Anticipating dengue outbreaks using a novel hybrid ARIMA-ARNN model with exogenous variables

Indranil Ghosh, Saurabh Gupta, Santu RanaarXiv 2023 · DOI · arXiv

TL;DR

Researchers have developed a novel hybrid forecasting model that combines classical statistical techniques (ARIMAX) with neural networks (ARNN) to predict dengue outbreaks. By using rainfall as a key external factor, this hybrid approach effectively captures both linear trends and complex non-linear patterns, often outperforming sophisticated deep learning models on limited epidemiological datasets.

What Problem Does This Paper Address?

Dengue fever is a massive global health challenge, with cases skyrocketing from roughly half a million in 2000 to over 5 million in 2019. Because there is no universal vaccine or specific medicine, public health authorities rely heavily on early warning systems to allocate resources and launch awareness campaigns.

Predicting dengue is notoriously difficult. Its incidence is influenced by environmental factors like rainfall and temperature, and the data contains both "linear" components (predictable trends) and "non-linear" components (complex, chaotic fluctuations). Traditional statistical models like ARIMA are great at the former but fail at the latter. Conversely, pure machine learning models can capture non-linearity but often require massive amounts of data that aren't always available in public health settings.

Methods

The authors propose a multivariate hybrid model called ARIMAX-ARNN. The "X" stands for exogenous variables—in this case, rainfall. The process works in two distinct phases:

  1. The Linear Phase (ARIMAX): The model first fits an Auto-Regressive Integrated Moving Average model with exogenous variables. This captures the linear relationship between past dengue cases, past errors, and rainfall.
  2. The Non-Linear Phase (ARNN): No model is perfect; the ARIMAX model leaves behind "residuals" (errors). The authors argue these residuals aren't just random noise—they contain hidden non-linear patterns. They use an Auto-Regressive Neural Network (a feed-forward network using lagged values as input) to model these residuals.

The final forecast is simply the sum of the ARIMAX prediction and the ARNN residual prediction. Crucially, the authors also provide a mathematical proof for the asymptotic stationarity of their model, ensuring it remains stable over time.

Datasets

The study utilizes two well-known weekly time series datasets from the NOAA Dengue Forecasting project:

  • San Juan, Puerto Rico: Data spanning 1990/1991 to 2008/2009.
  • Iquitos, Peru: Data spanning 2000/2001 to 2008/2009.

Both datasets include weekly dengue incidence and weekly rainfall measurements. The authors used a "sliding window" approach to split the data, gradually increasing the training set size to test the model's robustness across different forecast horizons (from 4 weeks up to 52 weeks).

Key Insights and Ideas

  • The "Residual" Goldmine: The core "aha moment" is that the error term of a classical model isn't just trash; it’s a signal that a neural network can decode. By letting the statistical model handle the "obvious" trends, the neural network can focus its power on the complex "noise."
  • Deep Learning Isn't a Silver Bullet: The paper compares the hybrid model against heavy hitters like Transformers and Temporal Convolutional Networks (TCNs). Interestingly, while Transformers occasionally performed very well, they were inconsistent. The authors suggest this is because deep learning models typically require much longer historical datasets than what is available for most disease outbreaks.
  • Rainfall as a Driver: Confirming previous ecological studies, the inclusion of rainfall as an exogenous variable significantly improved the forecasting capability, particularly in the hybrid framework.

Key Results

The hybrid ARIMAX-ARNN model was benchmarked against seven other models (including ARIMAX, ETSX, and various deep learning architectures) using Mean Absolute Error (MAE) and Root Mean Square Error (RMSE).

  • San Juan: The hybrid model showed superior performance and consistency, particularly in the 28, 32, and 36-week forecast horizons.
  • Iquitos: The model significantly improved upon standard ARIMAX forecasts for horizons of 20 weeks or longer.
  • Consistency: Unlike the deep learning models (TransformersX, NBeatsX), which showed volatile performance as the training data size changed, the hybrid model remained stable and competitive across almost all scenarios.

Limitations and Future Work

While the results are promising, the authors acknowledge a few areas for growth:

  • Limited Variables: This study focused primarily on rainfall. Future iterations could incorporate temperature, humidity, and socio-economic factors to further refine accuracy.
  • Data Scarcity: The inconsistency of the deep learning benchmarks highlights a general problem in epidemiology: models must be "data-efficient."
  • Generalization: The authors suggest this hybrid framework isn't just for dengue; it could be applied to other infectious diseases, stock market predictions, or weather forecasting—anywhere that exogenous variables and complex time-series patterns coexist.

Summary updated

Abstract

Dengue incidence forecasting using hybrid models has been surging in the data rich world. Hybridization of statistical time series forecasting models and machine learning models are explored for dengue forecasting with different degrees of success. In this paper, we propose a multivariate expansion of the hybrid ARIMA-ARNN model. The main motivation is to propose a novel hybridization and apply it to dengue outbreak prediction. The asymptotic stationarity of the proposed model has been established. We check the forecasting capability and robustness of the forecasts through numerical experiments. State-of-the-art forecasting models for multivariate time series data are compared with the proposed model using accuracy metrics. Dengue incidence data from San Juan and Iquitos are utilized along with rainfall as an exogenous variable. Results indicate that the proposed model improves the ARIMAX forecasts in some situations and closely follows it otherwise. The theoretical as well as experimental results reinforce that the proposed model has the potential to act as a candidate for early warning of dengue outbreaks. The proposed model can be readily generalized to incorporate more exogenous variables and also applied to other time series forecasting problems wherever exogenous variable(s) are available.
BibTeX
@misc{ghosh2023anticipating,
  author = {Indranil Ghosh and Saurabh Gupta and Santu Rana},
  title = {{Anticipating dengue outbreaks using a novel hybrid ARIMA-ARNN model with exogenous variables}},
  year = {2023},
  eprint = {2312.15270},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2312.15270},
}
Figure