IMO, the important part is Section 3.3 of the [paper](https://www.sciencedirect.com/science/article/pii/S1746809419301776 https://www.sciencedirect.com/science/

IMO, the important part is Section 3.3 of the [paper](https://www.sciencedirect.com/science/article/pii/S1746809419301776 https://www.sciencedirect.com/science/article/pii/S174680941...), particularly the image at [https://ars.els-cdn.com/content/image/1-s2.0-S1746809419301776-gr4_lrg.jpg](https://ars.els-cdn.com/content/image/1-s2.0-S1746809419301776-gr4_lrg.jpg) https://ars.els-cdn.com/content/image/1-s2.0-S17468094193017.... To my eye the difference in shape of the orange and green signals could also be found through more traditional signal processing/statistical means that machine learning. In a past job I did a combination of manual and machine-learning-based analysis of cardiac signals. We didn't have ECG, but did have PPG (blood flow) and PCG (sound) signals, and a pretty large study group. I recall there being one study participant who's signals were very clearly indicative of heart failure, enough that we raised the issue with our medical advisor about whether the subject should be deanonymized and contacted. In the paper they state that "the CHF subjects used in this study suffer from severe CHF only"; my suspicion is that a simpler, "hand rolled" model based on the features of the ECG could compete very well with this CNN approach for finding the same level of pathology in the ECG signal, without the "black box" of a CNN casting doubt on the technique.