Transcranial Doppler and Artificial Intelligence Prediction of Postoperative Delirium in Heart Valve Surgery

This is a single-center, prospective observational cohort study of 120 adults undergoing elective heart valve surgery under general anesthesia. All participants will receive the institution's standard anesthesia and analgesia protocol. No treatment will be assigned by the study.

Cerebral monitoring will include bispectral index (BIS), near-infrared spectroscopy (NIRS), and intermittent transcranial Doppler (TCD) measurements. TCD measurements will be obtained preoperatively at baseline (T0), at initiation of cardiopulmonary bypass (T1), during bypass (T2), during removal of the aortic cross-clamp (T3), and at separation from bypass (T4). Arterial blood gas parameters will be recorded at these time points and postoperatively, including a postoperative measurement and a measurement at 6 hours. The Standardized Mini-Mental State Examination will be administered preoperatively only to exclude participants with substantial baseline cognitive impairment; it will not be analyzed as an outcome.

The primary hypothesis is that a lower preoperative absolute mean middle cerebral artery blood flow velocity is associated with postoperative delirium. Exploratory analyses will assess whether intraoperative mean flow velocity normalized to baseline is associated with delirium in the direction of either hypoperfusion or hyperperfusion.

Postoperative delirium will be assessed daily on postoperative days 1 through 5 by an independent blinded evaluator using CAM-ICU in mechanically ventilated patients or the Turkish 3D-CAM in ward patients. Blinding will be maintained by separating the perioperative data and artificial intelligence prediction team, the delirium assessment team, and the statistician who merges the locked datasets for analysis.

Five general-purpose large language models will generate independent zero-shot delirium risk predictions for each participant using a standardized prompt and fixed recorded model versions. Predictions will be time-stamped and locked before the outcome data are merged. Model performance will be evaluated against the blinded reference standard using sensitivity, specificity, area under the receiver operating characteristic curve, and calibration. The primary clinical association will be analyzed using logistic regression, and model discrimination will be compared using the DeLong test.