A Unified Weighted Ensemble Framework for Simultaneous Early Sepsis Prediction and ICU Intervention Requirement Forecasting

Abstract

Sepsis is estimated to impact 48.9 million people globally annually and responsible for almost 11 million deaths, and is the leading cause of death in hospitalised patients in the UK and Ireland. Most existing machine learning and deep learning methods are dedicated to predicting a single outcome, have a learning strategy that considers only a single row, can possibly lead to inter-patient information leakage, and need a different model for each intervention in the ICUs, which limits clinical applicability. To overcome these drawbacks, a unified patient-level multi-task weighted ensemble approach for the simultaneous prediction of sepsis, mechanical ventilation, vasopressor therapy and renal replacement therapy is proposed in this study, based on the PhysioNet Computing in Cardiology Challenge 2019, containing 20,336 ICU admissions and 790,215 hourly observations. The proposed framework employs the first 24 hours of data from the ICU to represent a patient-level summary of their data, adds clinical feature engineering, task-specific weighted ensemble learning of XGBoost–LightGBM, and optimises the F2 score threshold to better balance precision and recall and to prevent inter-patient information leakage.

The mean AUROC and mean accuracy of experimental evaluation were 0.948 and 0.8946%, respectively, where AUC were 0.841, 0.980, 0.979, and 0.991 for sepsis, mechanical ventilation, vasopressor therapy, and renal replacement therapy, respectively. Comparative evaluation showed that the proposed framework achieved better results than six of the baseline machine learning and deep learning models, with a 0.133 AUROC improvement over the best-performing deep learning model and an AUROC standard deviation of < ±0.006 over 5-fold cross validation. The results show that the proposed framework has demonstrated its capabilities to accurately classify the risk level of a patient in the early phase of hospitalization in the ICU and guide interventions, with relevance for clinical use, being computationally efficient, clinically accurate, and clinically deployable.