EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Diagnostic performance of artificial intelligence in the syncope unit

Steven van Zanten, Thomas T. Boel, Jelle SY de Jong, Egbert M Koomen, Babette Bais, Ako Dara, Freek Giele, Christiaan Geertsma, Richard Sutton, Mike G Scheffer, Joris R de Groot, Frederik J de Lange

Exploration of Cardiology · 10.07.2026 · 4

DOI: 10.37349/ec.2026.1012114 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Exploration of Cardiology, online erschienen am 10.07.2026. Verfasst von Steven van Zanten und 11 weiteren Autor:innen. Die Arbeit nennt 50 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Aim

Artificial intelligence may support syncope evaluation, but reliability of language models in structured syncope care remains uncertain. We evaluated diagnostic performance, safety, and within-case consistency of Generative Pre-trained Transformer-5 (GPT-5) in patients with transient loss of consciousness (T-LOC).

Methods

This prospective cohort study included 55 patients evaluated in syncope units. GPT-5 and a syncope-expert assessed identical case information after core evaluation (CE: history-taking, physical examination, active standing test, and 12-lead electrocardiogram) and extended evaluation (EE: CE plus additional testing when indicated). An expert panel adjudicated the final diagnosis after 18 months. Outcomes were diagnostic yield, final-diagnosis inclusion rate, diagnostic precision score (DPS), cardiac diagnostic safety, and within-case consistency across five repeated GPT-5 runs.

Results

Of 55 patients, 54 had complete follow-up for performance analyses. Diagnostic yield was 94% for the syncope-expert at CE and EE, and 100% and 96% for GPT-5 at CE and EE, respectively. GPT-5 included the final diagnosis in 52% (CE) and 57% (EE) of cases, versus 67% for the syncope-expert. DPS remained negative for GPT-5 at CE (mean −0.03, SD 0.54) and EE (mean −0.01, SD 0.49). Among four final cardiac syncope cases, GPT-5 selected the final diagnosis in one case and the syncope-expert in three. First-diagnosis consistency across five GPT-5 runs was 69% after CE and 74% after EE.

Conclusions

GPT-5 generated diagnostic outputs frequently but showed limited precision, cardiac safety concerns, and within-case variability. Its role in syncope evaluation should remain supportive within clinician-led pathways rather than autonomous.

Zitieren

APA

Zanten, S. V., Boel, T. T., Jong, J. S. D., Koomen, E. M., Bais, B., Dara, A., Giele, F., Geertsma, C., Sutton, R., Scheffer, M. G., Groot, J. R. D., & Lange, F. J. D. (2026). Diagnostic performance of artificial intelligence in the syncope unit. Exploration of Cardiology, 4. https://doi.org/10.37349/ec.2026.1012114

Vancouver

Zanten SV, Boel TT, Jong JSD, Koomen EM, Bais B, Dara A, et al. Diagnostic performance of artificial intelligence in the syncope unit. Exploration of Cardiology. 2026;4. doi:10.37349/ec.2026.1012114

BibTeX
@article{Zanten2026,
  author = {Zanten, Steven van and Boel, Thomas T. and Jong, Jelle SY de and Koomen, Egbert M and Bais, Babette and Dara, Ako and Giele, Freek and Geertsma, Christiaan and Sutton, Richard and Scheffer, Mike G and Groot, Joris R de and Lange, Frederik J de},
  title = {Diagnostic performance of artificial intelligence in the syncope unit},
  journal = {Exploration of Cardiology},
  year = {2026},
  volume = {4},
  doi = {10.37349/ec.2026.1012114}
}
Kostenlos starten →