A machine learning approach for differential classification of chikungunya virus: an explainable AI study based on surveillance data.

PURPOSE: Mosquito-borne viral diseases such as chikungunya usually present similar clinical symptoms, which makes them hard to differentiate without laboratory confirmatory tests. Therefore, this work retrospectively developed interpretable machine learning models to improve the classification of chikungunya using routinely collected data. METHODS: Brazilian surveillance data, with 18,127 records, were analyzed, and eight different machine learning algorithms were trained and five-fold cross-validated. In addition, SHapley Additive exPlanations (SHAP) and permutation feature importance described model behavior. RESULTS: In univariable analysis, arthralgia (OR = 5.84, 95% CI 5.39-6.33) and fever (OR = 5.85, 95% CI 5.29-6.46) were most positively associated with Chikungunya. The macro-AUC values ranged from 0.89 to 0.90 for all models in the multiclass analysis. The Chikungunya-versus-Non-Chikungunya task demonstrated moderate discrimination. The Chikungunya-versus-Dengue scenario produced the highest performance, with an AUC of 0.97 in most models. CONCLUSION: The explainability mechanism characterized how fitted models utilized available clinical and demographic predictors, with acute symptom variables contributing prominently to predictions. The results demonstrate the potential of using an interpretable model-assisted approach for arboviral classification. Importantly, prospective, temporal, and external geographic validation, with harmonized data, are required before clinical decision-support or point-of-care implementation can be considered.