Computer-based detection of Atrial Fibrillation (AF) has been a relevant problem for 50 years. Interestingly, it is many times described as a "simple problem", due to the characteristic effect it has on the ventricular response and therefore on the RR interval series. However, this is just the tip of the iceberg, and under the same term we can find a variety of problems of very different nature and complexity, which are defined by diverse variables: AF type (paroxysmal, persistent, chronic), presence of other arrhythmias, source signals for detection (ECG, PPG, etc.), properties of those signals (single or multi-lead, short or long-term, in-hospital or ambulatory), etc. On the other hand, from the computer science and engineering perspectives, the variety of techniques and approaches is equally rich, ranging from purely knowledge-based expert systems to data-driven deep learning architectures; and from embedded, personalized continuous monitoring systems to large-scale, population wide analysis algorithms.
In this work we give an overview of the evolution of the field, focusing on the current state-of-the-art and discussing which problems can be considered solved, and which are the most relevant challenges that should be addressed by the scientific community in the near future. We give special importance to the discussion about what Machine Learning has contributed and will contribute to this field, with particular attention to the interpretability, complexity, and performance aspects. As a main conclusion, we will show that hybrid systems combining different Machine Learning techniques with domain knowledge have proven to have multiple quantitative and qualitative advantages over more puristic approaches.