Due to the latest advances in machine learning algorithms new deep learning-based approaches to the interpretation of 12-lead electrocardiograms have been developed, demonstrating the quality of diagnostics comparable to the expert one. In this paper, we propose several techniques increasing the quality of ECG classification by a deep neural network. The techniques include patient metadata incorporation, signal denoising and self-adaptive model training. The experimental validation of the approaches was carried out on a novel dataset containing 64198 standard ECG recordings obtained during routine medical practice. The conducted experiments demonstrated statistically significant quality growth compared to the baseline, supporting the further application of our findings.