A novel GA-ELM model for patient-specific mortality prediction over large-scale lab event data

Abstract Patient-specific mortality prediction models are an essential component of Clinical Decision Support Systems developed for caregivers in Intensive Care Units (ICUs), that enable timely decisions towards effective patient care and optimized ICU resource management. While high prediction accuracy is a fundamental requirement for any mortality prediction application, being able to so with minimal patient-specific data is a major plus point that can help in improving care delivery and cost optimization. Most existing scoring techniques and prediction models utilize a multitude of lab tests and patient events to predict mortality and also suffer from reduced performance when available patient data is less. In this paper, a Genetic Algorithm based Wrapper Feature Selection technique is proposed for determining most-optimal lab events that contribute predominantly to mortality, even for large-scale patient cohorts. Using this, an Extreme Learning Machine (ELM) based neural network is designed for predicting patient-specific ICU mortality. The proposed GA-ELM model was benchmarked against four popular traditional mortality scores and also state-of-the-art machine learning models for experimental validation. The GA-ELM model achieved promising results as it outperformed the traditional scoring systems by 11%–29% and state-of-the-art models by up to 14%, in terms of AUROC.