EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Nonstress Test Classification by Obstetric Clinicians and Large Language Models: A Comparative Study

Melis Altuğ Inan, Bilge Kapudere, Nisan Helin Donmez, Hanne Bulat Cim, Reyhan Ayaz Bilir, Abdulkadir Turgut

International Journal of Clinical Practice · 2026-01 · 2026 (1)

DOI: 10.1155/ijcp/7027485 ↗

Kurz gefasst

Fachartikel in der Zeitschrift International Journal of Clinical Practice, online erschienen am 15.11.2026. Verfasst von Melis Altuğ Inan und 5 weiteren Autor:innen. Die Arbeit nennt 24 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Objective

To compare nonstress test (NST) classification performance between obstetric clinicians and contemporary large language models (LLMs), and to evaluate whether deliberate reasoning strategies improve classification performance.

Methods

In this prospective cross‐sectional study, 110 obstetric clinicians (49 residents and 61 specialists) independently evaluated 15 deidentified NST tracings representing Categories I–III. Reference classifications were established by consensus of two maternal–fetal medicine specialists using guideline‐based criteria. The same tracings were assessed by ChatGPT (v5.2), Gemini 3, and Grok under standard and deliberate reasoning conditions. Category classification accuracy, component‐level accuracy, and agreement with the reference standard were evaluated.

Results

Specialists demonstrated higher category classification accuracy than residents (median 0.600 vs. 0.467; Mann–Whitney U = 950; p = 0.002). Clinician interobserver agreement was moderate (Fleiss’ κ = 0.399; 95% bootstrap CI, 0.195–0.523), with lower observed agreement in Category II than Category I tracings. Under standard conditions, category accuracy was 100.0% for Gemini (95% CI, 78.2–100.0) and 60.0% for both ChatGPT and Grok (95% CI, 32.3–83.7); corresponding Cohen’s κ values were 1.000, 0.196, and 0.313. Deliberate reasoning did not consistently improve performance: category accuracy was 73.3% for ChatGPT, 60.0% for Gemini, and 66.7% for Grok.

Conclusion

LLM performance in NST classification varied across models and analytic conditions. Deliberate reasoning did not provide a consistent benefit, and clinician performance also varied across cases, particularly in Category II tracings. These findings support cautious interpretation of LLM‐based NST classification and the need for broader external validation before routine clinical implementation.

Zitieren

APA

Inan, M. A., Kapudere, B., Donmez, N. H., Cim, H. B., Bilir, R. A., & Turgut, A. (2026). Nonstress Test Classification by Obstetric Clinicians and Large Language Models: A Comparative Study. International Journal of Clinical Practice, 2026(1). https://doi.org/10.1155/ijcp/7027485

Vancouver

Inan MA, Kapudere B, Donmez NH, Cim HB, Bilir RA, Turgut A. Nonstress Test Classification by Obstetric Clinicians and Large Language Models: A Comparative Study. International Journal of Clinical Practice. 2026;2026(1). doi:10.1155/ijcp/7027485

BibTeX
@article{Inan2026,
  author = {Inan, Melis Altuğ and Kapudere, Bilge and Donmez, Nisan Helin and Cim, Hanne Bulat and Bilir, Reyhan Ayaz and Turgut, Abdulkadir},
  title = {Nonstress Test Classification by Obstetric Clinicians and Large Language Models: A Comparative Study},
  journal = {International Journal of Clinical Practice},
  year = {2026},
  volume = {2026},
  number = {1},
  doi = {10.1155/ijcp/7027485}
}
Kostenlos starten →