• Paulo C S Neto, Attila L Rodrigues, Adriene Stahlschmidt, Lucas Helal, and Luciana C Stefani.
• From the Programa de Pós-graduação em Medicina: Ciências Médicas, Universidade Federal do Rio Grande do Sul (PCSN), Universidade Federal do Rio Grande do Sul (ALR), Programa de Pós-graduação em Medicina: Ciências Médicas, Universidade Federal do Rio Grande do Sul (AS), Hospital de Clínicas de Porto Alegre and Universidade Federal do Rio Grande do Sul (LH), Programa de Pós-graduação em Medicina: Ciências Médicas, Professor at Surgical Department -Universidade Federal do Rio Grande do Sul and Chief of Teaching Division of Hospital de Clínicas de Porto Alegre, Porto Alegre, Brazil (LCS).
• Eur J Anaesthesiol. 2023 May 1; 40 (5): 356364356-364.

BackgroundPostoperative delirium (POD) has a negative impact on prognosis, length of stay and the burden of care. Although its prediction and identification may improve postoperative care, this need is largely unmet in the Brazilian public health system.ObjectiveTo develop and validate a machine-learning prediction model and estimate the incidence of delirium. We hypothesised that an ensemble machine-learning prediction model that incorporates predisposing and precipitating features could accurately predict POD.DesignA secondary analysis nested in a cohort of high-risk surgical patients.SettingAn 800-bed, quaternary university-affiliated teaching hospital in Southern Brazil. We included patients operated on from September 2015 to February 2020.PatientsWe recruited 1453 inpatients with an all-cause postoperative 30-day mortality risk greater than 5% assessed preoperatively by the ExCare Model.Main Outcome MeasureThe incidence of POD classified by the Confusion Assessment Method, up to 7 days postoperatively. Predictive model performance with different feature scenarios were compared with the area under the receiver operating characteristic curve.ResultsThe cumulative incidence of delirium was 117, giving an absolute risk of 8.05/100 patients. We developed multiple machine-learning nested cross-validated ensemble models. We selected features through partial dependence plot analysis and theoretical framework. We treated the class imbalance with undersampling. Different feature scenarios included: 52 preoperative, 60 postoperative and only three features (age, preoperative length of stay and the number of postoperative complications). The mean areas (95% confidence interval) under the curve ranged from 0.61 (0.59 to 0.63) to 0.74 (0.73 to 0.75).ConclusionA predictive model composed of three indicative readily available features performed better than those with numerous perioperative features, pointing to its feasibility as a prognostic tool for POD. Further research is required to test the generalisability of this model.Trial RegistrationInstitutional Review Board Registration number 04448018.8.0000.5327 (Brazilian CEP/CONEP System, available in https://plataformabrasil.saude.gov.br/ ).Copyright © 2023 European Society of Anaesthesiology and Intensive Care. Unauthorized reproduction of this article is prohibited.

29 keywords...

hide…