Large Language Model-Based Localization of Premature Ventricular Contraction Origins: A Retrospective Diagnostic...

Area of Science:

• Cardiology
• Artificial Intelligence in Medicine
• Medical Imaging Analysis

Background:

• Accurate localization of premature ventricular contraction (PVC) origin from 12-lead electrocardiography (ECG) is crucial for effective catheter ablation planning.
• Current deep learning models like CNNs offer diagnostic potential but lack interpretability and require specific training.
• Investigating large language models (LLMs) for ECG image interpretation could provide a traceable diagnostic process for PVC origin localization.

Purpose of the Study:

• To evaluate the efficacy of LLM-based ECG image interpretation for binary left-versus-right PVC origin localization.
• To compare LLM approaches (one-shot and staged extraction) against a CNN-based model.
• To assess the interpretability and traceability of the LLM diagnostic process.

Main Methods:

• Retrospective study of 157 patients with successful PVC ablation, classifying ECGs as RIGHT-origin (103) or LEFT-origin (54).
• Comparison of three methods: CNN baseline, LLM one-shot, and LLM staged extraction with rule-based integration.
• Performance evaluation using PPV, NPV, recall, and AUC across five seeds.

Main Results:

• The CNN baseline model achieved PPV of 0.546 ± 0.058 and NPV of 0.793 ± 0.047 (AUC 0.595-0.730).
• The LLM staged extraction framework yielded comparable discrimination (AUC 0.720 ± 0.045) to the CNN model.
• A stricter threshold identified a potential high-PPV operating point, necessitating external validation.

Conclusions:

• LLM-based staged extraction shows potential for traceable, stepwise localization of PVC origins from 12-lead ECG images.
• The LLM approach demonstrated discrimination comparable to CNNs.
• Further prospective validation is required for the identified high-PPV operating point.