support the findings of this study are available from the corresponding author upon reasonable request.