Theses (Master In Health Informatics & Analytics)

Predicting hospital admission in the Emergency Department (ED) using essential triage information is vital to reduce the current steps of the work processes and the chances of missing values. We aimed to predict hospital admission using supervised machine learning methods of triage information from Medical Information Mart for Intensive Care IV ED (MIMIC-IV-ED). We analyzed the MIMIC-IV-ED database and selected basic demography and vital signs. Two datasets were created after replacing string values with numerical values. Subsequently, feature selection techniques were applied to identify the essential variables. The cleaned data were fit into three machine learning algorithms: (i) Logistic Regression, (ii) Decision Tree, and (iii) Random Forest. The performance was compared by using the area under the operating curve (AUROC). We included 420,848 ED data, of which 26,864 (6%) were excluded. The hospital admission risk was 39%. Logistic Regression and Random Forest without feature selection perform the best in both hot-encoded (AUROC 0.763 and 0.770, respectively) and label encoded datasets (AUROC 0.761 and 0.771, respectively). Although Decision Tree performs better with feature selection, its best result is still inferior to Logistic Regression and Random Forest. We developed a prediction modeling using essential triage information routinely collected at the triage. Although the result is promising, additional demography, such as age, pain score, and chief complaint would likely yield a better result. Implementing such minimalist information into the Electronic Health Record (EHR) could add extra value by giving healthcare providers information about the upcoming overcrowding issue.