For type 1 diabetes patients, maintaining the blood glucose (BG) at normal values is a challenging task due to e.g. variable insulin reactions, diets, lifestyles, emotional conditions, etc. Hyperglycemic and hypoglycemic events can generate various complications (e.g. diabetic ketoacidosis, retinopathy, neuropathy, etc.), so predicting BG values in time is of great importance for diabetes self-management. Herein, we propose a non-linear autoregressive neural network approach, based on the minimal dataset available from a continuous glucose monitoring (CGM) sensor, with an integrated measure of intra-patient BG variability. The method kept the balance between accuracy and complexity, allowing a fast response with no additional effort or discomfort for the patient.