AI-Assisted Prediction of Postoperative Delirium: External Validation of the Pre-Interventional Predictive Risk...

Postoperative delirium is a prevalent and serious complication in older surgical patients, linked to prolonged hospitalization, higher morbidity, and long-term cognitive decline.

To externally validate the Pre-Interventional Predictive Risk Assessment (PIPRA) model in a Japanese postoperative ICU cohort and examine the transportability of a preoperative delirium prediction model in a high-acuity clinical setting. Artificial intelligence-based prediction models have advanced perioperative risk assessment, yet the utility of the Pre-Interventional Predictive Risk Assessment model in critically ill populations remains unclear.

This retrospective study examined the external performance of Pre-Interventional Predictive Risk Assessment in a Japanese postoperative intensive care unit. Adult patients admitted after noncardiac, nonneurosurgical operations were screened for delirium using the Intensive Care Delirium Screening Checklist. The original Pre-Interventional Predictive Risk Assessment executable was applied without any recalibration or modification.

Among 146 eligible patients, 38 (26%) developed postoperative delirium. Pre-Interventional Predictive Risk Assessment yielded an area under the receiver operating characteristic curve of 0.60 (95% confidence interval, 0.49-0.71), indicating limited discriminative ability. Calibration analysis suggested reasonable alignment at lower predicted risks, with risk overestimation observed among higher-risk individuals. Patients with delirium had longer intensive care unit and overall hospital stays compared with those without postoperative delirium. Sensitivity analyses yielded area under the receiver operating characteristic curves of 0.58 (95% confidence interval, 0.49-0.68) for Intensive Care Delirium Screening Checklist ≥3 and 0.67 (95% confidence interval, 0.55-0.79) for Intensive Care Delirium Screening Checklist ≥5, indicating that variation in outcome threshold did not substantially alter overall interpretation.

In this postoperative intensive care unit cohort, direct application of a preoperative delirium prediction model showed limited transportability without recalibration. These findings highlight challenges in applying static risk models across heterogeneous clinical settings and support the need for context-specific validation and incorporation of intensive care unit-related predictors before clinical implementation.