BigLSTM: Recurrent neural network for the treatment of anomalous temporal signals. Application in the prediction of endotracheal obstruction in COVID-19 patients in the intensive care unit.
Journal:
Computers in biology and medicine
Published Date:
Apr 23, 2025
Abstract
Real-world applications, particularly in the medical field, often handle irregular time signals (ITS) with non-uniform intervals between measurements. These irregularities arise due to missing data, inconsistent sampling frequencies, and multi-sensor signals from different sources. Predicting outcomes using ISMTS is complex, especially when missing data is involved. This paper introduces the Binomial Gate LSTM (BigLSTM), a modular Recurrent Neural Network model designed to process ISMTS. Built on the LSTM network, BigLSTM integrates techniques for handling irregular time intervals and multiple sampling rates by injecting information redundancy. BigLSTM comprises five interconnected modules. Four are dedicated to information processing: Information Distribution, Central Computing, Predictive, and Time Axis Processing Modules. These modules ensure the redundancy of system, making it tolerant to missing data. The fifth module, LSTM Cells On/Off Control, manages the internal operations of the network. BigLSTM was tested on a critical clinical problem: predicting endotracheal obstruction in COVID-19 patients in intensive care units using ventilatory signals from 96 patients. BigLSTM achieved a mean validation mean squared error (MSE) of 0.028 for patients with obstructions and 0.2 for the entire dataset. Additionally, we analysed the prediction tendencies of the system, finding an advance trend of 3.87 days and a delay trend of 2.15 days for distant predictions (7 days), with shorter intervals for near predictions (48 h). BigLSTM provided an obstruction prediction, in the short-term, not earlier than the next 10.64 h, and not later than the next 6.8 days, with a confidence percentage of 95%, indicating its effectiveness in handling irregular time series data.