Ponentes
Descripción
Dengue is a globally significant endemic virus, with annual infections exceeding 100 million. In Paraguay, incidence patterns have repeatedly strained the health system, while current preventive measures remain broadly generalised and overlook regional heterogeneity. Prior research indicates that grouping geographic areas with similar incidence dynamics significantly improves predictive accuracy (Bogado Machuca, J.V. et al., CLEI Electron. J. 26, 2023). Building on this premise, this study introduces a novel mathematical framework based on a Clustering of Clusters (CofC) approach to forecast local dengue incidence using sets of epidemic curves. The proposed methodology seeks to identify underlying spatio-temporal patterns by sequentially grouping time series data. To account for temporal shifts and phase delays in epidemiological peaks, the Dynamic Time Warping (DTW) distance is employed as the core metric. Specifically, the time series are first clustered by locality, grouping geographic areas that share similar incidence dynamics. Subsequently, a second-level clustering is applied to group these localised medoids by epidemiological years, effectively capturing seasonal variations and long-term temporal trends. The mathematical objective function for finding the optimal configuration in any given clustering level minimizes the total intra-cluster variance, which is formalized as:\begin{equation}J = \sum_{k=1}^{K} \sum_{X_i \in C_k} DTW(X_i, \mu_k)\end{equation}Where $K$ represents the number of clusters, $C_k$ denotes the k-th cluster, $X_i$ is a local time series of dengue incidence, and $\mu_k$ is the medoid series representing the center of cluster $C_k$. By applying this operation hierarchically across spatial and temporal dimensions, the CofC framework effectively reduces the dimensionality of the predictive space while preserving critical regional dynamics (Davis, N. et al., in: Proc. 19th IEEE Int. Conf. Intell. Transp. Syst., pp. 223-228, 2016). The methodology is evaluated against single-level clustering and non-clustered baselines across regional, departmental, and national scales. Using multiple projection models, the forecasting performance is analysed on the aggregated macro-regions. Results demonstrate that the CofC approach improves the forecasting of local outbreaks by accurately identifying zones with homologous epidemiological behaviours across different years. This integration of nonlinear time series clustering and mathematical biology provides a computationally robust tool to support regionalised public health interventions and targeted vector control strategies. Agradecimientos: ARASY ESTR01-23.