Translating predictive models into clinical practice: Fast-and-frugal trees for postoperative delirium using routine data

Postoperative delirium (POD) is a frequent and serious complication associated with increased morbidity and mortality. Although predictive models for POD exist, they are rarely implemented in routine clinical practice, partly due to their complexity and data requirements. Simple and interpretable tools based on routinely available data may improve clinical applicability. In this retrospective cohort study, we evaluated Fast-and-Frugal Trees (FFTrees) for predicting POD using routine electronic health record data from 61,150 surgical patients treated between 2017 and 2020 at a large German university hospital. We assessed the transferability of previously developed FFTrees and compared their performance with logistic regression, classification and regression trees, random forests, and support vector machines. Updated FFTrees were derived for pre- and peri-operative prediction. Previously published FFTrees showed reduced performance in routine data but improved after recalibration. Updated FFTrees achieved balanced accuracies of 58% (pre-operative) and 60% (peri-operative), with performance comparable to more complex models. Although predictive accuracy was moderate across all models, FFTrees offer advantages in simplicity, transparency, and ease of implementation. As easily memorized decision rules, they can help clinicians organize familiar cues in an optimal order, supporting delirium risk assessment in routine practice where more complex tools are rarely used.