Dengue fever (abbreviated as `Dengue'), an arboviral disease caused by Aedes mosquito-borne dengue virus (DENV) remains a major public health concern in most tropical and subtropical regions. Early and accurate diagnosis is crucial for improving patient outcomes and controlling outbreaks. While recent machine learning (ML)-based methods leveraging advanced imaging data (e.g., Raman spectroscopy) have shown promise, they often require infrastructure, limiting their applicability in resource-constrained settings. To address this challenge, we present SPIDER, a novel model that combines Bi-directional Long Short-Term Memory (BLSTM) with Synthetic Minority Over-sampling Technique for Nominal and Continuous. (SMOTE-NC) for enhanced dengue detection using readily available clinical and laboratory data. SPIDER leverages BLSTM's capability to learn from structured data while addressing the inherent class imbalance problem through SMOTE-NC. While BLSTM networks are typically applied to sequential data, in this work we explore their capability to model tabular clinical and laboratory data for dengue detection. This demonstrates that BLSTM can effectively capture complex dependencies and deliver strong predictive performance even on non-time-series structured datasets. Second, we also present a novel dataset to test the effectiveness of the proposed approach which can also be used to benchmark future ML-based Dengue detection algorithms. We cannot share the dataset without permission, as they contain private medical records.

SPIDER significantly improved the dengue detection results and demonstrates competitive performance compared to existing relevant detection models, achieving an accuracy of 0.783784, recall of 0.783784, precision of 0.789039, and an F1 score of 0.783079. These results demonstrate SPIDER's potential as a robust, accessible model for early dengue detection, particularly in resource-constraint settings.