Home › Cardiac (non-arrest) › LLMs unreliable for cath-lab activation ECGs

Cardiac (non-arrest)

LLMs unreliable for cath-lab activation ECGs

Large language models fail to reliably predict emergent catheterization laboratory activation from prehospital electrocardiograms
Heart & Lung · Added 17 September 2026

Retrospective test of three multimodal LLMs against real cardiology cath-lab activation decisions on 615 prehospital ECGs found the best (Gemini, 95.3% sensitivity) had 9.4% specificity, and the standard ECG machine algorithm beat all three LLMs overall — general-purpose LLMs aren't reliable for STEMI pathway activation.

Study type

Other

Design

Retrospective diagnostic accuracy study comparing three multimodal LLMs and ECG machine algorithm against cardiology decisions

Participants

615 prehospital ECGs from 270 EMS encounters

Main finding

Gemini: sensitivity 95.3% but specificity 9.4%; ChatGPT and Claude sensitivity 68.1% and 67.2% with specificity 42.3% and 46.5%; ECG machine algorithm sensitivity 67.7%, specificity 64.2% — outperforming all LLMs overall.

Caveat

Retrospective single reference standard (cardiology activation) and no Australian validation; LLM performance may change with newer models.

Bottom line

Context only

Learning goal

Weigh the role of emerging decision-support tools against traditional ECG interpretation when activating the catheterisation laboratory.

More

• Topic: Cardiac (non-arrest)
• Daily scan: 17 September 2026

Related papers on Cardiac (non-arrest)

• Only 60% of paramedic-suspected STEMI confirmed

  Annals of Emergency Medicine · 9 Oct 2026 · Registry · Context only

  A retrospective Victorian statewide ambulance registry cohort (2020–25) followed up paramedic-suspected STEMI presentations and found STEMI confirmed in only 60% of those with a recorded discharge diagnosis, with 30-day mortality of 6.5%.

• Prehospital STEMI data transmission speeds reperfusion

  Clinics · 3 Oct 2026 · Cohort · Supports current practice

  In 409 Chinese STEMI patients, prehospital WeChat transmission of patient data to the PCI centre shortened symptom-to-first-medical-contact (75 vs 120 minutes), door-to-balloon (50 vs 67 minutes) and total ischaemic time (121.5 vs 150 minutes).

• 48-hour AF cardioversion threshold abandoned

  Emergency Medicine Australasia · 22 Sep 2026 · Guideline / statement · Context only

  EMA commentary arguing the once-conventional 48-hour elective cardioversion threshold for recent-onset atrial fibrillation is no longer supported

• Aortic emergencies: HR/BP drugs in transport

  Prehospital Emergency Care · 22 Sep 2026 · Cohort · Context only

  Cohort of 1,530 aortic emergencies in US critical care transport — esmolol, labetalol and nicardipine strategies hit HR/BP targets at similar rates, with esmolol alone showing the highest hypotension incidence (9.3%) and labetalol alone the lowest (2.1%)