ChatGPT-5, Gemini and DeepSeek ace pediatric infection questions, study finds

Pediatric infectious diseases have always demanded a particular kind of clinical vigilance. A fever in a newborn is not the same as a fever in a ten-year-old; a runny nose in an immunized toddler can hide something far more serious than it would in an adult. Now, as families arrive at clinics armed with answers generated by artificial intelligence, and as medical students quietly rehearse case presentations with chatbots instead of textbooks, pediatricians are asking a question that has moved from curiosity to necessity: how good are these systems, really, at the age-specific reasoning that defines their field?

A new study published in Pediatric Research offers one of the most structured answers to date. Kubra Aykac of Hacettepe University in Ankara and an international team spanning Korea, Japan, Poland, Hong Kong, and Germany evaluated three of the most widely used large language models — ChatGPT-5, Gemini 2.5 Flash, and DeepSeek-V3.2-Exp — across twelve carefully constructed pediatric infectious disease scenarios. The verdict was striking: all three models performed at near-perfect levels, and the differences between them were too small to reach statistical significance.

The design of the study reflects how clinical reasoning actually unfolds. The researchers selected six core pediatric infectious disease conditions, each represented by two scenarios that differed in patient age or illness severity, producing twelve cases in total. Each case was paired with five open-ended questions covering the essential domains of clinical thinking: etiology, clinical features, diagnostic evaluation, treatment options, and potential complications. That yielded sixty questions, and when each was posed to all three models, the exercise generated 180 responses, each of which was independently reviewed by two board-certified pediatric infectious disease specialists — 360 individual evaluations in all.

The scenarios were not drawn casually. The team applied three complementary criteria: clinical relevance and severity, epidemiologic representativeness, and educational applicability. High-acuity, time-critical conditions such as suspected acute bacterial meningitis sat alongside moderate-acuity presentations that commonly drive diagnostic or therapeutic errors. Respiratory and gastrointestinal infections — the leading causes of global pediatric infectious morbidity, as documented by the World Health Organization — anchored the case mix. And every scenario mapped to competencies universally taught in undergraduate medical curricula: focused history-taking, red-flag recognition, initial stabilization, first-line investigations, and escalation thresholds. Cases dependent on country-specific policies or resource-contingent pathways were deliberately excluded to preserve fairness across health systems.

Scoring followed a purpose-built Likert rubric covering three dimensions. Accuracy, rated on a six-point scale, measured factual correctness, clinical appropriateness, and alignment with established pediatric infectious disease guidelines. Completeness, on a three-point scale, assessed whether responses incorporated essential concepts and sufficient detail. Safety, a binary measure, flagged unsafe, misleading, or potentially harmful recommendations. Two specialists scored every answer while blinded to which model had produced it. An independent coordinator stripped away model names, platform branding, self-identifying phrases, timestamps, and interface artifacts, then reformatted everything into uniform plain text with standardized headings and randomized presentation order. Unblinding occurred only after primary analyses were finalized.

The results were remarkable for their consistency. ChatGPT-5 achieved perfect domain scores of 10.0 across all five clinical reasoning categories. DeepSeek-V3.2-Exp was nearly identical, scoring 10.0 in etiology, clinical findings, and complications, with only minor reductions in diagnostics (9.83) and treatment (9.92). Gemini 2.5 Flash showed somewhat more variability — 9.75 in diagnostics and 8.83 in treatment — while matching the other models perfectly in clinical findings and complications and scoring 9.83 in etiology. Across every domain-level comparison, the Friedman test found no statistically significant differences (all p > 0.05).

Model-level analyses told the same story. Accuracy scores were 6.0 for ChatGPT-5, 5.98 for DeepSeek-V3.2-Exp, and 5.85 for Gemini 2.5 Flash (p = 0.074). Completeness followed the same gradient: 3.0, 2.97, and 2.87 respectively (p = 0.066). Safety scores were 1.0 for both ChatGPT-5 and DeepSeek-V3.2-Exp and 0.97 for Gemini 2.5 Flash (Cochran’s Q p = 0.14). None of these differences reached statistical significance, suggesting that when given the same pediatric cases, the current generation of language models tends to produce similar clinical reasoning outputs.

The trajectory is as telling as the snapshot. In the team’s earlier study on pediatric pneumonia learning, ChatGPT-3.5 performed notably lower than DeepSeek-V3 and Gemini 2.0, especially on questions requiring age-specific reasoning or interpretation of etiologic clues. In the present analysis, the newer ChatGPT version reached scores similar to, and at times higher than, the other two systems. This pattern aligns with broader findings that updated model architectures with GPT-4-level reasoning now approach clinician performance in selected diagnostic tasks under controlled conditions, and that successive model versions incorporate broader medical knowledge, stronger safety mechanisms, and improved handling of clinical context.

Yet the authors are careful about what these numbers do and do not mean. Expert reviewers identified recurring qualitative gaps even among the high scores — particularly in age-specific decision thresholds such as vital sign cutoffs and hospitalization criteria, safety-oriented escalation criteria, and the depth of treatment recommendations, including dosing precision and follow-up planning. High numerical performance, the study emphasizes, does not necessarily equate to fully reliable clinical reasoning in pediatric practice. The rubric itself was designed primarily to flag clinically unsafe or clearly insufficient responses rather than to distinguish minor qualitative differences among high-performing models, and most scores clustered at or near the top of the scale — a ceiling effect that limits discrimination among models that are all performing well.

The study’s limitations are acknowledged candidly. Twelve predefined scenarios, however thoughtfully constructed, cannot reflect the full variety of situations encountered in everyday clinical practice. The cases were structured to evaluate clinical reasoning, management accuracy, and patient safety based on presented findings, but were not specifically designed to assess broad differential diagnosis generation in fully undifferentiated presentations. All evaluations relied on human reviewers, and despite predefined criteria, some subjective judgment is inherent to expert-based assessment. The consumer platforms used do not disclose specific version numbers and are updated continuously, so the exact underlying checkpoints cannot be independently verified — a reproducibility limitation the authors openly acknowledge.

What emerges is a picture of language models as competent but supervised assistants. The findings support their use in medical education and early clinical reasoning — helping trainees organize information, revisit disease summaries, and outline diagnostic steps — while making clear that their optimal role remains within clinician-supervised workflows. AI-generated outputs complement, rather than replace, human clinical judgment. As the authors note, a combined human-AI approach appears to be a safe and practical framework for integrating these systems into pediatric infectious disease practice, and future studies will require more sensitive evaluation tools and more clinically challenging scenarios to identify meaningful differences between current-generation models.

Subject of Research: Performance of large language models in pediatric infectious disease education: a multi-condition comparative study

Article Title: Performance of large language models in pediatric infectious disease education: a multi-condition comparative study

Article References: Aykac, K., Cubuk, O., Demir, O. O., Choe, Y. J., Kawata, S., Ludwikowska, K., Kwan, M. Y. W., Aydin, M., & Ozsurekci, Y. (2026). Performance of large language models in pediatric infectious disease education: a multi-condition comparative study. Pediatric Research. https://doi.org/10.1038/s41390-026-05578-0

Image Credits: AI Generated

DOI: 10.1038/s41390-026-05578-0

Keywords: Performance, large, language, models, pediatric, infectious, disease, education, multi-condition, comparative, scientific research