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

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.

Sensitivity was the percentage of image sets classified as successful by a reader among image sets classified as successful by the pragmatic single-expert reference. It was calculated separately for each of the five artificial-intelligence configurations and three expert anaesthesiologists in the 115 PIFB-treated participants. Possible values range from 0 to 100 percent; higher values indicate closer agreement with the reference.

Specificity was the percentage of image sets classified as not successful by a reader among image sets classified as not successful by the pragmatic single-expert reference. Patchy/inadequate and failed/no local anaesthetic were grouped as not successful. It was calculated separately for each of the five artificial-intelligence configurations and three expert anaesthesiologists in the 115 PIFB-treated participants. Possible values range from 0 to 100 percent; higher values indicate closer agreement with the reference.

Positive predictive value was the percentage of image sets classified as successful by a reader that were also classified as successful by the pragmatic single-expert reference. It was calculated separately for each of the five artificial-intelligence configurations and three expert anaesthesiologists in the 115 PIFB-treated participants. Possible values range from 0 to 100 percent; higher values indicate closer agreement with the reference.

Negative predictive value was the percentage of image sets classified as not successful by a reader that were also classified as not successful by the pragmatic single-expert reference. Patchy/inadequate and failed/no local anaesthetic were grouped as not successful. It was calculated separately for each of the five artificial-intelligence configurations and three expert anaesthesiologists in the 115 PIFB-treated participants. Possible values range from 0 to 100 percent; higher values indicate closer agreement with the reference.

Overall accuracy was the percentage of the 115 PIFB-treated image sets for which a reader's binary successful or not-successful classification matched the pragmatic single-expert reference. It was calculated separately for each of the five artificial-intelligence configurations and three expert anaesthesiologists. Possible values range from 0 to 100 percent; higher values indicate closer agreement with the reference.

Across all 120 image sets, linear-weighted Cohen kappa was calculated separately for each of the five artificial-intelligence configurations and three non-reference expert anaesthesiologists against the pragmatic single-expert reference using the ordered categories successful, patchy/inadequate and failed/no local anaesthetic. Values range from -1 to 1; 1 indicates perfect agreement, 0 indicates agreement expected by chance and negative values indicate less agreement than expected by chance. Each reader assessed each image set once.

Across all 120 image sets, Fleiss kappa was calculated separately among the five artificial-intelligence configurations and among the three non-reference expert anaesthesiologists using the categories successful, patchy/inadequate and failed/no local anaesthetic. Values range from -1 to 1; 1 indicates perfect agreement, 0 indicates agreement expected by chance and negative values indicate less agreement than expected by chance. Each reader assessed each image set once; within-reader and within-system repeatability were not evaluated.

Resting pain was reported verbally as a whole number on an 11-point Numerical Rating Scale from 0 (no pain) to 10 (worst imaginable pain); higher scores indicate worse pain. Scores were summarised as median and interquartile range at each scheduled time point. Assessment required the participant to be extubated and able to respond. The planned 6-hour assessment was applicable only when the participant had already been extubated.

Pain during coughing was reported verbally as a whole number on an 11-point Numerical Rating Scale from 0 (no pain) to 10 (worst imaginable pain); higher scores indicate worse pain. Scores were summarised as median and interquartile range at each scheduled time point. Assessment required the participant to be extubated and able to respond. The planned 6-hour assessment was applicable only when the participant had already been extubated.

In the 115 participants who received PIFB, rescue paracetamol administration was recorded as a participant-level yes/no measure and was to be summarised as the number and percentage of participants receiving at least one rescue dose. Rescue treatment was administered when NRS exceeded 4 at the attending physician's discretion.

In the 115 participants who received PIFB, rescue tramadol administration was recorded as a participant-level yes/no measure and was to be summarised as the number and percentage of participants receiving at least one rescue dose. Rescue treatment was administered when NRS exceeded 4 at the attending physician's discretion.

In the 115 participants who received PIFB, rescue pethidine administration was recorded as a participant-level yes/no measure and was to be summarised as the number and percentage of participants receiving at least one rescue dose. Rescue treatment was administered when NRS exceeded 4 at the attending physician's discretion.

Intensive-care length of stay was the recorded number of days from postoperative intensive-care admission until intensive-care discharge and was summarised as median and interquartile range in the PIFB-treated cohort.

Hospital length of stay was the recorded number of days from coronary artery bypass grafting until discharge from the index hospital admission and was summarised as median and interquartile range in the PIFB-treated cohort.

In the 115 participants who received PIFB, postoperative complications recorded during the index hospital admission were summarised as the number and percentage of participants with at least one event.

In the 115 participants who received PIFB, mortality was summarised as the number and percentage of participants who died before discharge from the index hospital admission.