Much of my research revolves around pattern recognition, specifically training algorithms to learn how to differentiate between the normal sinus rhythm of someone with atrial fibrillation (AF) and that of someone without AF. Although there are powerful machine learning algorithms that can essentially determine what features to look at on their own, the most robust algorithms often require initial parameters and measurements from the training dataset.
Before I can even begin trying to train an algorithm on what to look for, first I need to train myself. Most of my days in the initial few weeks involved just familiarizing myself with various EKG patterns of people with AF and people without AF. One method of doing this is through principal component analysis, a technique used to emphasize variation in a large dataset. In order to study waveform, I used an algorithm to determine the location of the high amplitude R peaks (which are the most easily identifiable EKG component of every heartbeat). After taking signal values of around 0.08 seconds before and after the R peak event, was able to overlay 30 of these plots in order to determine the variation in the waveform that surrounds the R peak. Furthermore, I visualized heart rate variability with RR vs. dRR plots, which basically plot the time interval between successive R peaks (RR-interval) and the differences between successive RR-intervals (dRR).
Currently, I am still trying to decide on the most robust measures of waveform and heart rate variability that I can feed into my automatic detection algorithm. Much of what I do is playing with data, trying to see if certain patterns exist by looking at different perspectives of the EKG signal. Next week we’ll be getting a handheld EKG machine from AliveCor that can give EKG readings in real-time, so hopefully I’ll be able to use that to look at a larger sample of normal sinus rhythm and to be able to test what type of noise (such as from exercise, breathing, electric noise) a handheld EKG machine can still produce a readable EKG with. At this point, I am just a few steps away from actually assembling the automatic detection algorithms for AF, so I am extremely excited to test how accurately the models actually work with real data.
