Abstract
Dengue and chikungunya are endemic arboviral diseases in many low- and middle-income countries, often co-circulating and presenting with overlapping symptoms that hinder early diagnosis. Timely differentiation is critical, especially in resource-limited settings where laboratory testing is unavailable. We developed and evaluated machine-learning (ML)- and deep-learning (DL) models to classify dengue, chikungunya, and discarded cases using a large-scale, real-world dataset of over 6.7 million entries from Brazil (2013-2020). After applying the Synthetic Minority Oversampling Technique (SMOTE) to address class imbalance, we trained six ML models and one artificial neural network (ANN) using only demographic, clinical, and comorbidity features. The Random Forest model achieved strong multi-class classification performance (Recall: 0.9288, the Area Under the Curve (AUC): 0.9865). The ANN model excelled in identifying chikungunya cases (Recall: 0.9986, AUC: 0.9283), suggesting its suitability for rapid screening. External validation confirmed the generalizability of our models, particularly for distinguishing discarded cases. Our models demonstrate high-accuracy in differentiating dengue and chikungunya using routinely collected clinical and epidemiological data. This work supports the development of Artificial Intelligence-powered decision-support tools to assist frontline healthcare workers in under-resourced settings and aligns with the One Health approach to improving surveillance and diagnosis of neglected tropical diseases.