A model supported in the clinical covid patient record using DL for pandemic preparedness

Abstract

This paper presents a model based on a Deep Learning approach to aid in improving our assessment of the risk of death in COVID-19 patients only based on their clinical record when admitted. It is aimed to provide an alternative fast tool for doctors and researchers to focus on a rapid selection of patients with a high likelihood of death, which is a critical aspect having in mind the growing infection dynamics of the current and perhaps future pandemics. The dataset used in this research is open-access, available for algorithm benchmarking, and represents the knowledge of the cases examined in Mexico before the immunization campaigns. The massive amount of information used to feed the algorithm provides robustness and aids in detecting the principal patterns involved in the data. The model is based on a Deep Neuronal Network, which uses different activation functions and several neurons in each hidden layer for getting a stable performance, and was tested in both a validation set and test set, obtaining a satisfactory and reliable accuracy of about 93 % for the survival prediction.