Clinical prediction models for postoperative delirium: from GLMs to Neural Networks

Izdar Abulizi, Philipps University Marburg

Co-authors: Jan Menzenbach, University Hospital Bonn; Maria Wittmann, University Hospital Bonn; Andreas Mayr, Philipps University Marburg

Abstract: Prediction modelling for patient-individual health outcomes requires careful consideration of interpretability and prediction accuracy. When discussing modelling choices, classical statistical models with structured, interpretable frameworks like GLMs, compete nowadays with more flexible unstructured models from the machine learning domain that mostly focus on prediction accuracy (e.g., tree-based approaches or neural networks). We tackle the question of what can be gained in predictive performance from increased model complexity when giving less focus on interpretability. We discuss these methodological questions by showcasing a highly relevant clinical application of predicting postoperative delirium from routine preoperative data. For this particular research question, advanced statistical models clearly outperformed the initial investigator model with expert-selected variables. However, penalized logistic regression models performed as well or even better than unstructured machine-learning approaches.