Optimal Prediction and Disease Severity Classification of Proteomic Survival in Pre and Post-Covid-19 Using Hybrid Machine Learning Approach

Abstract

Uncertainty surrounds the underlying mechanisms of the severe COVID-19 disease of 2019. The capability to detect COVID-19 through artificial intelligence techniques, particularly deep learning, will help to do so in the early stages, which will increase the likelihood that patients around the world will recover rapidly. The load on the healthcare system globally will be relieved as a result. Several thousand plasmas and serum proteins from COVID-19 patients and symptomatic controls are longitudinally analysed in this study to identify non-immune and immune proteins associated with COVID-19. The development of predictive models thus involves taking into account the topological variations across networks from different scenarios (survivors vs. non-survivors). As a result, the study's test subjects, who weren't included in the machine learning (ML) training, had high prediction accuracy. This study successfully predicted the existence of critically ill (CI) patients both before and after COVID-19 by using an MLM built on a synonymic network that incorporates measurements of several proteins. A rise in some acute phase and inflammatory proteins (IP) with time (e.g. ITIH3, SAA1; CRP, SAA2, LBP, SERPINA1, and LRG1) is related to the danger of death after COVID-19, while an upsurge of kallikrein (KLKB1), kallistatin (SERPINA4), thrombin (F2), Apo lipoprotein C3 (APOC3), GPLD1, and the protease inhibitor A2M, is associated with survival. The same clinical symptoms, such as dry cough, fever, squatness of breath, and others, are linked to both severe and critical patients. The lesion outlines are then retrieved from the COVID-19-contaminated regions after the entropy texture features have been extracted using a Gray-level co-occurrence Matrix (GLCM) to confirm the infected regions (IR). Further, the study implemented a variety of features using CT images with a CNN-based Inception V3 model for selection algorithms to filter significant features. Finally, construct a model of transfer learning (TL) using the VGGNet16 model which could capture and further classify the disease severity. Based on Matlab software, the suggested work is assessed. With a compassion of 96.7% and specificity of 98.2%, the results demonstrate that VGGNet16 is the most suitable TL model to identify COVID-19, nonetheless, it also exceeds the most advanced methods at the moment. The clotting system and accompaniment cataract are home to the bulk of proteins in the forecast model with high significance. This work shows that plasma proteomics (PP) can result in prognostic predictions that vastly outperform the present prognostic markers in critical care, respectively.