TL;DR
This paper introduces a novel deep learning framework, TSMixer-BiKSA, that combines Variational Mode Decomposition (VMD) with a dual-branch neural network to forecast COVID-19 trends. By effectively fusing multi-scale temporal features and external factors, the model achieves superior accuracy and robustness in multi-step predictions compared to existing methods.
What Problem Does This Paper Address?
Forecasting the spread of infectious diseases like COVID-19 is crucial for public health preparedness and policy-making. However, the complex, non-linear, and non-stationary nature of epidemic data poses significant challenges for traditional prediction methods. Existing approaches often struggle to capture multi-dimensional features and long-term dependencies, leading to insufficient accuracy, especially in multi-step forecasting. This paper aims to bridge this gap by proposing a more comprehensive model that can effectively handle these complexities.
Methods
The proposed TSMixer-BiKSA network is a multi-module deep learning framework designed for multi-step forecasting of COVID-19 cases. Its key components and novelties include:
- Dual-Branch Input: The model utilizes two parallel input branches.
- Branch 1 (Multi-feature Input): This branch incorporates daily new confirmed cases along with highly correlated external factors (identified through Pearson correlation and ExtraTrees analysis, such as hospital patients, ICU patients, new deaths, and new tests).
- Branch 2 (VMD Decomposition): This branch uses Variational Mode Decomposition (VMD) to decompose the daily new confirmed cases time series into several intrinsic mode functions (IMFs). VMD is chosen for its ability to adaptively decompose non-stationary signals into low-complexity components, preserving essential characteristics and reducing noise.
- TSMixer Module: This module, based on multilayer perceptrons (MLPs), alternately processes data along temporal and feature dimensions. It captures cross-time and cross-variable dependencies, enhancing feature mixing and temporal modeling.
- BiGRU Module: A bidirectional Gated Recurrent Unit (BiGRU) is employed to capture both forward and backward temporal dependencies, improving long-term dependency modeling and contextual representations.
- KAN Module: Kolmogorov-Arnold Networks (KAN) are integrated to extract high-order nonlinear features through layered nonlinear transformations, enhancing the model's ability to capture complex patterns.
- Self-Attention Mechanism (SA): A self-attention mechanism is used to adaptively weight and fuse features from both branches, optimizing information integration and improving prediction stability.
Datasets
The study utilizes a COVID-19 dataset from the Our World in Data repository on GitHub.
- Data Source: Publicly available data including new cases, total cases, total deaths, new deaths, ICU patients, hospital patients, total tests, new tests, total vaccinations, people vaccinated, new vaccinations, and new people vaccinated.
- Region: Italy was chosen as the target region for analysis due to its complex population distribution and mobility patterns.
- Time Span: The dataset spans from February 21, 2020, to October 12, 2021.
- Training/Testing Split: Data from February 21, 2020, to March 26, 2021 (400 days) was used for training, and data from March 27, 2021, to October 12, 2021 (200 days) was used for testing.
- Transfer Learning: For transfer learning experiments, the model was trained on Italy's data and then tested on US data. The US dataset was split into a training set (February 21, 2020, to March 26, 2021) and a test set (March 27, 2021, to October 12, 2021).
Key Insights and Ideas
- Multi-Scale Feature Fusion is Key: The paper strongly advocates for combining different types of features – raw time-series data, external factors, and decomposed signal components – to capture the multifaceted nature of epidemic spread.
- VMD's Power in Signal Decomposition: VMD proves effective in breaking down complex time-series data into simpler, more manageable components, which aids in capturing underlying patterns and reducing noise.
- Synergy of Deep Learning Modules: The combination of TSMixer (for temporal and feature mixing), BiGRU (for long-term dependencies), and KAN (for high-order nonlinearities) creates a powerful ensemble that addresses different aspects of time-series forecasting.
- Self-Attention for Adaptive Fusion: The self-attention mechanism is crucial for intelligently combining the diverse features extracted by the different modules, allowing the model to focus on the most relevant information.
- Transfer Learning for Generalizability: The model demonstrates strong transferability, performing well on US data even after being trained on Italian data, highlighting its robustness and potential for broader application.
Key Results
- Superior Accuracy: The proposed TSMixer-BiKSA network consistently outperformed baseline and existing hybrid models across various metrics (RMSE, MAE, MAPE, R²) for 1- to 3-day ahead forecasting. For instance, in a 7-day window with 3-step prediction, it achieved an eRMSE of 767.135, significantly lower than other methods.
- Improved Multi-Step Forecasting: The model showed substantial improvements in multi-step predictions, with eMAPE decreasing significantly and R² increasing, indicating better error reduction and fit.
- Effectiveness of VMD: Combining VMD decomposition with multi-feature input yielded better results than other decomposition methods like ICEEMDAN and EWT.
- Parameter Optimization: Ablation studies identified optimal parameter settings for TSMixer, BiGRU, and KAN modules, leading to the best performance.
- Robustness: The model demonstrated strong robustness and stability, especially in longer-term forecasting scenarios.
Limitations and Future Work
- VMD Decomposition Granularity: The paper acknowledges that decomposing training and test sets separately, while avoiding direct information leakage, can introduce boundary issues and inconsistencies in decomposition granularity due to different lengths. Future work could explore more advanced methods to mitigate these issues.
- KAN's Early Stage: The authors note that KAN is a relatively new architecture with limited literature and real-world applications. Its performance and stability require further evaluation and validation in broader contexts.
- Future Optimization: The authors suggest several avenues for future research:
- Advanced Decomposition: Exploring more advanced signal decomposition algorithms to further enhance feature extraction.
- Deep Reinforcement Learning: Incorporating deep reinforcement learning for dynamic and adaptive optimization of model parameters.
- Broader Applications: Extending the methodology to forecast trends of other infectious diseases to enhance generalizability and adaptability.
Summary updated