Neural Network-Based Prediction in Critical COVID-19 Patients

Overview

In the context of an emerging pandemic without an established prognostic scoring system, deep learning approaches can be used to quickly develop empirical prognostic models. This study aimed to present an artificial neural network (ANN) model to predict the duration of mechanical ventilation and mortality in COVID-19 patients at the intensive care unit. Methods: Data were collected from medical records of 113 COVID-19 patients who had followed up at the intensive care unit between February 2020 and June 2020. An ANN approach was used to predict the length of mechanical ventilation and mortality in COVID-19 patients by evaluating patients' clinical data (demographic, laboratory, and comorbidities).

Coronavirus disease 2019 (COVID-19) has led to an unprecedented burden on intensive care units (ICUs), particularly due to high rates of respiratory failure requiring invasive mechanical ventilation. Early identification of patients at risk for prolonged mechanical ventilation and mortality is crucial for optimizing resource allocation and clinical decision-making. This retrospective cohort study aimed to develop and evaluate an artificial neural network (ANN) model to predict mechanical ventilation duration and in-hospital mortality among COVID-19 patients admitted to the ICU. After approval by the Gaziantep University Clinical Research Ethics Committee (Decision No: 2024/07, Date: 17.01.2024), data from 113 adult patients admitted to the ICU between February 1, 2020 and June 30, 2020 were retrospectively analyzed. Demographic characteristics, comorbidities, vital signs, laboratory parameters, severity scores (e.g., APACHE, SOFA), treatment modalities, and clinical outcomes were extracted from medical records. Artificial neural network models were developed using commercially available software (Alyuda NeuroIntelligence, Alyuda Research Inc., Los Altos, CA, USA). Multiple training algorithms, including Quick Propagation, Conjugate Gradient Descent, Limited Memory Quasi-Newton, Online Backpropagation, and Batch Backpropagation, were tested. Model performance was evaluated using 10-fold cross-validation. Predictive accuracy for mortality and correlation performance for mechanical ventilation duration were calculated. Classical statistical methods, including multiple linear regression and binary logistic regression, were also performed for comparison. The primary objective was to assess the predictive performance of ANN models for ICU mortality. A secondary objective was to evaluate ANN performance in estimating mechanical ventilation duration. This study was conducted in accordance with the Declaration of Helsinki.