Oropouche, Dengue, and Chikungunya differential diagnosis using Brazilian surveillance data: development and validation of predictive models

Technical Summary: Differential Diagnosis of Oropouche, Dengue, and Chikungunya Using Brazilian Surveillance Data

Problem Statement
Since 2024, Brazil has experienced a significant outbreak of Oropouche Fever (OF), expanding from its traditional Amazonian range to southeastern states like Espírito Santo. While predictive models exist for Dengue and Chikungunya, there is a critical gap in tools capable of differentiating between Dengue, Chikungunya, and the emerging OF within a clinical and surveillance context. These arboviruses share heterogeneous and overlapping clinical presentations (e.g., fever, myalgia, arthralgia), complicating differential diagnosis. Furthermore, co-infections occur in approximately 2.1% of cases globally, and existing literature offers limited data on distinguishing OF from Dengue, with previous attempts yielding poor performance. The study addresses the need for validated clinical prediction models to aid in the diagnostic work-up and epidemiological monitoring of these co-circulating viruses.

Methodology
The study utilized surveillance data from the Espírito Santo state health system in Brazil, covering the period from March 2023 to July 2025.

• Data Source: The dataset comprised 465,280 observations derived from electronic notification forms. Cases were classified into four outcomes: Dengue, Chikungunya, Oropouche Fever (OF), and "discarded" (cases with no confirmed diagnosis). Zika cases were included in the initial data but excluded from final modeling due to unacceptable model performance.
• Case Definition: Outcomes were determined via clinical-epidemiological confirmation (epidemiological link + compatible symptoms) or laboratory confirmation (RT-qPCR or specific IgM).
• Predictors: The models incorporated 28 potential predictors extracted from notification forms, including demographic data (age, sex, race, education, address), clinical signs (fever, headache, vomiting, arthralgia, rash, etc.), and comorbidities (diabetes, hypertension, etc.).
• Modeling Approach: The researchers employed Random Forest regression models, developing a separate model for each outcome (Dengue, Chikungunya, OF) rather than a single multiclass model, due to the complexity of co-infections.
  • Data Split: Data was randomly divided into a training set (two-thirds) and a validation set (one-third).
  • Feature Selection: Predictors were selected using a "shadow attribute" method, comparing variable importance against shuffled "shadow" attributes to retain only those outperforming random noise.
  • Handling Imbalance: Class imbalance was managed by weighting observations inversely to outcome density.
  • Calibration: Predictions were calibrated using the Platt scaling method to improve probability estimates.
• Evaluation: Performance was assessed using ROC AUC, R², calibration metrics (intercept, slope, error metrics), and sensitivity-specificity trade-offs.

Key Results

• Dataset Composition: The final analyzed dataset included 261,691 Dengue cases (56.6%), 18,676 Chikungunya cases (4.0%), 12,174 OF cases (2.6%), and 179,115 discarded cases (38.6%).
• Predictor Selection:
  • Dengue Model: Retained 24 predictors. The most discriminatory features included leukopenia, vomiting, fever, headache, retro-orbital pain, petechiae, and the week of symptom onset.
  • Chikungunya Model: Retained 25 predictors. Key discriminators were intense arthralgia, arthritis, petechiae, and back pain.
  • OF Model: Retained 16 predictors. Unlike the other models, epidemiological features (week of first symptoms, age, education, address, skin color) were the most relevant discriminators, rather than specific clinical symptoms.
• Model Performance:
  • Dengue: ROC AUC of 0.726 in the validation set. It demonstrated the best calibration metrics but the lowest discrimination among the three.
  • Chikungunya: ROC AUC of 0.851 in the validation set, showing good discrimination.
  • Oropouche (OF): ROC AUC of 0.896 in the validation set, exhibiting the highest discrimination but less optimal calibration at higher probability ranges.
• Decision Support: The study defined specific probability thresholds to categorize predictions into "uncertain," "inconclusive," and "conclusive" ranges to assist users in decision-making. For instance, the Dengue model suggested an inconclusive range of 0.47 to 0.67.

Significance and Claims
The authors claim that this research successfully identifies the most discriminative predictors for Dengue, Chikungunya, and OF, filling a gap in the literature regarding OF differentiation. The study validates three distinct predictive models with moderate to very good performance, which are intended to serve as decision-support tools rather than replacements for clinical judgment or laboratory confirmation.

The significance of the work lies in its potential to:

1. Optimize Diagnostic Work-ups: By providing probabilistic estimates based on readily available clinical and epidemiological data, the models can help prioritize patients for confirmatory laboratory testing, particularly in resource-limited settings.
2. Enhance Surveillance: The models can aid in the early detection and monitoring of arbovirus outbreaks, especially as OF expands into new geographic areas.
3. Address Emerging Threats: This is one of the few studies to attempt the clinical modeling of Oropouche Fever, acknowledging its growing epidemiological importance in Latin America.

The authors emphasize that the models are context-specific to the Brazilian surveillance data and may require updates if the seasonality or clinical presentation of OF changes. They explicitly state that the models do not replace laboratory work-ups but aim to refine the rational use of complementary tests and support clinical reasoning during the co-circulation of multiple arboviruses. A web application (Shiny app) implementing these models is made available for public use.