Predicting the evolution of COVID-19 mortality risk: A Recurrent Neural Network approach

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Epidemiology

Background:

• COVID-19 pandemic caused significant mortality in Spain by December 2020.
• Health decision support systems are vital for managing pandemics.
• The study addresses the need for predictive tools in critical healthcare situations.

Purpose of the Study:

• To apply Deep Learning for COVID-19 patient mortality prediction.
• To develop an interpretable Recurrent Neural Network (RNN) model.
• To enhance healthcare decision-making through accurate patient outcome prediction.

Main Methods:

• Utilized two datasets of COVID-19 patients from Spanish hospitals.
• Constructed temporal event sequences from clinical data.
• Trained and ensembled RNN models with attention mechanisms, performing extensive hyperparameter tuning and cross-validation.

Main Results:

• RNN models outperformed baseline classifiers (Support Vector Classifier, Random Forest).
• Ensemble models significantly improved prediction sensitivity and stability.
• Performance was assessed across various prediction timelines relative to hospital admission and outcome.

Conclusions:

• Demonstrated the feasibility of deep learning for predicting patient clinical outcomes.
• Developed a robust, interpretable RNN-based model for healthcare decision support.
• The system effectively handles sparse and heterogeneous real-world clinical data.