This work aims to enhance classification performance by generating new data through a Generative Adversarial Model for two problems: the Heart Failure (HF) diagnosis and the HF severity estimation problem. The conditioned by class generated data added to the real data are exploited in a classical training-test learning framework using eight conventional machine learning classifiers. The data are provided in the framework of the Kardiatool project by the University College Dublin (UCD) and the University Hospital of Ioannina, 2nd Department of Cardiology (UOI) from 487 subjects with demographic, laboratory, medication, risk factor, medical history, and physiological information. Classification performance of 95.97% and 90.23% in terms of accuracy is achieved when exploiting the generated data for the HF diagnosis and the HF severity estimation problem, respectively. In both cases, the average performance across the conventional classifiers rises by 2.43% and 7.39% in accuracy and the F1-score improves by 2.47% and 8.78%.Clinical Relevance- Automated patient classification (in terms of NYHA class) can become a valuable tool for HF patients management, overcoming the problem of subjective evaluation of patient condition. Towards this direction, synthetic data generation and interpretability are critical keys.