A Bayesian Approach to Explore Risk Factors for Respiratory Dysfunction in Intensive Care Unit Patient

Abstract

Respiratory dysfunction and failure are common in the intensive care unit (ICU); they are often the primary reasons for ICU admission and affect length of stay, mortality, and cost. However, diagnosing respiratory dysfunction requires arterial blood gas values to calculate the partial pressure of arterial oxygen (PaO2) to a fraction of inspired oxygen (FiO2) or P/F ratio. These intermittent blood gas values may be difficult to obtain in some patients or where financial resources are limited. Its varying etiologies and lack of other specific biomarkers make diagnosing difficult without this measurement. Thus, in this study, we investigate commonly available parameters in the ICU for the classification of respiratory dysfunction without arterial blood gas values using a Bayesian network, an unsupervised structural learning method. Clinical data from selected patients in the Medical Information Mart for Intensive Care (MIMIC) III v1.4 database is used to create and validate these models. Bayesian network generated using the taboo order algorithm showed a satisfying performance in the classification of respiratory dysfunction. Results are compared to standard diagnosis with P/F ratio. The predictor variables selected could stratify respiratory dysfunction with 80% accuracy and 94% sensitivity. Hence, without using arterial blood gas values, these parameters could identify respiratory dysfunction in 90% of cases using Bayesian networks.