Predicting COVID-19 in the UAE Using Machine Learning

Abstract

As of 17th August 2022, the United Arab Emirates (UAE) have recorded over 1 million Covid-19 cases since the start of the pandemic with over 2000 deaths and the numbers are still increasing. To combat Covid-19, UAE has taken many initiatives, which are: surveillance and contact tracing by introducing apps such as Al Hosn, containment of spread by limiting the gathering of people, online schooling and remote work, closure of public places and sanitation drives. The question then arises how well the above-mentioned initiatives worked in reducing the spread. This thesis aims to analyze the trends in Covid-19 in the UAE by predicting future number of cases based on the recorded history as well as the addition of policies and vaccinations to see the effect of such policies. This is done by using well-established Machine Learning models such as LASSO regression, Exponential Smoothing, and Deep Learning models such as LSTM, LSTM-AE and bi-directional LSTM-AE. The data used to train the various models is acquired from the UAE government, Federal Competitiveness and Statistics Centre (FCSC) and consists of numerical attributes such as number of Confirmed Cases, Recovered Cases, Deaths, Tests, and Vaccinations. This data set publicly available and is updated every day. To further analyze the trend to see if it changes after the vaccination drives and the above-mentioned initiatives are applied, an additional categorical attribute that describes whether an event has taken place, such as a national holiday or a sanitization drive was created. After doing initial analysis to understand the nature of the data, the models were trained and tested with different combinations of attributes, and it was found that the Univariate LSTM model with an input of 5-day history of Confirmed Cases performed the best with an RMSE of 275.85. This was an improvement of over 30% from the current state of the art related to the UAE. Other than the LSTMs, it was also found that the bi-directional LSTMs performed relatively well.