Comparative Analysis and Application of Deep Neural Networks in Covid-19 Prediction

Abstract

Machine learning specifically deep or reinforcement learning is a growing field. With the advent of the SARS COV 2 pandemic, its application for Covid 19 prediction has continued to evolve as traditional prediction approaches are rendered inaccurate or outdated with the challenges paused as we globally try to understand the trend of this global pandemic. Deep learning coupled with bio-statistical theory and approaches provide a new way of tackling the prediction challenge in Covid 19.This new approach factors in complex variation in variables that is non-linear, multivariate and with multiple independent variables. We assess the promise entailed in automated machine learning, SIR & Hybrid SIR Models referred to as SEIR (Susceptible-Exposed-Infected-Recovered) Models and LSTM RNN (Long-Short Term Recurrent Neural Networks. These three approaches directly inform CDC, WHO, major entities sharing prediction results on the pandemic and individual government health organs globally. We explore why they exhibit efficiency in arriving at predictions as the variables, geography and demographics fed into each keep varying. The criticality of the assessment we arrive at is rigorously tested and validated using K-Fold validation, Mean Accuracy Prediction Error (MAPE) and we also plot receiver operating characteristic (ROC) curves the results are later on exhibited showing Auto-ML.