There are multiple challenges when it comes to detection of atrial fibrillation (AF) in simplified home-based ECG measurements. In Cardiolund Research we create software algorithms that overcome these challenges, and enables automatic ECG analysis of simplified home-based measurements, for screening purposes.
One challenge is that data collected from the general population does not look like the test databases often used in research, which tends to contain signals with 50% AF and 50% normal sinus rhythm. For a research database, a simple RR variability approach provides an acceptable discrimination performance. The same technique is not applicable on databases containing hundreds of thousands, or even millions, of home-based ECG recordings from a general population of people in their seventies or eighties. Such unrefined database may contain many other types of arrhythmias and deviations, which has to be dealt with separately, and not be confused with AF.
In a general population database, containing more than 80 000, 30s Lead I, recordings from 75 year olds, we found over 1 200 signals with different combinations of bigemini, trigemini, frequent SVES, or frequent VES. These are deviations that all may be mistaken for AF when using a simple RR variability approach and that together are more common than the around 250 cases with verified AF. In addition, the database also contains many cases with AV block II, sinus arrhythmia, and faster or slower regular shorter episodes of beats that are identified by the Cardiolund software.
Another challenge is that users perform measurements in their home without assistance from trained professionals. A user may change focus, get distracted, move their body, breathe heavily, etc. while recording. This introduce large disturbances in the signal, and variations in signal amplitude throughout the recording. Such disturbances and variations, makes it hard for the QRS detector to avoid additional detections, which again makes it difficult to accurately determine the RR sequence.
In the above mentioned 80 000 signal database, there were approximately 800 signals with very poor quality, and a couple of thousands of signals with some degree of influence from disturbances in the RR series. The difficult task for the software algorithm is to discern between such irregular sequences and a true AF sequence. And, this is where the Cardiolund ECG Parser excels, it provides a robust estimate of the RR series, despite the presence of disturbances.
Some may choose to simplify the problem of detecting AF in ECG signals containing severe disturbances, by reducing the problem to only search for longer sequences of AF. But, we believe that the shorter sequences are equally important to find. Our aim with the ECG Parser is to not miss a single case of AF, and to even identify micro-AF candidates of lengths down to only a few seconds. This does of course affect the specificity of actual AF in the AF candidate category. However, with our present tuning of the ECG Parser, this approach results in a documented 88% workload reduction, while sorting out normal rhythms and poor quality cases. The remaining 12%, contains all other deviations of interest divided into AF, fast/slow/flutter episodes, AVblock II, bigemini, trigemini, and frequently occurring SVES/VES.