Predicting Readmissions or Post-Discharge Mortality After Cardiac Surgery with Machine Learning Using an Australian Database.

AIM: This study aimed to create machine learning algorithms using the Australian & New Zealand Society of Cardiac & Thoracic Surgeons (ANZSCTS) Database that can predict readmissions or post-discharge mortality within 30 days of cardiac surgery. METHOD: Data from 54 public and private Australian hospitals from January 2017 to December 2021 were included in this study. Only coronary artery bypass grafting (CABG), valvular heart surgery, or aortic surgery cases were included for analysis. The primary outcome of this study was the performance of our machine learning models measured using sensitivity, specificity, positive and negative predictive values, accuracy, area under the receiver operating characteristic curve (AUROC), and area under the precision-recall curve. RESULTS: Of the 178,252 patients in the database, 61,721 CABG, valvular heart surgery, and aortic surgery cases were identified. The incidence of 30-day readmission or post-discharge mortality was 10.3% (6,351/61,721). The best-performing model was the deep neural network (AUROC, 0.615). A calibrated version of this model predicted one in five patients at higher risk of readmissions or post-discharge mortality compared with the lowest risk group (2.4× increase in incidence rate and an odds ratio of 2.7 [2.2-3.2]). CONCLUSIONS: Using machine learning, models developed from the Australian ANZSCTS Database demonstrated limited predictive ability for readmissions or post-discharge mortality. The ANZSCTS Database may require additional informative variables before models become sufficiently accurate. Individual hospitals or small hospital networks may be best placed to incorporate routinely collected information (such as electronic medical records) alongside ANZSCTS data to create more accurate models.