The work includes statistical methodology development of heterogeneous treatment recommendation when missing data exist. We would like to recommend individualized glucose (A1c) control levels for complex patients with diabetes accounting for the patient heterogeneity. However, the A1c measurements across time are subject to missing data, where the missingness could be potentially informative about the values and the health risk conditions. We propose Bayesian latent profile models to infer the latent risk categories. The recommendations can be extended to multiple points, as dynamic treatment regime. Two main strategies will be investigated. First, using multiple imputation to handle the missing data and then perform comparative effectiveness research to evaluate the tight control effect. Second, integrating the missing data analysis into the model for treatment recommendations. We will develop the computational algorithms and convert to an open source R package.