Deep-Risk: Deep Learning-Based Mortality Risk Predictive Models for COVID-19

Nada M Elshennawy; Dina M Ibrahim; Amany M Sarhan; Mohamed Arafa

doi:10.3390/diagnostics12081847

Deep-Risk: Deep Learning-Based Mortality Risk Predictive Models for COVID-19

Diagnostics (Basel). 2022 Jul 30;12(8):1847. doi: 10.3390/diagnostics12081847.

Authors

Nada M Elshennawy¹, Dina M Ibrahim^{1

2}, Amany M Sarhan¹, Mohamed Arafa¹

Affiliations

¹ Department of Computers and Control Engineering, Faculty of Engineering, Tanta University, Tanta 31733, Egypt.
² Department of Information Technology, College of Computer, Qassim University, Buraydah 51452, Saudi Arabia.

Abstract

The SARS-CoV-2 virus has proliferated around the world and caused panic to all people as it claimed many lives. Since COVID-19 is highly contagious and spreads quickly, an early diagnosis is essential. Identifying the COVID-19 patients' mortality risk factors is essential for reducing this risk among infected individuals. For the timely examination of large datasets, new computing approaches must be created. Many machine learning (ML) techniques have been developed to predict the mortality risk factors and severity for COVID-19 patients. Contrary to expectations, deep learning approaches as well as ML algorithms have not been widely applied in predicting the mortality and severity from COVID-19. Furthermore, the accuracy achieved by ML algorithms is less than the anticipated values. In this work, three supervised deep learning predictive models are utilized to predict the mortality risk and severity for COVID-19 patients. The first one, which we refer to as CV-CNN, is built using a convolutional neural network (CNN); it is trained using a clinical dataset of 12,020 patients and is based on the 10-fold cross-validation (CV) approach for training and validation. The second predictive model, which we refer to as CV-LSTM + CNN, is developed by combining the long short-term memory (LSTM) approach with a CNN model. It is also trained using the clinical dataset based on the 10-fold CV approach for training and validation. The first two predictive models use the clinical dataset in its original CSV form. The last one, which we refer to as IMG-CNN, is a CNN model and is trained alternatively using the converted images of the clinical dataset, where each image corresponds to a data row from the original clinical dataset. The experimental results revealed that the IMG-CNN predictive model outperforms the other two with an average accuracy of 94.14%, a precision of 100%, a recall of 91.0%, a specificity of 100%, an F1-score of 95.3%, an AUC of 93.6%, and a loss of 0.22.

Keywords: COVID-19 detection; deep learning; machine learning; mortality and severity risk.

Grants and funding

This research received no external funding.