Development of a machine learning model to predict intensive care unit bed demand for adult elective surgical patients at a large United Kingdom National Health Service Trust

Background Elective surgical admissions form a growing share of demand for ICU beds, a constrained resource. Capacity planning for these admissions is feasible, but hospitals often lack reliable systems estimating daily elective surgical ICU bed demand before the day of surgery. Comprehensive clinical review of all elective cases is impractical, so planning relies on subjective preassessment processes of variable reliability. This study aimed to develop a machine learning model predicting elective surgical ICU bed demand using electronic health record data to improve on current electronic bed demand estimation at a large UK National Health Service (NHS) Trust. Methods Using a retrospective dataset comprising 38 656 elective inpatient surgeries occurring at three sites in a large UK NHS trust between 1 May 2019 and 31 December 2023, we developed two tree-based machine learning models predicting ICU admission after elective surgery: one using only basic, objective clinical data (CoreML) and one using additional preassessment data (FullML). Individual predictions were aggregated to forecast ICU bed demand. Performance was validated retrospectively and prospectively. Results At our large UK NHS Trust, in a prospective evaluation, only 71.6% of elective surgical cases admitted to ICU after surgery had an ICU bed electronically requested. In this evaluation, the CoreML model predicting ICU admission at an individual level 1 day before surgery achieved an area under the receiver operator curve of 0.88. It outperformed the current electronic indicator of aggregate elective surgical ICU bed demand 1 day before surgery at two sites handling 72% of inpatient elective surgery (root mean square error, 1.28 vs 1.64 at site A; 0.76 vs 1.16 at site C). CoreML outperformed FullML in aggregate prediction at all sites in prospective evaluation; however, importantly in retrospective evaluation, the converse was true. Conclusions We demonstrate that aggregating individual-level ICU admission predictions for elective surgeries provides a bed demand estimate that improves on the current electronic bed demand indicator 1 day before surgery at two out of three sites conducting the majority of inpatient elective surgery at our large UK NHS Trust. We demonstrate the importance of prospective validation, in which the more parsimonious model was the best performing.