Prediction of the Number of COVID-19 Confirmed Cases Based on K-Means-LSTM

TL;DR

This paper introduces a novel K-Means-LSTM hybrid model that improves short-term COVID-19 case prediction accuracy in Louisiana by combining feature selection, clustering, and deep learning techniques, outperforming traditional SEIR models.

Contribution

The paper presents a new combined algorithm integrating K-Means, Xgboost, and LSTM for more accurate COVID-19 case forecasting, addressing overfitting issues in existing models.

Findings

K-Means-LSTM achieved RMSE of 601.20, significantly lower than SEIR's 3615.83.

The hybrid model effectively captures regional similarities for better predictions.

The approach reduces overfitting and improves short-term forecast accuracy.

Abstract

COVID-19 is a pandemic disease that began to rapidly spread in the US with the first case detected on January 19, 2020, in Washington State. March 9, 2020, and then increased rapidly with total cases of 25,739 as of April 20, 2020. The Covid-19 pandemic is so unnerving that it is difficult to understand how any person is affected by the virus. Although most people with coronavirus 81%, according to the U.S. Centers for Disease Control and Prevention (CDC), will have little to mild symptoms, others may rely on a ventilator to breathe or not at all. SEIR models have broad applicability in predicting the outcome of the population with a variety of diseases. However, many researchers use these models without validating the necessary hypotheses. Far too many researchers often "overfit" the data by using too many predictor variables and small sample sizes to create models. Models thus developed are unlikely to stand validity check on a separate group of population and regions. The researcher remains unaware that overfitting has occurred, without attempting such validation. In the paper, we present a combination algorithm that combines similar days features selection based on the region using Xgboost, K Means, and long short-term memory (LSTM) neural networks to construct a prediction model (i.e., K-Means-LSTM) for short-term COVID-19 cases forecasting in Louisana state USA. The weighted k-means algorithm based on extreme gradient boosting is used to evaluate the similarity between the forecasts and past days. The results show that the method with K-Means-LSTM has a higher accuracy with an RMSE of 601.20 whereas the SEIR model with an RMSE of 3615.83.