MXLPred: A Natural Language Processing Deep Learning Model for ICU Patients’ Data

Early estimation of a patient’s mortality rate and length of stay is crucial for saving lives. With the launch of Electronic Health Records (EHRs) in the healthcare industry, many studies have been done on clinical decision-making with EHR. The publicly accessible dataset, the Medical Information Mart for Intensive Care IV (MIMIC-IV) database, is used in this study. We propose a prediction of mortality and hospital stay length Mortality Prediction Ensemble Model with XLNet and GRU (MXLPred) framework to address the three main challenges of accurate medical condition prediction tools and data visualisation. First, we have created an ensemble model from eXtreme Language Net (XLNet) and Gated Recurrent Unit (GRU) learning algorithms to predict the first 24-hour and 90-day mortality rates during a patient’s Intensive Care Unit (ICU) stay. We have utilised label-wise FastText embedding to identify medical entities from patient discharge summary notes, which has improved the clinical embedding and mortality estimates. We have also fine-tuned the model for better prediction results. Second, the Extreme Gradient Boosting (XGBoost) model for regression is utilised to predict the length of ICU stay of patients admitted in the ICU. Third, BigQuery is used to visualise the patient’s current and past medical history to aid in quick decision-making. We compare the above prediction estimates with other state-of-the-art algorithms. We demonstrate that the final ensemble MXLPred shows improved outcomes regarding better Area Under the Receiver Operating Characteristic curve (AUROC) and Mean Squared Error (MSE).