Predict Postoperative Length of Stay for Isolated Coronary Artery Bypass Graft Patients Using Machine Learning

Abstract

Background and Objective: Predictive Analytics (PA) is new trending approach in the field of healthcare delivery. It focusses on patient centeredness, proactive care, quality improvement and advancing medical knowledge. PA uses data mining approach and build the prediction model through the development of Machine Learning algorithms. Coronary Artery Bypass Grafting (CABG) is an open-heart surgery and common procedure for Coronary Heart Disease treatment. Due to the invasive nature of the procedure it is a good practice to anticipate the Postoperative Length of Stay (PLoS) for better patients’ preparation and recovery management. The aim of this study is to develop and evaluate a prediction model to predict PLoS for iCABG patients, using Machine Learning techniques, and to identify the predictors with highest contribution to the model. Methods: This is a Retrospective study which use historic data of adult patients underwent isolated CABG from Saud Albabtain Cardiac Centre, Dammam. After data pre-processing and data imputation using kNN method, five prediction models were build using the algorithms of Naïve Bayes, Decision Tree, Random Forest, Logistic Regression and k Nearest Neighbour. Data imbalance was managed using stratified 10-k cross validation, oversampling, undersampling, Both and ROSE. Models were evaluated using cross validation and Hold-Out technique and the AUC was used to compare models’ performance. Finally, variable selection process was conducted using Information Gain and Boruta method. Models were built using Orange and R statistical software. Results: Dataset contains 35 distinct attributes and 621 instances were used to develop the models. Total of 38 models were developed using both software, and Random Forest model with “Both” sampling method produced (AUC=0.80) and F1 score (0.77) and recall (0.75), and it was selected as the best fit model. Attributes found to be highly correlated with the PLoS are age, PA systolic pressure, complication, and intra-aortic balloon pump. Conclusion: This study demonstrates the ability to predict PLoS for iCABG patients using patient specifications and pre/intra-operative measures. In addition it proved the high predictability of EuroSCORE II and age in PLoS. Further studies are needed to develop a model that approximate PLoS in form of continuous data.