Predictions For COVID-19 With Deep Learning Models of Long Short-Term Memory (LSTM)

Introduction

Coronavirus disease (COVID-19) is a dangerous and infectious disease caused by Severe Acute Respiratory Syndrome Coronavirus-2 (SARS-CoV2) has become a pandemic that has shaken the whole world. Currently, in June 2020, according to the data posted by Johns Hopkins University of Baltimore, Maryland USA and New York Times (NYT), more than 181 million cases have been reported across the world, while deaths tragically climbed to 3.93 million. The United States leads the world in total cases at 33.6 million, and the deaths toll around 604,000. Fortunately, based on the latest data from the Centers of Disease Control and Prevention (CDC), nearly 131 million people are fully vaccinated, and 66% of the total population have received at least one dose. Daily new cases plummed to a relatively low number by the protection of the vaccine. However, the delta variant, also known as B.1.617.2, is still one of the concern viruses as it seems to be around 60% more transmissible and may cause more severe symptoms. Due to the delta variant, despite being fully vaccinated, breakthrough infections could occur and rapidly increase new cases reported in Israel and India.

The World Health Organization (WHO) warns that COVID-19 could be spread starting with one person, then exaggerated through contact and respiratory spay. Pandemic has always been individually but shared worldwide. According to the character of this virus, many countries have to close their borders and enforce a partial or complete lock-down which had a devastating impact on tons of industry workers. The sudden lock-down often creates a large wave of problems in terms of agriculture, transportation, especially for third-world countries and low-income communities. Overwhelmed hospitals and lack of oxygen made the patient unable to have enough treatment and finally lost their lives. As the pandemic destroyed standard lifestyle, governments urgently know whether the peak of this pandemic is passed away? When is the proper time to reopen? Under this circumstance, prediction of the COVID-19 new cases and deaths becomes crucial to help the governments and healthcare systems prepare in advance about what kind of policy is more appropriate for their situation and check whether the policy undertaken is effective or not.

Prominent computational and statistical models have been unrealistic as the dynamic transmission of the epidemic, highly nonlinear, long interval (several days and months), and high variance, is hard to describe by traditional methods. Thanks to enough good data and computational resources, machine learning and deep learning techniques have recently gained lots of progress, known as solving more complex models and acquiring relationships without providing too much prior information. Recurrent Neural Networks (RNN) is a powerful tool to model sequence data by its unique connection structure and has been widely used in recognition tasks (Robinson et al., 2002; Graves et al., 2013; Sak et al., 2014). Based on the Recurrent Neural Network, Long Short-Term Memory (LSTM) (Sepp et al., 1997) adds a feedback connection to its structure. It makes it possible to process data sequentially while keeping robustness against the long-term dependency issue. LSTM is also well-known for its outstanding performance among handwriting recognition, speech recognition, and making predictions based on time series data.

Benefit from the high accuracy prediction behavior of deep learning. Several studies were published about modeling the diagnosing and spreading procedure. COVID-19 infection forecasting of all states in India has been investigated through the Long Short-Term Memory (LSTM) and predicted the new cases of the next few days (Arora et al., 2020; Chandra et al., 2020; Shahid et al., 2020; Shastri et al., 2020). The mutation rate of COVID-19 has been modeled and forecasted using the Recurrent Neural Networks (RNN) based technique (Pathan et al., 2020). As the current data of confirmed cases and death is available to the public, we consider using the relative model on this up-to-date data and conducting our prediction and discussion after the fitting procedure.