A hybrid machine learning framework to predict mortality in paralytic ileus patients using electronic health records (EHRs)

Paralytic ileus (PI) is the pseudo-obstruction of the intestine secondary to intestinal muscle paralysis. Causes of PI include electrolyte imbalances, gastroenteritis (inflammation or infection of the stomach or intestines), overuse of medications, abdominal surgery, etc. Predicting mortality in PI patients hospitalized to ICU is crucial for assessing the severity of illness and adjudicating the value of treatment strategy and resource planning. We have developed a Statistically Robust Machine Learning based Mortality Prediction framework namely SRML-MortalityPredictor that could potentially help intensivists, surgeons, and other medical professionals to carefully plan treatment strategies for critically ill PI patients. We used MIMIC III v1.4, a publicly available ICU database to extract patients data (with age > 18 years old) admitted to the ICU with paralytic ileus (PI) as their primary illness. At phase 1, the SRML-MortalityPredictor framework uses univariate statistical analysis to filter out those risk factors which are not associated with the label of the data. Subsequently, it uses the risk survival statistical methods such as cox-regression and Kaplan–Meier survival analyses. The cox-regression analysis provides the hazard ratio about the potential PI risk factors that later used in conjunction with the Kaplan–Meier analysis to generate a rank order list (highest to lowest risk factors). At phase 2, we used several machine learning classification approaches such as linear discriminant analysis (LDA), Gaussian naive bayes (GNB), decision tree (DT) model, k-nearest neighbor (KNN), and support vector machine (SVM) to find the one with highest predictive power using the rank order features extracted at phase 1. We have evaluated the SRML-MortalityPredictor framework and recorded the accuracy, sensitivity, specificity, receiver operating characteristic (ROC) curves, and area under the curve (AUC) scores for each model. The SRML-MortalityPredictor framework with support vector machine (using RBF kernel) showed better performance and yielded an accuracy score: 81.30% and AUC score: 81.38% while predicting mortality in PI patients. We demonstrated a feasible framework for the mortality risk prediction in PI patients admitted to the ICU. The proposed framework could potentially be helpful for intensivists in clinical decision making. Further research is necessary to incorporate more risk factors associated with PI patients to ensure the adaptability of SRML-MortalityPredictor at the bedside.