Development and internal validation of a gradient-boosted trees model for prediction of delirium after surgery and anesthesia (the BioCog study)

IMPORTANCEPostoperative delirium (POD) is a multietiological condition and affects 20% of older surgical patients. It is associated with poor clinical outcome and increased mortality. OBJECTIVEWe aimed to develop and validate a risk prediction algorithm for POD based on a multimodal biomarker database exploiting preoperative data (predisposing factors) and procedural factors as well as perioperative molecular changes associated with POD (precipitating factors). DESIGNBioCog is a prospective cohort study conducted from November 2014 to April 2017. Patients were followed up for seven postoperative days after surgery for POD. Gradient-boosted trees (GBT) with nested cross-validation was used for POD prediction. SETTINGPatients aged [≥]65 years were enrolled at the anesthesiologic departments of two tertiary care centers. EXPOSUREAll patients underwent surgery with an expected duration of at least 60min. Clinical, neuropsychological, neuroimaging data and blood were collected and clinically well established as well as non-established biomarkers (e.g., gene expression profiling) were measured pre- and postoperatively. MAIN OUTCOMEPOD according to DSM 5 until the seventh postoperative day RESULTS184 of 929 (20%) patients experienced POD. A GBT algorithm using both preoperative data, characteristics of the intervention and postoperative changes in laboratory parameters achieved the highest area under the curve (0.83, [0.79; 0.86]) with a Brier score of 0.12 (0.12; 0.13). CONCLUSIONS AND RELEVANCEModels combining predisposing factors with precipitating factors predict POD best. Non-routine laboratory data provide useful information for POD risk prediction, providing relevant results for future studies on the molecular factors of POD. In addition, possibly relevant molecular mechanisms contributing to the development of POD were identified, mostly indicating a dysregulated postoperative immune response. This study constitutes the basis for future hypothesis-driven analyses or implementation of prediction expert system for clinical practice.