DOI: 10.12688/wellcomeopenres.26850.1 ISSN: 2398-502X

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

Richard A. Armstrong, Paul Yousefi, Ben Gibbison, Golam M Khandaker, Tom R Gaunt
Show PDF Cite
Background Whilst high-dimensional omic profiles from biobanks can predict long-term disease outcomes, their ability to predict postoperative outcomes is unknown. We investigated whether adding metabolomic and proteomic data from historic samples to standard clinical variables improved the prediction of postoperative complications. Methods We analysed data from 158,156 UK Biobank participants who underwent major surgery. The primary outcomes were postoperative atrial fibrillation, acute kidney injury, acute myocardial infarction, delirium, stroke and surgical site infection. We developed machine learning models (elastic net penalised regression) to compare baseline clinical variables against integrated single- and multiomic datasets. To address sample size constraints in high-dimensional omic subsets, we employed transfer learning from related non-postoperative domains. Results 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 provided no clinically meaningful improvement in discriminative performance metrics for any clinical phenotypes. Transfer learning from the non-postoperative domain improved model performance and stability but did not outperform baseline clinical models. Conclusions The addition of metabolomic and proteomic data from samples collected at a temporal distance from surgery does not improve preoperative risk prediction compared to standard clinical variables. The lack of incremental predictive value suggests that long-term biological trajectories have become manifest in routine clinical data by the time of surgery. The success of transfer learning from non-postoperative settings may indicate shared biological risk between chronic and acute phenotypes.

More from our Archive