Predictive techniques based on artificial intelligence have demonstrated their ability to establish risk predictions for different pathologies associated with clinical variables in order to generate systems that facilitate medical decision-making. Because many of these prediction models do not produce good results when predictor values ​​are missing, disease progression often cannot be detected early. The purpose of this research is the implementation of a CDSS based on data imputation methods to train machine learning models that allow the prediction of the GFR in a period of 2 years with intervals of 6 months. The methodology will be divided into two stages (i) the preparation of the data made up of 482 patients with CKD in stages 1 to 4 in a period of more than 2 years of follow-up in the program, where the filtered cleaning and imputation of the variables are carried out selected to feed the models (ii) implementation of a prediction model over time and evaluation of average MAE error metrics in the different configurations of the hyperparameters of the models. Among the approaches evaluated, the Convolutional Neural Networks (CNN) performed better with a test and validation set with two different architectures for one for the periods of 6th 12th and another for 18th to 24th with a MAE 5.48, 6.46, 6.76, 6.53 for the periods respectively in the validation data.