Framework and Preliminaries
The global impact of the COVID-19 pandemic, characterized by its extensive societal, economic, and environmental challenges, escalated with the emergence of variants of concern (VOCs) in 2020. Governments, grappling with the unpredictable evolution of VOCs, faced the need for agile decision support systems to safeguard nations effectively. This article introduces the Variant-Informed Decision Support System (VIDSS), designed to dynamically adapt to each variant of concern’s unique characteristics.
Dataset Description and Preprocessing
The basis of our analysis in this study lies in the COVID-19 world dataset sourced from Our World in Data (https://ourworldindata.org/coronavirus). This extensive dataset spans the entire duration of the COVID-19 pandemic, with daily updates meticulously recorded up to March 5, 2024.
In this study, the investigation revolves around the analysis of VOCs, with a particular focus on the integration of COVID-19 variants data sourced from ECDC (https://www.ecdc.europa.eu/en/publications-data/data-virus-variants-covid-19-eueea) and GISAID (https://www.gisaid.org). This comprehensive dataset encompasses the incidence of cases within each country, categorized by specific VOCs and subvariants.
Before exploring the details of the dataset features, it is imperative to delineate the preprocessing procedures undertaken:
- In the COVID-19 variant dataset, all subvariants are relabeled with their corresponding major VOCs.
- The COVID-19 variant dataset is streamlined to focus solely on the investigation of the five major VOCs: Alpha, Beta, Gamma, Delta, and Omicron.
- Features in the COVID-19 world dataset are aggregated based on the year-week timeframe, aligning with the temporal resolution in the COVID-19 variant dataset.
- The COVID-19 world dataset is filtered to exclusively include countries present in the COVID-19 variants dataset.
- Both datasets are merged based on country and year-week features, fostering a comprehensive alignment of information.
- The merged dataset is further aggregated based on country and VOCs.
- Null columns are eliminated, and any missing values are imputed with the median.
MADM Techniques
Utilizing MADM techniques necessitates the application of both weighting and ranking methods. In this context, we employ the Criteria Importance Through the Intercriteria Correlation (CRITIC) method for weighting the criteria and the Combined Compromise Solution (CoCoSo) technique to rank the alternatives.
The CRITIC method assigns weights to attributes based on a decision matrix, ensuring coherence without contradictions. It offers advantages such as accounting for interdependencies between attributes and effectively quantifying qualitative attributes.
On the other hand, the CoCoSo ranking method integrates a hybrid model, combining simple additive weighting (SAW) and exponentially weighted product (EWP) approaches. It serves as a versatile tool for generating compromise solutions in decision-making scenarios.
Transfer Learning
Transfer learning is a prevalent machine learning approach that involves leveraging a model previously trained on a specific task and applying it to a new, related task. This technique is particularly useful in mitigating the challenges associated with training, especially when dealing with limited datasets.
Transfer learning encompasses four distinct types: instance-based transfer, feature-based transfer, parameter-based transfer, and relational-based transfer. These approaches provide a foundational framework for understanding various methods in transfer learning and serve as a platform for the development of innovative techniques.
The Variant-Informed Decision Support System (VIDSS)
To facilitate understanding, we present a detailed explanation of the VIDSS framework in Fig. 2. The process initiates with the acquisition of a comprehensive dataset, integrating information on the spread of COVID-19 associated with each VOC and the relevant features of countries affected.
- The dataset is categorized, with each subset filtered by individual VOCs.
- To establish a foundation for subsequent analyses, the relationships between two VOC datasets are investigated.
- For VOC 1, the CRITIC method is employed to weight each feature.
- The CoCoSo method is applied to assess countries’ performance based on these weights, assigning scores accordingly.
- Parameters for the neural network are selected, and the model is trained.
- Feature importance is examined, elucidating the influential factors in the virus spread.
- The model’s parameters are saved for future VOC models.
For VOC 2, analogous actions are undertaken to calculate scores for each country. The novel criterion, the Relative Performance Index (RPI), is introduced, considering a country’s performance relative to its past and its growth compared to other countries.
Transfer learning is applied, utilizing parameters trained in the prior VOC model. The subsequent training of the model determines the most crucial features. Policymakers can then derive essential insights to refine their strategies, leveraging the tools provided by the VIDSS framework: feature importance, RPI, and the model of spread.
Results
In this section, the outcomes of the VIDSS framework are meticulously examined, with a structured discussion on the various VOCs in a specific order based on their chronological emergence: Alpha, Beta, Gamma, Delta, and Omicron.
Alpha Variant
The neural network model is initialized with random parameters, and after the training phase, the learned weights and biases are stored for use in the subsequent variant. The study employs a rigorous K-fold cross-validation approach with k=10 to robustly assess model performance across multiple folds.
Feature importance is vividly portrayed through SHAP plots, shedding light on the significance of various features in the model’s decision-making process. Simultaneously, the CRITIC-CoCoSo integrated approaches yield scores for each country, offering valuable insights into their performance.
Beta Variant
The concept of transfer learning is applied, utilizing the weights and biases derived from the Alpha model as initial parameters for the neural network. Feature importance is represented through SHAP plots, and the RPI criterion is derived after scoring via MADM methods.
Gamma, Delta, and Omicron Variants
The analysis of the remaining variants follows a similar approach, maintaining the interconnected relationships between datasets and consistently applying the comprehensive neural network parameter settings outlined in Table 3.
Managerial Insights
The feature importance analysis across different COVID-19 variants reveals critical insights into the pandemic’s dynamics and response measures’ effectiveness. The consistent importance of weekly hospital admissions and ICU patients underscores the need for robust healthcare infrastructure to manage severe cases and prevent healthcare system overload.
The evolving significance of vaccination-related features highlights the profound impact of vaccination campaigns in mitigating the spread and severity of COVID-19. As the pandemic progresses and vaccination efforts intensify, the importance of total vaccinations and people vaccinated becomes more pronounced, indicating the shifting focus towards achieving herd immunity and reducing overall transmission.
The inclusion of testing metrics in the Beta variant emphasizes the critical role of testing infrastructure in promptly identifying and isolating cases, and curbing the spread. These dynamic shifts in feature importance offer valuable insights for shaping and refining pandemic management policies based on evolving scenarios and key determinants.
Conclusion and Future Research Directions
The VIDSS framework offers a systematic and innovative approach to understanding and managing the spread of VOCs, providing policymakers with indispensable tools for navigating the complex landscape of pandemics. By incorporating MADM techniques and transfer learning, the framework enables dynamic adaptation to the unique characteristics of each VOC, enhancing forecast accuracy and informing effective policy decisions.
While the VIDSS framework presents promising potential, certain limitations must be acknowledged. The system’s reliance on data availability and accuracy, as well as the rapid evolution of VOCs, pose challenges in maintaining its relevance and adaptability. Future research should consider integrating additional features, enhancing spatial and temporal dimensions, and developing mechanisms for real-time integration of emerging data on new variants and subvariants.
By addressing these limitations and continuously refining the VIDSS framework, researchers and policymakers can capitalize on its capabilities to navigate the complexities of COVID-19 and other similar pandemics, ultimately leading to more effective and resilient strategies in safeguarding nations.
