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

Background:

In December 2020, the COVID-19 disease was confirmed in 1,665,775 patients and caused 45,784 deaths in Spain. At that time, health decision support systems were identified as crucial against the pandemic.

Methods:

This study applies Deep Learning techniques for mortality prediction of COVID-19 patients. Two datasets with clinical information were used. They included 2,307 and 3,870 COVID-19 infected patients admitted to two Spanish hospitals. Firstly, we built a sequence of temporal events gathering all the clinical information for each patient, comparing different data representation methods. Next, we used the sequences to train a Recurrent Neural Network (RNN) model with an attention mechanism exploring interpretability. We conducted an extensive hyperparameter search and cross-validation. Finally, we ensembled the resulting RNNs to enhance sensitivity.

Results:

We assessed the performance of our models by averaging the performance across all the days in the sequences. Additionally, we evaluated day-by-day predictions starting from both the hospital admission day and the outcome day. We compared our models with two strong baselines, Support Vector Classifier and Random Forest, and in all cases our models were superior. Furthermore, we implemented an ensemble model that substantially increased the system’s sensitivity while producing more stable predictions.

Conclusions:

We have shown the feasibility of our approach to predicting the clinical outcome of patients. The result is an RNN-based model that can support decision-making in healthcare systems aiming at interpretability. The system is robust enough to deal with real-world data and can overcome the problems derived from the sparsity and heterogeneity of data.