Baker Oras, Ziran Zahra, Mecella Massimo, Subaramaniam Kasthuri, Palaniappan Sellappan
Faculty of Computing and Emerging Technology, Ravensbourne University London, London SE10 0EW, UK.
Department of Computer, Control and Management Engineering, Sapienza University of Rome, Via Ariosto 25, 00185 Rome, Italy.
Int J Environ Res Public Health. 2025 Apr 4;22(4):562. doi: 10.3390/ijerph22040562.
This study proposes a data-driven approach to leveraging large-scale COVID-19 datasets to enhance the predictive modeling of disease spread in the early stages. We systematically evaluate three machine learning models-ARIMA, Prophet, and LSTM-using a comprehensive framework that incorporates time-series analysis, multivariate data integration, and a Multi-Criteria Decision Making (MCDM) technique to assess model performance. The study focuses on key features such as daily confirmed cases, geographic variations, and temporal trends, while considering data constraints and adaptability across different scenarios. Our findings reveal that LSTM and ARIMA consistently outperform Prophet, with LSTM achieving the highest predictive accuracy in most cases, particularly when trained on 20-week datasets. ARIMA, however, demonstrates superior stability and reliability across varying time frames, making it a robust choice for short-term forecasting. A direct comparative analysis with existing approaches highlights the strengths and limitations of each model, emphasizing the importance of region-specific data characteristics and training periods. The proposed methodology not only identifies optimal predictive strategies but also establishes a foundation for automating predictive analysis, enabling timely and data-driven decision-making for disease control and prevention. This research is validated using data from New Zealand and its major trading partners-China, Australia, the United States, Japan, and Germany-demonstrating its applicability across diverse contexts. The results contribute to the development of adaptive forecasting frameworks that can empower public health authorities to respond proactively to emerging health threats.
本研究提出了一种数据驱动的方法，利用大规模的新冠肺炎数据集来加强疾病早期传播的预测建模。我们使用一个综合框架系统地评估了三种机器学习模型——自回归积分移动平均模型（ARIMA）、先知模型（Prophet）和长短期记忆网络（LSTM），该框架纳入了时间序列分析、多变量数据整合以及多准则决策（MCDM）技术来评估模型性能。该研究聚焦于每日确诊病例、地理差异和时间趋势等关键特征，同时考虑了数据限制以及不同场景下的适应性。我们的研究结果表明，长短期记忆网络和自回归积分移动平均模型始终优于先知模型，在大多数情况下，长短期记忆网络具有最高的预测准确率，尤其是在20周数据集上进行训练时。然而，自回归积分移动平均模型在不同时间框架内表现出卓越的稳定性和可靠性，使其成为短期预测的可靠选择。与现有方法的直接对比分析突出了每个模型的优势和局限性，强调了特定区域数据特征和训练周期的重要性。所提出的方法不仅确定了最优预测策略，还为自动化预测分析奠定了基础，能够为疾病控制和预防提供及时且基于数据的决策依据。本研究使用来自新西兰及其主要贸易伙伴——中国、澳大利亚、美国、日本和德国的数据进行了验证，证明了其在不同背景下的适用性。这些结果有助于开发适应性预测框架，使公共卫生当局能够积极应对新出现的健康威胁。
