An autoregressive integrated moving average and long short-term memory (ARIM-LSTM) hybrid model for multi-source epidemic data prediction

Benfeng Wang, Yuqi Shen, Xiaoran Yan, Xiangjie Kong

TL;DR

This paper introduces a novel hybrid model, ARIMA-LSTM-FC, that combines traditional time-series forecasting (ARIMA) with deep learning (LSTM) and a Bayes-Attention mechanism to predict epidemic trends using multiple data sources. The model significantly improves prediction accuracy by effectively fusing diverse data, outperforming individual models and existing hybrid approaches.

What Problem Does This Paper Address?

The COVID-19 pandemic highlighted the critical need for accurate and timely epidemic forecasting to inform public health responses and mitigate economic damage. However, predicting epidemic trajectories is challenging due to their inherent complexity, including periodic dependencies and dynamic, non-linear variations. Relying solely on confirmed case data can be insufficient due to data scarcity and potential reporting lags. While hybrid models combining ARIMA and LSTM have shown promise, effectively integrating diverse auxiliary data sources (like mobility data) to enhance prediction accuracy remains a significant challenge. This paper aims to address this by proposing a multi-source data fusion approach to improve epidemic trend prediction.

Methods

The core of the proposed methodology is the ARIMA-LSTM-FC model:

  • ARIMA-LSTM Hybrid: The model first leverages the strengths of both ARIMA (for linear and periodic patterns) and LSTM (for complex, non-linear patterns) to forecast epidemic trends. The time series data is segmented into low-volatility and high-volatility components. ARIMA is used for the low-volatility part, while LSTM handles the high-volatility part. The predictions from both are then combined.
  • Bayes-Attention Fusion Mechanism: This is the novel contribution for integrating multi-source data. After obtaining initial predictions from individual data sources (e.g., confirmed cases, population mobility) using the ARIMA-LSTM model, a fully connected (FC) layer with a Bayes-Attention mechanism is employed. This mechanism dynamically allocates attention weights to different data sources, allowing the model to learn which sources are most relevant for prediction at any given time. It then fuses these weighted predictions into the primary case data.

Datasets

The study utilizes publicly available datasets:

  • Cumulative Confirmed Case Numbers: Data for New York State from March 2, 2020, to February 24, 2023.
  • Daily Population Mobility Trajectory Numbers: Data from February 15, 2020, to October 15, 2022.

These datasets were sourced from the Johns Hopkins University website and a U.S. COVID-19 pandemic population mobility dataset. The data underwent outlier cleaning and mean substitution. The first 80% of the data was used for training, and the remaining 20% for testing.

Key Insights and Ideas

  • Synergy of ARIMA and LSTM: Combining ARIMA's ability to capture linear trends and seasonality with LSTM's capacity for complex non-linear pattern recognition is crucial for accurately modeling epidemic dynamics.
  • The Power of Multi-Source Data: Auxiliary data, such as population mobility, contains valuable information that, when effectively integrated, can significantly improve the accuracy of epidemic forecasts beyond what case data alone can provide.
  • Dynamic Attention is Key: A static fusion of data sources is less effective than a dynamic approach. The Bayes-Attention mechanism allows the model to adaptively weigh the importance of different data sources, leading to more robust predictions.
  • ARIMA-LSTM-FC Outperforms: The proposed hybrid model with data fusion consistently outperforms individual ARIMA, LSTM, and even other advanced hybrid models, demonstrating its effectiveness.

Key Results

The ARIMA-LSTM-FC model achieved superior performance across all evaluated metrics on the test set:

  • RMSE: 0.022 (best)
  • MAPE: 0.42 (best)
  • MAE: 0.032 (best)
  • R²: 0.993 (best)

These results significantly outperform other models, including standalone LSTM, GRU, Transformer, and CNN-ARIMA-LSTM, as well as the simpler ARIMA-LSTM model without data fusion. The paper highlights that the ARIMA-LSTM-FC model's prediction performance surpasses that of the ARIMA-LSTM model, and the latter's performance exceeds that of standalone LSTM or ARIMA models, confirming the benefits of both hybridization and multi-source data fusion.

Limitations and Future Work

  • Limitations: The paper focuses on COVID-19 data from New York State. While the authors suggest universal applicability, further validation on diverse geographical regions and different epidemic outbreaks would strengthen this claim. The study also acknowledges that while the Bayes-Attention mechanism dynamically allocates weights, it does not impose constraints on the data, which might be a consideration for future work.
  • Future Work: The authors plan to propose more efficient models for fusing multi-source data. They also aim to explore predicting epidemic development trends in countries where case data is scarce, suggesting a focus on leveraging alternative data sources in data-limited scenarios.

Summary updated

Abstract

The COVID-19 pandemic has far-reaching impacts on the global economy and public health. To prevent the recurrence of pandemic outbreaks, the development of short-term prediction models is of paramount importance. We propose an ARIMA-LSTM (autoregressive integrated moving average and long short-term memory) model for predicting future cases and utilize multi-source data to enhance prediction performance. Firstly, we employ the ARIMA-LSTM model to forecast the developmental trends of multi-source data separately. Subsequently, we introduce a Bayes-Attention mechanism to integrate the prediction outcomes from auxiliary data sources into the case data. Finally, experiments are conducted based on real datasets. The results demonstrate a close correlation between predicted and actual case numbers, with superior prediction performance of this model compared to baseline and other state-of-the-art methods.
BibTeX
@article{wang2024autoregressive,
  author = {Benfeng Wang and Yuqi Shen and Xiaoran Yan and Xiangjie Kong},
  title = {{An autoregressive integrated moving average and long short-term memory (ARIM-LSTM) hybrid model for multi-source epidemic data prediction}},
  journal = {PeerJ Comput. Sci.},
  year = {2024},
  doi = {10.7717/peerj-cs.2046},
}
Figure