Abstract
During the early stages of the COVID-19 pandemic, Case Fatality Rates (CFR) varied significantly across nations. Identifying the underlying drivers was challenging due to sparse, noisy, and inconsistent data. This paper introduces MFL_COVID19, a machine learning framework utilizing Regularised Multi-task Feature Learning (MTFL) to identify critical country-specific factors. By employing Fused Sparse Group Lasso (FSGL), the model achieves superior predictive accuracy and stability compared to traditional regression methods. Key findings indicate that the Power Distance Index (PDI) and Diabetes mortality rates were primary predictors of early-phase mortality.
Introduction
In the first months of 2020, the global community observed stark differences in COVID-19 mortality. While Italy and Spain faced CFRs exceeding 10%, countries like Japan and Germany remained below 4%. Understanding these disparities required analyzing a complex array of 27 indicators across demographics, healthcare, and culture. However, standard statistical models struggled with the "small data" nature of early pandemic reports, which were often incomplete and inconsistent.
Methods: Multi-task Feature Learning
The authors propose a Regularised Multi-task Feature Learning (MTFL) approach to overcome data sparsity. Instead of treating each day or region as an isolated event, the model treats CFR prediction at different time points as related "tasks" that share common features.
The core of this framework is the Fused Sparse Group Lasso (FSGL). This regularization technique selects a consistent set of factors across time while enforcing "temporal smoothness," preventing unrealistic daily oscillations in factor importance. Additionally, a three-layered hybrid feature selection process (combining Pearson correlation, Recursive Feature Elimination, and tree-based models) prunes irrelevant data before the final analysis.
Datasets
The study analyzed the first 42 days of the outbreak across 12 countries (including the UK, China, Italy, and Singapore) and 17 Chinese provinces. Data was aggregated from diverse sources such as Worldometer, the WHO, and the World Bank to capture a holistic view of national contexts.
Results and Key Insights
The MTFL approach significantly outperformed traditional methods. The FSGL model achieved a modified root Mean Square Error (rMSE) of 91.95, compared to 150.08 for Ridge regression and 146.98 for Lasso.
As shown in the comparison above, FSGL provides a much cleaner and more interpretable feature set by removing irrelevant noise while maintaining consistency across the temporal dimension.
Critical Factors Identified
- Power Distance Index (PDI): A cultural metric reflecting societal hierarchy. High-PDI countries (often in Asia) saw more effective implementation of lockdowns and suppression measures.
- Diabetes Mellitus Mortality Rate: Identified as a more consistent predictor of CFR than raw hospital bed counts.
- Gender Rate: Demographic distribution played a significant role.
- SPAR_detect Capability: A country’s inherent ability to detect health emergencies under International Health Regulations.
Conclusion and Future Work
MFL_COVID19 demonstrates that cultural and specific health indicators were more influential in early pandemic mortality than previously assumed. While the model is constrained by the quality of reported case counts, it provides a robust framework for extracting insights from noisy global data. Future research will focus on integrating epidemiological simulations (like SEIR models) to further refine the baseline for CFR estimates.
Summary updated