A carregar...
Publicação
Multi-output machine learning for prediction of postoperative outcomes after cardiac surgery using patient blood management biomarkers
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 858.23 KB | Adobe PDF |
Orientador(es)
Resumo(s)
Background/Objectives: Postoperative complications following adult cardiac surgery are biologically interrelated, yet most machine learning models predict single outcomes. We developed an explainable multi-output model integrating routinely collected clinical variables and patient blood management (PBM) biomarkers to predict multiple postoperative outcomes simultaneously, with complementary mono-output analyses for selected endpoints. Methods: This retrospective single-center cohort included 1414 adults undergoing cardiac surgery. In total, 513 complete cases were analyzed. Thirteen outcomes were modeled, including major binary complications and ICU/ward length of stay. An initial 80:20 train–test split was used only for algorithm screening across six candidate multi-output regressors and training-set-defined feature subsets. The selected regressor was then evaluated across five random states, and global permutation feature importance was used for multi-output explainability. Mono-output binary analyses using the selected regressor and the same training-set-only feature-selection workflow were evaluated along with accuracy, precision, recall/sensitivity, and F1-scores. Results: The Decision Tree Regressor was selected. Across five random states, global multi-output performance was R2 = 0.83, MSE = 1.296, RMSE = 1.132, MAE = 0.298, and MAPE = 0.128. Based on global multi-output permutation importance, creatinine, ferritin, platelet count, estimated glomerular filtration rate, preoperative red blood cell units, and EuroSCORE II were ranked the highest. Atrial fibrillation had the lowest mono-output F1-score (0.719), whereas acute kidney injury, postoperative bleeding, infection, and 1-year hospital readmission yielded F1-scores of 0.928, 0.970, 0.963, and 0.975, respectively. Conclusions: This proof-of-concept study shows the feasibility of explainable multi-output modeling for postoperative outcomes after adult cardiac surgery using clinical and PBM variables. However, external validation is required prior to clinical use.
Descrição
Palavras-chave
Cardiac surgery Machine learning Multi-output regression Patient blood management Postoperative complications
Contexto Educativo
Citação
Editora
MDPI