Title:A Machine Learning Approach to Predict In-Hospital Mortality in
COVID-19 Patients with Underlying Cardiovascular Disease using Artificial
Neural Network
Volume: 18
Issue: 4
Author(s): Samaneh Sabouri, Mohammad Hossein Khademian, Mehrdad Sharifi, Razieh Sadat Mousavi-Roknabadi and Vahid Ebrahimi*
Affiliation:
- Department of Biostatistics, School of Medicine, Shiraz University of Medical Sciences, Shiraz, Iran
Keywords:
Artificial neural network, cardiovascular disease, COVID-19, machine learning, cohort study, ROC curve.
Abstract:
Background: Machine learning algorithms, such as artificial neural networks (ANN),
provide more accurate predictions by discovering complex patterns within data. Since COVID-19
disease is prevalent, using advanced statistical tools can upgrade clinical decision making by identifying
high risk patients at the time of admission.
Objective: This study aims to predict in-hospital mortality in COVID-19 patients with underlying
cardiovascular disease (CVD) using the ANN model.
Methods: In the current retrospective cohort study, 880 COVID-19 patients with underlying CVD
were enrolled from 26 health centers affiliated with Shiraz University of Medical Sciences and followed
up from 10 June to 26 December 2020. The five-fold cross-validation method was utilized to
build the optimal ANN model for predicting in-hospital death. Moreover, the predictive power of
the ANN model was assessed with concordance indices and the area under the ROC curve (AUC).
Results: The median (95% CI) survival time of hospitalization was 16.7 (15.2-18.2) days and the
empirical death rate was calculated to be 17.5%. About 81.5% of intubated COVID-19 patients
were dead and the majority of the patients were admitted to the hospital with triage level two (54%).
According to the ANN model, intubation, blood urea nitrogen, C-reactive protein, lactate dehydrogenase,
and serum calcium were the most important prognostic indicators associated with patients’
in-hospital mortality. In addition, the accuracy of the ANN model was obtained to be 83.4%, with a
sensitivity and specificity of 72.7% and 85.6%, respectively (AUC=0.861).
Conclusion: In this study, the ANN model demonstrated a good performance in the prediction of
in-hospital mortality in COVID-19 patients with a history of CVD.