Five Multimodal AI Configurations Versus Anaesthesiologists for Pecto-Intercostal Block Classification in an Archived-Image Reader Study of a Prospective CABG Cohort

Background/Objectives: We compared five general-purpose multimodal AI configurations with anaesthesiologists classifying sonographic pecto-intercostal fascial plane block spread. Methods: This archived-image reader study used a prospective cohort of 120 ASA III adults undergoing isolated coronary artery bypass grafting at one centre. In June 2026, five free AI configurations and three experts classified identical sets of 8–10 left-sided frames. Direct binary comparisons used each expert’s labels in 115 block recipients; five no-block controls were separate. The senior expert reviewed expert labels before setting the final reference, precluding independent expert–AI accuracy comparison. Reference-conditioned balanced accuracy was defined as mean sensitivity and specificity. Results: Direct balanced agreement was 42.3–53.9% for four configurations and 55.4–64.9% for Kimi K2.6. AI asserted spread in 17/25 no-block classifications; all 20 human classifications indicated absence. Final-reference balanced accuracy was 76.9–91.9% for experts and 44.6–66.9% for AI, conditional on the expert-informed reference. Pain analyses were exploratory. Conclusions: Direct comparisons establish neither anatomical accuracy nor independent expert superiority. Limitations include one centre, operator and ultrasound system; selected frames; free access; no raw AI outputs, reproducible a priori power calculation or non-recruitment-day eligibility log.