Machine Learning Approach of Logistic Organ Dysfunction Score Prediction with Data Acquired from Bedside in ICU

Abstract

Intensive Care Unit (ICU) is a high-stakes environment in hospitals, where patients are at a higher risk than in other departments, such as organ dysfunction, and are monitored with devices, along with more care workers. Effectively predicting patient status in an ICU is a critical task serving patient health and resource allocation. Logistic Organ Dysfunction Score (LODS), calculated with weighted variables of the worst values in the first 24 hours, is an organ dysfunction scoring system that reflects the severity level of organ systems, and can be converted to the probability of mortality in a certain period. However, LODS calculation requires some laboratory results, such as bilirubin, which costs time and money. Effective prediction of LODS value could measure the patient’s overall organ dysfunction situation and calculate the probability of mortality for the patient, providing doctors with assistance in adjusting treatment. Machine learning can utilize large amounts of data and existing algorithms to train effective models for highly accurate prediction tasks. There are some studies on predicting organ dysfunction with bedside data and some Electronic Health Records (EHR) information, including demographic information and laboratory results. This thesis proposes machine learning models, trained with the Medical Information Mart for Intensive Care (MIMIC) -IV database, to predict total LODS with data that can be acquired bedside in the first 12 hours of ICU stay, to save time and assist doctors in treatment. The model with the best performance utilized eight features and was trained using XGBoost. It achieved a mean absolute error (MAE) of 1.4173 and a root mean square error (RMSE) of 1.8222. These models enhance the practicality and ease of application of LODS, while providing evidence-supported calculated probabilities of mortality. Moreover, this study fills the gap of predicting LODS.