Learning accurate personalized survival models for predicting hospital discharge and mortality of COVID-19 patients

Area of Science:

• Epidemiology
• Biostatistics
• Health Informatics

Background:

• The COVID-19 pandemic (emerged December 2019) caused significant global healthcare strain, overwhelming medical facilities and necessitating efficient resource allocation.
• Effective patient triage and resource management became critical due to high hospitalization and mortality rates, straining intensive care units and auxiliary resources.

Purpose of the Study:

• To develop and evaluate learned survival prediction models, termed Individual Survival Distributions (ISDs), to assist healthcare professionals in decision-making for patient triage and resource allocation.
• To assess the predictive performance of various combinations of covariates (age, sex, location, chronic conditions) and demographic/economic data in ISD models for COVID-19 outcomes.

Main Methods:

• Created multiple data subsets from a public COVID-19 epidemiological dataset.
• Developed and compared several Individual Survival Distribution (ISD) models using various covariates, including demographic and economic information.
• Evaluated model performance using multiple metrics to determine the most effective survival prediction algorithm.

Main Results:

• The gradient boosting Cox machine algorithm significantly outperformed competing techniques in predicting both hospital discharge and COVID-19 mortality.
• The study identified key covariates and data augmentation strategies that enhance the accuracy of survival prediction models.
• Extensive experiments and comparative analyses validated the superior performance of the selected algorithm.

Conclusions:

• Learned survival prediction models, particularly the gradient boosting Cox machine, offer a valuable tool for optimizing medical resource allocation and patient management during pandemics.
• The developed models and curated datasets provide a foundation for future research in predictive modeling for infectious diseases.
• Availability of curated datasets and code facilitates reproducibility and further advancements in healthcare informatics.