Diagnostic accuracy of artificial intelligence versus 263 pediatric clinicians for childhood exanthems

Clinical Snapshot

Key Findings

• P Value: P < .001 for specialist vs resident comparison

• Effect Size: ChatGPT: 86.9% accuracy (53/61 cases), Gemini: 82.0% (50/61), Copilot: 72.1% (44/61)

• Primary Outcome: Diagnostic accuracy of AI models versus pediatric clinicians

• Nnt Or Sensitivity: ChatGPT and Gemini exceeded specialist 95% CI upper bound (47.17%); disease-specific accuracy ranged 0-100%

• Confidence Interval: Not provided for AI models

Clinical Application

Requires standardized clinical photography, structured data input, and integration with existing clinical workflows. Physician oversight essential for context-dependent diagnoses. Relevant to Australian pediatric practice but would require validation in local population. Integration with existing telehealth frameworks could enhance rural/remote diagnostic support. TGA approval pathway unclear for diagnostic AI tools. Pediatric patients presenting with exanthematous diseases in settings where clinical photographs and structured data collection are feasible

Abstract

Pediatric exanthematous diseases pose diagnostic challenges because clinical presentations overlap. To determine whether current artificial intelligence (AI) models achieve diagnostic accuracy within or above the performance distribution of pediatric residents and specialists for common rash-associated diseases. Participants and AI models were evaluated against definitive diagnoses confirmed by clinical features, laboratory findings, and consensus of two pediatric infectious disease specialists. A volunteer sample of 263 pediatric clinicians: 107 residents (years 1 through 4) and 156 specialists. Each clinician completed a blinded multiple-choice questionnaire with a clinical photograph and accompanying clinical data per case. The same cases were presented to three AI models: ChatGPT, Gemini, and Copilot. Among 263 clinicians (107 residents, 156 specialists), specialists scored higher than residents (median, 46 [IQR, 42-50] vs 41 [IQR, 36-46]; P < .001; r = 0.32). ChatGPT correctly diagnosed 53 of 61 cases (86.9%), Gemini 50 (82.0%), and Copilot 44 (72.1%). Both ChatGPT and Gemini exceeded the upper bound of the specialist population median 95% CI (47.17). All three AI models scored above the resident 95% CI upper bound (42.76). Disease-level accuracy ranged from 0% (insect bites, all models) to 100% (9 conditions, all models). Fourth-year residents scored higher than first- and second-year residents (P = .001; ε2 = 0.13).Conclusions: AI models given clinical data alongside images matched or exceeded specialist-level performance for pediatric exanthems. Accuracy varied by disease; failures clustered in conditions that require contextual reasoning. Physician oversight remains necessary where AI accuracy is lowest. What is Known • Childhood exanthematous diseases pose significant diagnostic challenges due to their overlapping clinical presentations. • Artificial intelligence models are becoming increasingly proficient in accurately diagnosing these conditions. What is New • In this diagnostic accuracy study of 61 cases evaluated by 263 clinicians and 3 artificial intelligence models, ChatGPT (86.9%) and Gemini (82.0%) exceeded the 95% CI upper bound of the specialist population median. Disease-level accuracy ranged from 0 to 100% across models • Artificial intelligence models given clinical data alongside images can match or exceed specialist-level accuracy for common pediatric exanthems, but failures in context-dependent diagnoses require physician oversight.

References

1. 1.Gençeli, M., Akcan, M., Soran, Ö. B., Çokbiçer, A., Saraç, U., Üstüntaş, T., Yücel, M., Doğan, M., Yılık Kömür, E., Gençeli, S., Yılmaz Dağlı, H., Sarı, M., Kılıç, A. O., Şahin, S., & Akkuş, A. (2026). Diagnostic accuracy of artificial intelligence versus 263 pediatric clinicians for childhood exanthems. *European Journal of Pediatrics*. https://doi.org/10.1080/00031305.1978.10479236