The COVID-19 pandemic, first detected in late 2019 in China, rapidly spread globally, becoming a major public health crisis. By February 2022, over 410 million infections and nearly 5.9 million deaths were reported. Mass vaccination programs, initiated globally in 2020, aimed to combat the pandemic. This study addresses the question of whether these programs led to a decrease in confirmed COVID-19 cases. The researchers chose to employ a Long Short-Term Memory (LSTM) network, a deep learning method known for its effectiveness in time-series analysis, to predict COVID-19 cases in the top ten countries with the highest vaccination rates. LSTM was selected due to its ability to handle long-term dependencies in time-series data, overcoming limitations of simpler Recurrent Neural Networks (RNNs). The study aimed to analyze COVID-19 case trends both before and after the implementation of mass vaccination programs to assess their effectiveness. Previous studies have utilized various deep learning methods for COVID-19 prediction, but this study distinguishes itself by focusing specifically on the comparative analysis before and after the implementation of widespread vaccination efforts. This analysis is critical in evaluating the success of global pandemic control strategies and providing valuable data-driven insights for future pandemic preparedness.

Several studies have leveraged machine and deep learning techniques to predict COVID-19 trends. These studies employed diverse methods such as ARIMA, Cubist Regression, Random Forest, Ridge Regression, Support Vector Regression, stacking ensemble learning, Bayesian Regression Neural Network, k-Nearest Neighbors, Quantile Random Forest, and various LSTM architectures (including bi-directional and convolutional LSTM). These studies showed varying levels of success, with some methods demonstrating superior predictive accuracy compared to others. The existing literature highlights the potential of deep learning in forecasting COVID-19 case numbers, but a focused study examining the direct effect of mass vaccination campaigns using LSTM networks was missing. This study fills that gap by focusing on this specific and impactful question, utilizing the well-established LSTM method as the foundation for the analysis.

The study used COVID-19 confirmed case data from a Johns Hopkins University GitHub repository, focusing on the ten countries with the highest vaccination rates (China, India, USA, Brazil, Indonesia, Japan, Pakistan, Vietnam, Mexico, and Germany). Data from January 22, 2020, to February 12, 2022, were used. Two datasets were created: 'All Time' (all data) and 'Before Vaccination' (data before February 1, 2021, assumed vaccination program start date). Data were split into 80% training and 20% testing sets, with a 14-day prediction window. Data normalization was performed using MinMaxScaler. A five-layer LSTM network (two LSTM layers, two Dropout layers, and one Dense output layer) was built using Keras and TensorFlow. The model was trained using Mean Squared Error (MSE) as the loss function. Model performance was evaluated using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE). For comparison, the same architecture was used with Vanilla RNN. The equations for MAE, RMSE, and MAPE calculations are provided. The LSTM model was also used to project the future trend of COVID-19 confirmed cases one period ahead to analyze whether there is an upward or downward trend.

The LSTM model effectively predicted COVID-19 confirmed cases across the ten countries studied. The 'All Time' scenario (including post-vaccination data) yielded better accuracy (average MAPE 5.977%) than the 'Before Vaccination' scenario (average MAPE 10.388%). The LSTM model outperformed the Vanilla RNN model, particularly regarding MAPE scores (5.977% vs 7.772% for 'All Time', and 10.388% vs 19.305% for 'Before Vaccination'). Most countries showed a downward trend in predicted future cases in the 'All Time' scenario, suggesting a positive impact from mass vaccination. However, India and Mexico showed slight upward trends. In the 'Before Vaccination' scenario, only Germany showed a significant upward trend projection, highlighting the impact of the vaccination campaign. Detailed MAE, RMSE, and MAPE values for each country in both scenarios are provided in Table 4 and the trend percentage for each country for both scenarios are provided in Table 6.

The findings largely support the hypothesis that mass vaccination programs significantly contributed to reducing the spread of COVID-19 in most of the ten countries studied. The LSTM model's ability to accurately predict case numbers and trends, especially in the 'All Time' scenario, provides compelling evidence of the vaccine's effectiveness. The contrasting trends observed in India and Mexico in both the 'All Time' and 'Before Vaccination' scenarios suggest that factors beyond vaccination, such as government response speed, vaccination program implementation, testing capacity increases, community support, and the emergence of variants like Omicron, should be considered. The superior performance of LSTM over Vanilla RNN underscores the importance of choosing appropriate time-series prediction models capable of handling long-term dependencies. The results highlight the crucial role of data-driven decision-making in pandemic management. The detailed analysis offered in this study can inform policymakers and healthcare systems.

This study demonstrates the efficacy of LSTM networks in forecasting COVID-19 cases and analyzing the impact of mass vaccination programs. While the results primarily suggest a positive correlation between vaccination and reduced case numbers, exceptions highlight the complexities of pandemic dynamics. Future research could focus on investigating the identified exceptions and integrating additional factors such as social behavior, healthcare resource allocation, and emerging variants into predictive models. This study supports the continued adoption of vaccination strategies as a crucial tool in combating pandemics and advocates for data-driven decision-making in public health policy.

The study employed relatively simple five-layer LSTM networks, focusing on applicability rather than optimization of the model. The performance metrics used did not directly measure trend movements; more advanced techniques like Directional Statistics could provide more refined analysis. The assumed start date of vaccination programs might not be entirely accurate for all countries. The emergence of new variants post-vaccination could also influence predictions.