Determining latent features and forecasting of COVID-19 hospitalisations in Malaysia using a national patient assessment data platform: a study of machine learning modelling against expert system

COVID-19 had a severe impact on Malaysia, as cases increased dramatically as the pandemic spread. In order to combat the pandemic, the Ministry of Health has established a number of standard operating procedures (SOP) and started operating COVID-19 Assessment Centers (CAC). This study compares the expert system created using the current patient evaluation standards to the capabilities of machine learning approaches in capturing the potential of being admitted directly or during home quarantine, based on the different clinical symptoms and age group. Boruta is a feature selection method that is employed to rank and extract significant characteristics. Treatment for imbalance has been carried out by under-sampling with K-Means and over-sampling with SMOTE. It appeared that the machine learning method using Random Forest would perform better than the expert systems. There are five performance metrics used in this study, i.e. accuracy, precision, recall, F1-score, and specificity. This study focused to maximize the true positive rate while minimize the false negative rates, it is to make sure that the patient who really need to be hospitalized will not be missed out. Therefore, recall becomes the main evaluation metrics when comparing the machine learning model and the expert system. The results shown that the recall score for machine learning approach is vastly higher then of expert systems. For age group 18-59, machine learning has 32.75% recall more than the expert system to predict if a patient requires direct admission, while for age group more than 60, the recall of machine learning is 18.11% more than expert system. In addition, to predict if a patient require admission during their home quarantine due to their health deterioration, machine learning recorded 76.72% recall more than the expert system for patient aged 18 to 59, and 70.59% difference for patient more than 60 years old. This supports the potential application of machine learning for clinical decision making for COVID-19 patients.