GENERATIVE-DISCRIMINATIVE HYBRID MODELING FOR BINARY CLASSIFICATION IN PUBLIC HEALTH

This work proposes and evaluates a hybrid binary classifier for public health data with binary attributes, such as clinical signs and symptoms. Classification problems with exclusively binary attributes present specific challenges, such as limitation in the individual expressiveness of variables and greater sensitivity to dependencies between attributes. The proposed approach combines the Bernoulli Naive Bayes generative model, which assumes a simple causal structure between diagnosis and symptoms, with logistic regression, a discriminative model widely used for prediction. The final prediction is constructed as a weighted combination of the outputs of the two models, with weights adjusted by minimizing logistic loss, resulting in a second-order optimization problem. The methodological motivation is based on the complementarity between the statistical robustness of generative models and the discriminatory power of discriminative models. The model is applied to a real dataset from DATASUS, focusing on the classification of Severe Acute Respiratory Syndrome (SARS). The results are evaluated for accuracy, probabilistic calibration, and robustness in different subsets of the population. The proposal stands out for incorporating a structural reading of the data, bringing together statistical modeling and causal learning in public health contexts.