Postoperative delirium (POD) is a complication affecting up to 50% of older surgical patients. Early identification of at-risk patients is critical for targeted prevention.

In this prospective cohort study, we validated the performance of the PIPRA machine-learning model for predicting POD across four hospitals in Switzerland and Germany. POD incidence (defined as a 4AT score of ≥4 points or an ICDSC score of ≥4) was assessed twice daily for 5 days in 1,179 consecutive patients aged 60 years or older undergoing noncardiac, non-intracranial surgery. Risk of POD was estimated using the PIPRA algorithm. A prespecified subgroup analysis was conducted in patients receiving a structured delirium prevention program, and predictive performance was assessed using discrimination (AUC), calibration, and the Brier score.

174 patients (14.8%) developed POD. PIPRA discriminated with an AUC of 0.73 (95% CI: 0.69–0.77), compared to the AUC of the original development dataset (0.80 (95% CI: 0.77–0.82)). In patients receiving structured POD prevention, PIPRA discriminated with an AUC of 0.64 (95% CI: 0.55–0.73), and in patients with no prevention, it discriminated with an AUC of 0.76 (95% CI: 0.71–0.80).

In this large, multicenter cohort, the PIPRA model demonstrated acceptable, albeit not perfect, discrimination for predicting POD in older surgical patients and was able to stratify risk in current perioperative practice. However, the imperfect calibration, particularly at higher risks, limits the direct use of absolute risk estimates. Local recalibration and dedicated implementation studies are required before considering its general implementation to guide targeted delirium prevention.