ABSTRACT
Nowadays there is a world pandemic of a challenging respiratory illness, COVID-19. A large part of COVID-19 patients evolves to severe or fatal complications and require an ICU admission. COVID-19 mortality rate approaches 30% due to complications such as obstruction of the trachea and bronchi of patients during the ICU stay. An endotracheal obstruction occurring during any moment in a COVID-19 patient ICU stay is one of the most complicated situations that clinicians must face and solve. Therefore, it is very important to know in advance when a COVID-19 patient could enter in the pre-obstruction zone. In this work we present an intelligent computing solution to predict endotracheal obstruction for COVID-19 patients in ICU. It is called the Binomial Gate LSTM (BigLSTM), a new and innovative deep modular neural architecture based on the recurrent neural network LSTM. Its main feature is its ability to handle missing data and to deal with time series with no regular sample frequency. These are the main characteristics of the BigLSTM information environment. This ability is implemented in BigLSTM by an information redundancy injection mechanism and how it copes with time control. We applied BigLSTM with first wave COVID-19 patients in ICU of Complejo Hospitalario Universitario Insular Materno Infantil. Encouraging results, even while working with a very small data set, indicate that our developed computing solution is going forwards towards an efficient intelligent prediction system which is very appropriate for this kind of problem. © 2021, Springer Nature Switzerland AG.