Abstract
Introduction: Dengue fever poses a major global public health challenge, making early outbreak forecasting essential for effective intervention. While traditional time-series models provide baseline forecasts, they inadequately capture spatial coherence. Advanced approaches, including long short-term memory (LSTM) and spatial or Bayesian state-space models, better represent spatiotemporal dynamics but frequently require strong assumptions or high computational costs. However, dynamic mode decomposition (DMD) provides a low-rank, equation-free framework for extracting dominant dynamical patterns from complex datasets. Therefore, this study assessed the effectiveness of DMD in characterizing and forecasting dengue transmission dynamics across Sri Lanka.
Methods: Monthly dengue incidence data (2010-2020) from 26 cities were Z-score normalized and analyzed using standard DMD, with no exclusions. The dataset was partitioned into a pre-outbreak period (2010-2016) and an outbreak/post-outbreak period (2017-2020), with the final six months of each period reserved for testing. Then, forecast accuracy was evaluated using six-month-ahead root-mean-square error (RMSE) while not employing imputation, cross-validation, or rolling-window validation.
Results: DMD identified low-rank coherent modes corresponding to long-term trends and multi-year oscillations. In the pre-outbreak period, the stationary mode (~40%) was dominated by Colombo and Gampaha (R≈0.898), whereas during the outbreak, dominance shifted toward Jaffna (>12%). Moreover, six-month forecasts achieved high accuracy (RMSE: 0.04±0.03 pre-outbreak; RMSE: 0.005±0.006 post-outbreak), outperforming LSTM and autoregressive integrated moving average (ARIMA) models.
Conclusion: Overall, DMD effectively captures remarkable spatiotemporal dengue dynamics and enables accurate short-term forecasting. Its computational efficiency and minimal assumptions make it a promising tool for real-time dengue early warning and broader infectious disease surveillance.