LLM-Standardised CTA Impressions for PAD Revascularisation Planning: Blinded Single-Centre Agreement, Accuracy, and Efficiency Study

Abstract

Purpose: Lower-extremity CT angiography (CTA) reports are often narrative and slow to interpret for peripheral arterial disease (PAD) revascularisation planning. Large language models (LLMs) may standardise key anatomic descriptors. Our study is aimed to evaluate agreement, diagnostic accuracy, and decision-making efficiency of an LLM that converts narrative CTA reports into a predefined structured impression.

Material and Methods: In this retrospective, blinded, single-centre study (June–October 2025), consecutive adults undergoing CTA for PAD were included. Of 94 eligible patients, 86 (142 limbs) were analysed after exclusions for poor image quality (n=5) and incomplete reports (n=3). The LLM output a limb schema capturing ≥50% stenosis/occlusion, femoropopliteal lesion-length class (<5 cm, 5–15 cm, ≥15 cm), popliteal trifurcation involvement, and distal runoff (0–3). Comparator impressions were produced by an image-only expert reader. Reference standard was consensus of two vascular radiologists. Agreement used Cohen’s kappa (weighted for runoff). Timing per limb compared use of the LLM impression versus the full narrative report.

Results: Agreement between LLM and expert was near-perfect for ≥50% stenosis (κ=0.82) and occlusion (κ=0.88), and substantial for lesion-length class (κ=0.76), trifurcation involvement (κ=0.83), and distal runoff (weighted κ=0.74). Versus reference standard, LLM accuracy was 94% for ≥50% stenosis and 96% for occlusion. Median decision time decreased from 96 s (IQR 74–122) to 28 s (IQR 21–39) per limb (P<0.001).

Conclusion: An LLM can standardise CTA impressions for PAD planning, preserving diagnostic performance while improving efficiency by ~1 minute per limb.