692 A Machine Learning Approach to Predict the Postoperative Length of Stay After Coronary Artery Bypass Grafting Using Preoperative Characteristics. (12th October 2021)

• Record Type: Journal Article
• Authors: Bruno, V D; Guida, G; Jones, C; Bates, M; Di Tommaso, E; Rajakaruna, C
• Is Part Of: British journal of surgery. Volume 108:Supplement 6(2021)
• Publication Date: 2021-10-12
• Subjects: Surgery -- Periodicals; 617.005
• DOI: 10.1093/bjs/znab258.048
• Languages: English
• ISSNs: 0007-1323

Abstract

Abstract: Aim: Lengthy hospital length of stay (LOS) has a direct impact on healthcare costs. We aimed to design predictive models of prolonged LOS after coronary artery bypass grafting (CABG) with only preoperative characteristics and machine learning (ML) strategies. Method: In a single centre retrospective analysis, 2, 082 consecutive patients underwent first-time elective/urgent CABG: 1, 262 has a short postoperative LOS (≤ 6 days) while the remaining 820 had a long LOS (> 6 days). 70/30 training/testing ratio and resampling methods were used, and cross-validation was conducted. Results: The two groups differ significantly in terms of pre-operative variables: short LOS patients were younger (p < 0.01), more frequently male (p < 0.01) with lower BMI (p < 0.01) and better angina class (p < 0.01) and NYHA class (p < 0.01). Moreover, they had lower incidence of hypertension (p = 0.04), COPD (p < 0.01) and PVD (p < 0.01). The Logistic Euroscore was also better in this group (median 0.02 vs 0.03, p < 0.01). The predictive abilities of the ML models were as follows: logistic regression: Area under the Curve (AUC) = 0.71, accuracy = 0.69; Generalized additive model: AUC= 0.7, Accuracy = 0.68; Random Decision Forest: AUC = 0.7, Accuracy = 0.68; Naïve Bayes AUC = 0.63, Accuracy 0.58. Conclusions: Developing a reliable predictive model with only pre-operative variables proved to be difficult, but several preoperative characteristics have a significant impact on the probability of prolonged LOS after CABG. Larger studies are needed to investigate the possibility of developing a reliable predictive model that would help to improve surgical planning.