<p>Multiclass logistic regression is a widely used technique for classification tasks, aiming to predict categorical outcomes based on a set of covariates. In this paper, we extend the multiclass logistic regression framework to handle missing covariates that follow a Gaussian mixture distribution, presenting a comprehensive approach for inference and prediction. Specifically, we propose a hybrid method that combines the Expectation Maximization (EM) and the Stochastic Approximation EM (SAEM) algorithms to iteratively estimate model parameters while accounting for missing data and mixture covariate structures. This approach avoids the complex simulation steps typically associated with SAEM while enhancing accuracy. Additionally, model selection is performed using multiple information criteria (BIC, AIC and ICL), and standard errors are estimated to assess the uncertainty of parameter estimates. Extensive simulations based on synthetic data and real-world datasets—including the Indian Pines hyperspectral dataset from AVIRIS and the Iris dataset—demonstrate the effectiveness of our method in handling missing covariates and improving predictive performance in multiclass logistic regression settings.</p>