Truace
Subscribe

AI chatbot reads heart scans to locate irregular beats in 157 patients, matching another AI method

Study title: Large Language Model-Based Localization of Premature Ventricular Contraction Origins: A Retrospective Diagnostic Accuracy Study

Researchers tested whether large language models could interpret 12-lead ECG images to distinguish left- versus right-sided origins of premature ventricular contractions in 157 patients who had undergone successful catheter ablation. By August 2026 they reported a staged extraction framework that produced a traceable stepwise process and a continuous score whose discrimination was numerically similar to a CNN baseline.

Read the study abstract

Background Accurate localization of premature ventricular contraction (PVC) origin from 12-lead electrocardiography (ECG) is important for procedural planning in catheter ablation. Although convolutional neural network (CNN)-based models have shown promising diagnostic performance, they require task-specific training and remain limited in interpretability. We evaluated whether large language model (LLM)-based ECG image interpretation could perform binary left-versus-right PVC origin localization from 12-lead ECG…

By the Truvace desk1 min read1 sourceEvidence: Moderate · 60/100 — Observational study · n = 157 · 1 independent sourcePeer-reviewedTRV-2026-0629· cite & verify
Large Language Model-Based Localization of Premature Ventricular Contraction Origins: A Retrospective Diagnostic Accuracy Study
Hospital Universitario Doctor Peset, Valencia 01 by 19Tarrestnom65. CC BY-SA 4.0 · https://creativecommons.org/licenses/by-sa/4.0

In brief

The finding matters because accurate pre-procedural localization can inform ablation planning, and a traceable LLM approach could address interpretability limits of task-specific CNNs. Uncertainty remains about generalizability, as the high-PPV threshold was set from training data and the authors noted it warrants prospective external validation.

Main points

  1. Retrospective study of 157 patients who underwent successful catheter ablation, classified as RIGHT-origin (n = 103) or LEFT-origin (n = 54) by final ablation site.
  2. Compared CNN baseline (PPV 0.546 b1 0.058, NPV 0.793 b1 0.047, AUC 0.595 to 0.730) to LLM one-shot and LLM staged extraction with deterministic rule-based integration.
  3. Staged extraction generated continuous rule-based score with AUC 0.720 b1 0.045 versus CNN 0.712 b1 0.054 across five independent seeds.

The gain

LLM-based staged extraction framework localized PVC origin as left versus right from 12-lead ECG images with discrimination comparable to a CNN baseline while providing a traceable stepwise diagnostic process.

The rundown

The authors retrospectively analyzed 12-lead ECG images from 157 patients with successful ablation, labeling each case by ablation site. They evaluated performance using PPV, NPV, recall, and PPV + NPV across five seeds, comparing a CNN baseline to LLM approaches.

The staged extraction approach combined LLM-based feature extraction with deterministic rule-based integration to produce a continuous score and a stricter threshold that increased PPV at the expense of recall, as reported in the results.

Sources

  1. Peer-reviewedJournal of Cardiovascular Electrophysiology2026-08-01

The debate