Knowledge, Machine Learning and Atrial Fibrillation: More Ingredients for a Tastier Cocktail

Fifty years after the publication of the first algorithms for the automatic detection of Atrial Fibrillation (AF), this cardiac condition is still the most studied from the computer science and engineering perspectives. Machine learning techniques are widely applied to a variety of problems, including detection, characterization, prediction and simulation, in general with promising results. In the last years, the Big Data + Deep Learning binomial is getting most of the attention in academia and industry, but on many occasions this approach fails on capitalizing all the knowledge acquired in previous decades of research. This article, written as a companion to the keynote with the same title presented in the CinC 2020 conference, tries to illustrate the importance of exploiting expert knowledge and classical approaches in synergy with the most advanced deep learning methods, which by themselves have fundamental limitations. The discussion is built around the AF detection problem and the conclusions extracted from the Physionet/CinC Challenge 2017, but the main points can be relevant in other problems for which humans have a better answer than computers, and this answer can be described.