Relational Learning Improves Prediction of Mortality in COVID-19 in the Intensive Care Unit

Area of Science:

• Medical Informatics
• Machine Learning
• Computational Biology

Background:

• Traditional machine learning models struggle to predict Coronavirus-19 (COVID-19) outcomes from electronic health records (EHR) due to limitations in capturing complex data inter-connectivity.
• Effective prediction of COVID-19 patient mortality in intensive care units (ICU) requires advanced analytical frameworks.

Purpose of the Study:

• To propose a novel framework using relational learning and a heterogeneous graph model (HGM) for predicting COVID-19 patient mortality.
• To leverage diverse EHR data from a large patient population across multiple hospitals for robust model training.

Main Methods:

• Developed a heterogeneous graph model (HGM) incorporating Long Short-Term Memory (LSTM) for time-varying EHR data.
• Implemented a Skip-Gram relational learning strategy in the output layer, replacing traditional softmax, to compare patient and outcome embeddings.
• Utilized EHR data from over five hospitals in New York City, encompassing a diverse patient cohort.

Main Results:

• The proposed relational learning-based HGM demonstrated superior performance in predicting mortality across various time windows.
• Achieved higher area under the receiver operating characteristic curve (auROC) compared to comparator models in all prediction time windows.
• Showcased dramatic improvements in recall, indicating a more effective identification of at-risk patients.

Conclusions:

• The heterogeneous graph model (HGM) with relational learning effectively captures complex patterns in EHR data for COVID-19 mortality prediction.
• This novel framework offers a significant advancement over traditional machine learning methods for critical care outcome prediction.
• The approach holds promise for improving patient management and resource allocation in ICUs during pandemics.