Using machine learning and sequential characteristics in prediction of cardiac arrest: a case study of Uganda Heart Institute, Central Uganda

Abstract

Background: Research on cardiac arrest and its associated risk factors has considerably increased in recent decades. With that, several machine learning models have been implemented to predict the onset of cardiac arrest from available patient data in various settings. However, local machine learning studies in the prediction of cardiac arrest in Uganda have barely been undertaken to enable health workers to identify those individuals likely to suffer cardiac arrest and implement corrective measures before the potential likelihood of fatality. We, therefore, sought to identify a novel approach to the prediction of cardiac arrest in adult patients at a Uganda Heart Institute. Methods: Data from patient’s files at the Uganda Heart Institute spanning from January 1, 2012, to December 31, 2021, was collected for a retrospective study. This included information from adult inpatient and out patient’s records regarding occurrences of cardiac arrest. Multivariate clinical features such as diastolic blood pressure and systolic blood pressure were extracted from patient files and monitoring charts, as well as post-mortem findings. Three machine learning algorithms were employed to develop predictive models for cardiac arrest, with the best-performing model selected. Data compilation and cleaning were done using MS Excel and Python software, respectively. Model performance was assessed using confusion matrix output and area under the curve (AUC) metrics for comparison. Results: The study established that the Random Forest model was the most effective in predicting the likelihood of cardiac arrest. The random forest algorithm (AUC 0.90) outperformed the Support Vector Machine (AUC 0.84) and the logistic regression algorithm (AUC 0.72). Conclusions: This study demonstrates the potential of machine learning models in accurately predicting the likelihood of cardiac arrest. Hence demonstrating their potential to enhance patient care by identifying individuals at higher risk and aiding in implementation of proactive interventions.