Cardiovascular diseases are still one of the topmost causes of deaths worldwide. The 12-lead electrocardiography is an extremely important non-invasive modality for diagnosis of CVDs. Deep leaning and artificial intelligence increasingly recur in automatic diagnosis systems; however, most of the approaches that exist are based on semi- supervised learning, where a lot of unlabeled data takes place. But in the medical field, large amounts of annotated ECG data are available to use in supervised learning. This work adopts a robust supervised learning framework for ECG-based predicting CVD. We address the challenge of detection of multiple co-occurring CVDs within a single ECG recording by adapting advanced preprocessing techniques and model architectures commonly deployed in semi-supervised and multi- label learning models. Our model is annotated ECG signals-efficient and accurately predictive while utilizing labeled data. The strong signal preprocessing methods plus scalability of deep architecture towards multi-label classification will be a representation of some of the best contributions.