Abstract
The clinical differentiation of Chikungunya from other acute febrile illnesses poses a significant diagnostic challenge due to the substantial symptom overlap among co-circulating arboviruses such as Dengue and Zika, particularly in resource-constrained settings where laboratory confirmation is frequently delayed or unavailable. This study proposes an interpretable machine learning framework for the prediction of laboratory-confirmed Chikungunya cases versus suspected cases that were subsequently discarded after laboratory investigation, using a large-scale, nationally representative dataset of 12,709 notification records derived from Brazil's Information System for Notifiable Diseases (SINAN), spanning 2013 to 2020 and encompassing 26 clinical symptom, comorbidity, and sociodemographic features. Four supervised learning algorithms-Logistic Regression, Random Forest, XGBoost, and LightGBM-were systematically evaluated under 5-fold stratified cross-validation, with Random Forest achieving the highest discriminative performance (AUC = 0.785 ± 0.005) and XGBoost demonstrating the best probability calibration reliability. Considering the trade-off between discrimination and calibration, XGBoost was identified as the most suitable model for potential clinical deployment. To ensure clinical transparency, a structured three-level SHapley Additive exPlanations (SHAP) interpretability analysis was conducted on both gradient boosting models, encompassing global feature importance ranking, feature-level directional and interaction effects, and local patient-level prediction decomposition. The SHAP analysis consistently identified arthralgia (mean |SHAP|: 0.7407 in XGBoost, 0.7489 in LightGBM) and fever (0.5043 and 0.4881, respectively) as the two dominant predictors of Chikungunya confirmation, followed by age, education level, and rash, while comorbidities contributed negligibly to case discrimination. Cross-model validation between XGBoost and LightGBM revealed highly concordant feature importance rankings and directional effect patterns, confirming that the identified clinical predictors are robust and independent of algorithmic choice. These findings demonstrate the practical value of combining machine learning with SHAP-based explainability for supporting clinical triage of suspected arboviral cases, providing a transparent, evidence-based diagnostic support tool that aligns data-driven insights with established clinical knowledge and enables personalized patient-level explanations at the point of care. Importantly, the model distinguishes confirmed Chikungunya from discarded suspected cases rather than from confirmed infections with other specific arboviruses, and its predictions should be interpreted within this operational context.