Data-driven AI approach narrows diagnostic gap between dengue and chikungunya

Dengue and chikungunya, the two mosquito-borne diseases that frequently circulate at the same time, share the same Aedes vector, and present with nearly identical early symptoms such as fever, headache, muscle pain, and joint discomfort. In many regions where outbreaks are most severe, laboratory confirmation is either delayed or unavailable, increasing the risk of misdiagnosis, delayed care, and avoidable complications.

A new large-scale artificial intelligence study now suggests that routine clinical and epidemiological data may be enough to reliably differentiate dengue from chikungunya, even in highly imbalanced real-world conditions. The findings point to a potential shift in how frontline health workers screen arboviral infections in under-resourced settings, particularly during peak transmission seasons when health systems are stretched thin.

The research, titled AI-Assisted Differentiation of Dengue and Chikungunya Using Big, Imbalanced Epidemiological Data, was published in Tropical Medicine and Infectious Disease and is based on more than 6.7 million anonymized case records from Brazil collected between 2013 and 2020.

Leveraging millions of real-world cases to address a persistent diagnostic gap

The study was designed to address a critical limitation in earlier artificial intelligence research on arboviral diseases. While previous models often relied on small, balanced datasets or binary classification tasks, real-world surveillance data are rarely balanced and rarely limited to simple yes-or-no diagnoses. Dengue cases typically far outnumber chikungunya cases, and a substantial share of reported cases ultimately remain inconclusive or discarded after review.

To reflect this reality, the researchers used a national Brazilian dataset containing more than 4.3 million dengue cases, 325,000 chikungunya cases, and over 2.1 million discarded cases. Laboratory variables were deliberately excluded to mirror the conditions faced by clinicians working in remote or resource-limited settings. Instead, the models were trained on demographic information, reported symptoms, and comorbidities that are routinely collected during patient intake.

One of the defining features of the study was its approach to extreme class imbalance. Chikungunya represented the smallest and most clinically risky group to miss, yet it accounted for only a fraction of total cases. The researchers evaluated several imbalance-handling techniques before selecting a synthetic oversampling method that allowed minority class patterns to be learned without discarding real dengue cases. This decision proved central to model performance and reflects a growing recognition that imbalance handling is as important as algorithm selection in public health AI.

Six machine learning models were developed and tested alongside a deep learning artificial neural network. The machine learning models included tree-based and distance-based approaches commonly used in clinical prediction tasks. All models were evaluated using metrics designed to reflect real diagnostic performance rather than simple accuracy, which can be misleading in imbalanced datasets.

The results showed that a Random Forest model consistently delivered the strongest overall performance. It achieved high sensitivity and discriminative power across all three classes, including dengue, chikungunya, and discarded cases. Its ability to correctly identify discarded cases was particularly strong and remained stable across validation and internal test sets, suggesting that the model generalizes well beyond its training data.

The artificial neural network, while less balanced in overall performance, demonstrated exceptional sensitivity for chikungunya. Its ability to identify nearly all chikungunya cases positions it as a potential screening tool in outbreak scenarios where the cost of missed diagnoses is high, even if some false positives are tolerated.

Clinical and epidemiological signals drive AI performance

High diagnostic accuracy can be achieved without laboratory markers, which are often unavailable during the early stages of infection or in rural clinics. Instead, the most informative inputs combined clinical symptoms with demographic and epidemiological context.

Age emerged as a meaningful predictor, with chikungunya cases tending to occur in slightly older populations compared to dengue. Gender also played a role, with women more frequently represented among chikungunya cases, aligning with earlier epidemiological observations. Geographic variables, including location of residence and infection, contributed significantly, underscoring the importance of spatial context in arboviral transmission patterns.

Symptom profiles further helped the models distinguish between diseases, even when overlap was substantial. Fever, headache, joint pain, rash, nausea, vomiting, and retro-orbital pain all contributed to classification decisions. Comorbidities such as hypertension, diabetes, and autoimmune conditions were included, though their predictive weight was lower than that of demographic and symptom-related features.

Notably, some variables traditionally considered critical for clinical decision-making were missing from the dataset, including the number of days since symptom onset and markers of disease severity. Despite these gaps, the models achieved performance levels that exceeded those reported in earlier multi-class studies using smaller or artificially balanced datasets.

This finding challenges the assumption that epidemiological variables add little value to clinical diagnosis. Instead, the study shows that when processed at scale through machine learning, demographic and temporal signals can meaningfully improve diagnostic differentiation, especially when laboratory confirmation is unavailable.

The study also highlights the trade-offs between machine learning and deep learning approaches in public health contexts. While deep learning models can automatically learn complex patterns, they may struggle to balance sensitivity and precision when data are noisy and imbalanced. Tree-based machine learning models, by contrast, demonstrated strong robustness and interpretability advantages, making them more suitable for deployment in decision-support tools.

Implications for public health surveillance and frontline care

In many dengue-endemic countries, health workers must make rapid triage decisions with limited information, particularly during outbreaks that overwhelm laboratory capacity. Misclassification can lead to inappropriate monitoring, delayed hospitalization, or missed opportunities to prevent complications.

By demonstrating that large-scale AI models can reliably distinguish dengue from chikungunya using routine data, the study points to a practical path forward for digital decision support in low-resource environments. Such tools could be deployed on basic computer systems or mobile devices, allowing clinicians to input patient information and receive probabilistic guidance in real time.

The ability of the Random Forest model to accurately identify discarded cases also has implications for surveillance efficiency. Reducing false positives can help public health authorities focus limited resources on true cases, improving outbreak tracking and response planning.

The study also acknowledges its limitations. The reliance on a single oversampling technique limits direct comparison with studies that use alternative imbalance strategies. The absence of interpretability analysis means that while feature importance is known at a high level, the precise contribution of individual variables to specific predictions remains unclear. The authors note that future work should incorporate explainability tools to support clinical trust and adoption.

Despite these constraints, the scale and realism of the dataset set this study apart. Few previous efforts have attempted multi-class arboviral classification on millions of records while preserving the imbalance inherent in real surveillance data. By incorporating an internal test set rather than relying solely on cross-validation, the researchers also strengthened the credibility of their performance claims.

The authors suggest expanding the framework to include other arboviral diseases such as Zika and yellow fever, further enhancing its value for integrated disease surveillance. Such expansion would align with One Health objectives, which emphasize coordinated human, environmental, and vector-based monitoring.