Five Multimodal AI Configurations Versus Expert Anaesthesiologists for Classifying Pecto-Intercostal Fascial Plane Block Spread in CABG

Bursa Yuksek Ihtisas Training and Research Hospital120 enrolled

Overview

This prospective observational cohort with a nested archived-image reader study compared the sonographic pecto-intercostal fascial plane block classifications of five general-purpose multimodal artificial-intelligence configurations and three expert anaesthesiologists against a separate senior anaesthesiologist's single-expert reference. Adults undergoing elective isolated coronary artery bypass grafting through median sternotomy were enrolled consecutively on days when the consultant anaesthesiologist responsible for block performance and ultrasound acquisition was available. Standardised left-sided ultrasound videos were recorded in 120 consented participants. Bilateral pecto-intercostal fascial plane block was performed in 115 participants; five participants who declined the block consented to standardised ultrasound imaging and served as no-block image controls. Eight to ten representative de-identified frames from each video were assessed independently by all readers. Associations between reference-defined sonographic spread and postoperative pain were secondary and observational.

This single-centre prospective observational cohort included 120 consenting adults undergoing elective isolated coronary artery bypass grafting through median sternotomy. At the end of surgery, after complete sternal and skin closure, 115 participants received bilateral ultrasound-guided pecto-intercostal fascial plane block with 20 mL of 0.25% bupivacaine on each side (40 mL and 100 mg in total). Five participants declined the block but consented to Ethics Committee-authorised standardised ultrasound imaging, storage and subsequent de-identified human/AI assessment; no needle was inserted and no local anaesthetic was administered in this group. Ultrasound videos were acquired from the left side with an M5 portable ultrasound system and a 7L4s linear-array transducer at a nominal frequency of 7.5 MHz using standardised settings. Eight to ten representative frames were selected from each video across the pre-block view, needle advancement, local-anaesthetic spread and final post-injection view, as applicable. The same frame set and fixed participant order were shown to every reader. After participant follow-up had ended, five general-purpose multimodal AI configurations, three expert anaesthesiologists and one separate senior reference anaesthesiologist independently classified each image set as successful, patchy/inadequate or failed/no local anaesthetic. The evaluated configurations were Gemini 3.5 Flash, Claude Sonnet 4.6, GPT-5.5 Instant in ChatGPT, Grok 4.3 and Kimi K2.6. They were accessed through free, non-subscription consumer accounts using their standard web interfaces and default response and data-use settings. For each system-participant pair, a new conversation was opened, the same fixed prompt and image set were submitted, and the first response was recorded without regeneration or re-prompting. The visible conversation was deleted before the next assessment, although this did not establish provider-side deletion. Each configuration assessed each participant once; within-system repeatability was not evaluated. The number and identities of the five consumer configurations were fixed before any reader output was observed; no bespoke AI model was developed or trained. The separate senior anaesthesiologist was a pragmatic single-expert reference and not an objective gold standard. The study was not designed as an equivalence or non-inferiority analysis. Postoperative Numerical Rating Scale pain at rest and during coughing was scheduled at 6, 12, 24 and 48 hours. The 6-hour assessment was applicable only to participants who had already been extubated at that time. Rescue analgesic administration during the first 48 hours, intensive-care and hospital length of stay, complications and mortality were also recorded. Pain associations were secondary, observational and non-causal. The ClinicalTrials.gov record was first submitted on 6 May 2026 and first posted on 12 May 2026, after the study had begun but while recruitment, image acquisition and clinical follow-up were still ongoing. Registration was therefore retrospective relative to study start. Image acquisition ended on 27 May 2026 and participant clinical follow-up ended on 1 June 2026.

Outcomes

Primary Outcomes

Balanced Accuracy of Binary PIFB Classification Against a Single-Expert Reference

Among 115 participants who received pecto-intercostal fascial plane block, binary classifications from each of five general-purpose multimodal artificial-intelligence configurations and three expert anaesthesiologists were compared separately with the classification of a senior anaesthesiologist serving as a pragmatic single-expert reference. Successful was the positive category; patchy/inadequate and failed/no local anaesthetic were combined as not successful. Balanced accuracy was calculated as (sensitivity + specificity)/2 and reported from 0 to 1, with higher values indicating closer agreement with the reference. The reference was not an objective gold standard, and no equivalence or non-inferiority hypothesis was tested.