Artificial Intelligence Evaluation of Fascial Plane Block Quality and Postoperative Pain in Cardiac Surgery

Patients scheduled for elective open heart surgery via median sternotomy were enrolled after obtaining written informed consent, including separate consent for AI-based image analysis. All patients underwent routine anesthetic management including BIS monitoring, intra-arterial cannulation, standard anesthesia induction, orotracheal intubation, and central venous catheterization.

Fascial Plane Block Application:

Following general anesthesia induction, fascial plane blocks routinely performed in our clinic were applied by a blinded anesthesiologist using a high-frequency linear ultrasound probe (10-15 MHz). The blocks applied included SAPB + PIFB combination or bilateral parasternal (PIFB) block, depending on the surgical decision. All blocks were performed using 0.25% bupivacaine (total 30-40 mL bilaterally) under sterile conditions with an in-plane technique.

Image Recording Protocol:

Block images were recorded in a standardized manner including: pre-block anatomical scanning, video recording during needle placement, local anesthetic spread images, and post-injection final images. All images were recorded in DICOM format (minimum 1920x1080 resolution) with standardized depth, gain, and focus settings. Patient identifiers were anonymized. From each recording, a standardized series of approximately 8-10 consecutive representative frames was compiled and presented identically to all assessors.

AI Assessment:

In accordance with the current approved protocol (dated 29 April 2026), five pre-trained, publicly available large vision-language models (Gemini 3.5 Flash, Claude Sonnet 4.6, ChatGPT 5.5, Grok 4.3, Kimi AI 2.6) were included as independent observers. Models were used as-is, without fine-tuning, retraining or any task-specific adaptation. Images were presented via standardised medical instructions (zero-shot prompting): for each patient a standardised series of approximately 8-10 consecutive anonymised frames was submitted in a new, independent session under an identical prompt, and each model returned a single 3-point classification.

The originally registered description (a bespoke convolutional neural network of U-Net/ResNet architecture, trained on expert-labelled reference images with 70/20/10 partitioning, evaluated by landmark-recognition accuracy, Dice coefficient and a 0-100 similarity score) was transcribed from a superseded version of the protocol. No bespoke model was developed, trained, or evaluated at any stage of the study, and none of those metrics were computed.

Block Quality Scoring:

Recorded ultrasound images were scored independently and blindly by the five AI models, by three anaesthesiologists experienced in fascial plane blocks, and by an adjudicator, using a 3-point scale:

• Anatomically successful / optimal block
• Anatomically patchy / inadequate block
• Anatomically incorrect block All assessors were blinded to one another, to the AI outputs, and to clinical outcomes. Note: the approved protocol defines the scale in the reverse direction (1 = incorrect, 3 = successful). As conducted, the inverted coding above was applied uniformly to all assessors and to the analysis dataset; agreement analyses are unaffected by the direction of the labels.

Postoperative Follow-up:

NRS pain scores (at rest and with coughing) were assessed at 0, 6, 12, 24 and 48 hours. Assessment at 0 and 6 hours was not evaluable because all patients remained intubated and sedated at these time points; analyses use the 12, 24 and 48 hour assessments. Additional parameters recorded included total analgesic consumption, side effects (nausea, vomiting, pruritus), complications, mobilization time, ICU stay, and hospital length of stay.

Statistical Analysis:

Correlation between block success scores and postoperative pain scores was assessed using Spearman correlation. Factors affecting pain scores were evaluated by multiple linear regression. Agreement between assessors was quantified with linear weighted Cohen's kappa and Fleiss' kappa. Diagnostic-accuracy analyses against the adjudicator were not prespecified and are exploratory.

Sample Size:

The initially approved protocol specified 80 patients for 80% power at a 5% significance level, based on pilot data. An amendment petition dated 29 April 2026 requested an increase to 120 patients, stating that this raises statistical power for the primary correlation analysis from 80% to above 90% and secures sufficient analysable cases against image loss. The Ethics Committee decision (E-31234050-514.99-313724142, 6 May 2026) approved the increase in the approved sample size to 120. The figure of "95% power" in the original registration entry did not correspond to any protocol document and has been corrected. 120 patients were enrolled; 115 received a fascial plane block and were analysed.

Corrections and deviations:

This record was first submitted on 6 May 2026, after completion of data collection; registration was therefore retrospective. Prospective registration is not mandated for observational studies under ICMJE criteria. The protocol was approved by the institutional Ethics Committee on 13 August 2025 (2024-TBEK 2025/08-14), before the first patient was enrolled, and the design was prospective.

• Registration error. The AI methodology in the original registration entry was transcribed from a superseded version of the protocol. The current approved protocol, dated 29 April 2026 and submitted to the Ethics Committee with the amendment petition, specifies pre-trained large vision-language models used as independent observers via zero-shot prompting. The study as conducted followed that protocol. The Ethics Committee decision of 6 May 2026 refers in its own text to the evaluation of AI models (ChatGPT, Gemini). This entry has been corrected accordingly.
• Timing of the amendment. The amendment petition was submitted on 29 April 2026 and the Ethics Committee decision is dated 6 May 2026; enrolment had by then been completed, so the increase from 80 to 120 approved participants was granted after the enrolment period had closed. The covering letter for the petition characterised the change as affecting sample size only; the amended protocol attached to it describes the change in methodology in full. All participants gave written informed consent, including separate consent for AI-based analysis of anonymised images, and all blocks were part of routine care.
• Assessors and reference standard. The approved protocol designates expert anaesthesiologist assessment as the gold standard. As conducted, three blinded experts scored the images independently and a fourth, more senior anaesthesiologist served as adjudicator and reference standard. Diagnostic-accuracy analyses against the adjudicator were not prespecified and are exploratory.
• Scale direction. The approved protocol defines the 3-point scale as 1 = incorrect to 3 = successful. As conducted, the direction was inverted and applied uniformly to all assessors and to the analysis dataset.
• Pain assessment. NRS at 0 and 6 hours was not evaluable as all patients remained intubated and sedated.
• Sample size. The "95% power" figure in the original entry did not correspond to any protocol document and has been corrected.

Inclusion criteria

• Age between 18-80 years
• Patients scheduled for elective cardiac surgery (CABG, valve surgery, combined procedures)
• Patients undergoing median sternotomy
• ASA physical status II-III
• Patients capable of providing informed consent, including separate consent for AI-based image analysis

Exclusion criteria

• Emergency surgery
• Redo sternotomy
• Allergy to local anesthetics
• Coagulopathy (INR >1.5, platelet count <80,000/mm³)
• Infection at the block application site
• Chronic opioid use (>3 months)
• Cognitive dysfunction
• Pregnancy
• ASA physical status IV and above
• Minimally invasive surgery performed via thoracotomy