Using multiomic data to predict postoperative complications after major surgery in the UK Biobank cohort

IntroductionPostoperative complications after major surgery have substantial impacts on morbidity and resource utilisation. We investigated whether adding high-dimensional metabolomic and proteomic data to standard clinical variables would improve the prediction of a range of postoperative complications. MethodsWe analysed data from UK Biobank, a large prospective cohort study. Participants who underwent major surgery and had metabolomic and/or proteomic data were included. The primary outcomes were postoperative atrial fibrillation, acute kidney injury, acute myocardial infarction, delirium, stroke and surgical site infection. We trained machine learning models (elastic net penalised regression) with a range of feature sets to predict these outcomes. For outcomes where sample sizes were below recommended levels for predictive modelling, we employed transfer learning from the non-postoperative domain. We compared the predictive performance (AUROC, sensitivity, specificity) of models using only baseline clinical variables with those integrating single- and multiomic datasets. ResultsThe dataset included 158,156 individuals undergoing qualifying surgery. The numbers of cases with omic data varied across outcome phenotypes and feature sets: metabolomic: 144-1596, proteomic: 27-289 and multiomic: 15-219. Baseline clinical models achieved robust predictive performance (AUROC 0.72-0.88, sensitivity 0.71-0.80). The addition of metabolomic and/or proteomic features, using a variety of integration approaches, provided no clinically meaningful improvement in performance across any of the clinical phenotypes. Transfer learning from the non-postoperative domain improved model performance and stability but did not outperform baseline clinical models. ConclusionsThe addition of metabolomic and/or proteomic data from samples collected at a temporal distance from surgery does not improve pre-operative risk prediction compared to standard clinical variables. The lack of incremental predictive value likely reflects the extended gap between biobank sampling and the surgical event. The success of transfer learning from non-postoperative settings suggests shared biological risk between chronic and acute phenotypes.