This work focuses on comparing ECG feature extraction methods to enable a two-step approach for diagnostic modeling. Traditional feature extraction methods rely on handcrafted features extracted from fiducial points within the ECG waveform, such as the P-wave, QRS complex, and T-wave. However, these methods often fail to capture the full complexity of the signal. Recent advances in end-to-end deep learning enable automatic feature extraction, but at the cost of large datasets and substantial computational resources. Alternatively, an autoencoder is trained to reconstruct ECG signals by compressing the waveform into a latent space representation, which forces the model to learn and retain the most critical morphological features. This latent representation could subsequently be used as input for secondary lightweight machine learning models. This work evaluates the information retention of an LSTM autoencoder by comparing its reconstruction performance with that of a traditional fiducial point-based method. The median root mean square error (RMSE) for reconstruction from fiducial points is 0.223, while the autoencoder achieves a notably lower median RMSE of 0.033. Additionally, a quality score analysis shows that the optimal trade-off between latent space size and performance is achieved at a latent space size of 11. The autoencoder approach offers a promising alternative to traditional feature extraction techniques and a computationally efficient alternative to end-to-end predictive deep learning. The source code is publicly available at https://github.com/Computational-Biology-TUe/ae_waveform_compression .