Machine Learning-Based Pre-Diagnosis Tools in Emergency Departments: Predicting Hospitalization, Mortality and Triage Acuity

With demographic growth, aging populations, and the recent advent of the Covid-19 pandemic, emergency department staff are increasingly overwhelmed, leading to a decline in the quality of patient care and an alarming rise in hospital mortality rates. Thus, there is a growing need for accurate and reliable pre-diagnostic tools to optimize patient flow and improve the quality of care. While many artificial intelligence-based tools have been tested for pre-diagnosis in emergency services, they are still not effective enough. Moreover, exploring the use of additional indicators to enhance preliminary diagnosis should also be considered. This paper aims to predict the Emergency Severity Index-a standard reference for triaging patients-along with other indicators that clinicians can use to refine preliminary diagnoses, particularly regarding patient hospitalization and mortality. To achieve this, we utilized the MIMIC-IV (Medical Information Mart for Intensive Care) and MIMIC-IV ED (Emergency Department) datasets, applying several machine learning techniques. We propose a new pipeline for these prediction tasks, employing Polynomial Features for feature extraction and Principal Component Analysis for dimensionality reduction. Our proposed methods achieved high accuracy rates, with hospitalization prediction reaching 97.00%, mortality pre-diction at 86.41 %, and triage acuity prediction scoring of 99.80% using Logistic Regression, Random Forest, and Decision Tree classifiers.