Dealing with the database variability problem in learning from medical data: an ensemble-based approach using convolutional neural networks and a case of study applied to automatic sleep scoring