AI-based Models for SARS-CoV-2 Severity Scores using Multiple Chest X-Ray Image Features

As SARS-CoV-2 threatens global public health, computer aided detection has been frequently proposed to assist medical professionals in screening patients for the infection. Based on the state-of-the-art, clinical images, such as chest Xrays (CXR), have been used in artificial intelligence processes like convolutional neural networks in detecting the virus. Additionally, no study has proposed a classification model that accurately predicts the overall SARS-CoV-2 severity scores (OSS) of patients using multiple CXR image features. Therefore, the present study introduces a model that predicts the patients' OSS using their CXR image features. The dataset consisted of 1007 CXR images opacity and geographic extent scores-generated by COVIDNet models in addition to the airspace disease grading scores as the input features while the OSS served as the output label. The dataset was preprocessed using the rescaling, binning, and data splitting techniques to increase the models' predictive accuracy. Among all the trained classification models, the Neural Network-based Trilayered model achieved the highest accuracy of 81.19%, in predicting the OSS specifically for mild severity score. Significantly, the study proves the potential of ML models in accurately predicting the OSS based on CXR image features; however, it is recommended to create a more balanced and larger dataset, and consult more radiologists for the CXR images' OSS validation.

Automated summary

This study introduces a model that predicts the overall severity scores of patients with SARS-CoV-2 using their chest X-ray image features. The results show the potential of machine learning models in accurately predicting the severity of the disease based on image features.