<p>Dengue fever is one of the main public health problems worldwide because of the complexity of its spatiotemporal transmission and the co-circulation of four antigenically different serotypes (DENV-1 to DENV-4). The identification of outbreaks and serotype classification in an accurate way are hindered by the diffused nature of the data sources, scarcity of case records with too many zeros, and high genomic heterogeneity. Accordingly, the present paper presents an integrated deep learning architecture CNN-GNN-ZI-LSTM, which can jointly model genomic, epidemiological and environmental data to aid dengue outbreak detection and serotype prediction. The convolutional neural network (CNN) is used to extract discriminative sequences from dengue genomic sequences, a zero-inflated long short-term memory network is used to extract temporal dynamics of outbreak dynamics in the spin of sparse surveillance information, and a graph neural network (GNN) is integrated to sepi features and temporal relationships between place. Experiments, using c. 2000 genomic samples and c. 10 000 WHO outbreak reports, to validate the proposed framework show that the fps framework outperforms the state-of-the-art baselines and achieves 97.20% accuracy with balanced precision-recall characteristics as well as robust discriminative performance (AUC = 0.96). Ablation studies confirm the incremental performance provided by each and every module that validate the significance of modelling zero-inflation and adopting a graph-based integration for feature. The proposed framework thus provides a scalable basis for serotype aware dengue early warning systems and provides the basis for targeted public health interventions.</p>