Electrocardiogram (ECG) processing systems based on deep learning offer potential for advanced cardiac analysis. However, these systems often encounter significant challenges, such as the scarcity of labeled data, which affects their performance, reliability, and integration into clinical practice. This study aims to address these challenges by proposing an Active Learning (AL) methodology to optimize data labeling, reducing annotation effort while improving model performance. We evaluate the AL approach across three distinct applications: (1) sinus rhythm beat classification using synthetic data; (2) clinical severity of noise classification with a long-term ECG monitoring repository acquired under real conditions; and (3) cardiac wave delineation using a gold-standard dataset with expert annotations from the publicly available PhysioNet QT Database (QTDB) and the Lobachevsky University ECG Database (LUDB). In each classification task, our proposed AL framework integrates a neural network based on an autoencoder that generates a visualizable latent space for explainability into the decision-making process. The system iteratively selects the most informative instances using a margin sampling strategy in the latent space and incorporates them into the training process to refine performance. Results demonstrate that the AL approach consistently outperforms random sample selection in precision, recall, and F1-score. Additionally, ECG in-line analysis shows that models trained with the AL strategy outperform those from previous studies, even when trained on smaller subsets of the experimental datasets. This approach can reduce the labeling workload of clinicians, helping to efficiently increase labeled data, improve model performance, foster confidence in decision support systems, and advance ECG analysis applications.