Multi-class autoencoder-ensembled prediction model for detection of COVID-19 severity

Abstract

Coronavirus pandemic has hampered human life all over the world with several serious casualties and numerous deceased cases. Though the pace of the pandemic slowed down owing to the vaccination process, still new mutant variants of the virus continue to evolve. Our research aims on developing an Autoencoder (AE) based multi-class prediction model for detecting the severity of coronavirus. Our proposed model is based on certain symptoms, such as fever, dry cough, fatigue, difficulty in breathing, aches, sore throat, nasal congestion, diarrhea, etc. Our AI driven prediction model can efficiently identify distinct classes of severity, which includes no-infection, mild, moderate, and severe cases where a patient may require intensive medical care. We have considered machine learning models, including logistic regression, support vector machine, and decision trees; and recurrent deep neural network model for performance comparison with proposed Autoencoder (AE) based models. Our autoencoder-based logistic regression, decision trees, and support vector machine improves the prediction performance and accuracy towards anticipating the spread of coronavirus. It has been observed that AE-based deep learning models outperformed some of the existing machine learning and deep learning techniques with a maximum accuracy of 99.66%. Hence, AE-based models were found to be more suitable for multi-class prediction of large-scale patient-centric data in terms of F1-score, accuracy, precision, recall, model testing, training loss, and other validation statistics.