From Rhythm to Remedy: Music Meets Cardiac Electrophysiology

What inspired the concept of using music as both a conceptual and digital theranostic tool in electrophysiology?

I have been working in music technology for over 25 years. I started combining music, mathematics, and engineering in the early days of the digital music revolution, the movement that gave rise to Spotify, Apple Music, and such.

Amidst all the activity in music tech, I was also a patient. I have had cardiac arrhythmia since I was a child and have had two ablations to take care of the arrhythmias. My heart is just very creative at expressing itself. So, I decided that I would do something about it. In music technology, working with rhythms, frequencies, and signals are things that are part of the trade. When I was in the cath lab, I was captivated by all the beautiful signals, and I decided I would apply what I know about how to handle and analyze music signals to heart signals. And so that’s how I got started.

Most people don’t realize this, but there’s just so much in common between music signals and heart signals that the methods that are developed independently in these two fields are actually easily translatable. You don’t have to do very much to apply a technique in one to the other, and you get a whole other kind of analysis that gives you new perspectives and novel ways to understand the patterns and the behaviours. Such cross-domain transfers are an untapped resource, bringing knowledge, methods and technologies that have been developed in one field to the other. For example, in music, the technologies to understand and represent rhythms have been developed over hundreds of years and have barely been applied to cardiology. Cardiology has developed its own ways of looking at rhythms, ways that are valid and very useful. But why not consider alternative ways to look at the signals?

How can musical structures or principles help clinicians better understand or interpret complex electrophysiological patterns?

Let me give you two examples because this is a big subject and it’s hard to cover everything. Music structures can be categorized into those that have to do with frequencies, like harmonies, and those having to do with time, like rhythms. Each of these are major areas of research, and researchers devote their whole lives to writing books about topics in each area.

The idea for the first example came about when I was preparing to go into the cath lab. The young doctor-in-training asked me what I did. When he learnt it had to do with music, he said, “Oh, you teach music! I love music. I was in charge of our Christmas party games in the Department of Cardiology and I picked out pieces of music to play to my colleagues and had them guess which arrhythmia that was.” And then I thought, what could he have done? If it were atrial flutter, which has a baseline rate of 300 beats per minute, it could be some fraction of that, like 150 if every other beat gets through, or 75 if the third beat gets through. And so, he could have found pieces of music, recordings that were at those tempi.

But we can do much more. I knew from having worked on music generation that we can find patterns that match the rhythms of arrhythmia exactly. In music generation, we had used existing rhythmic structure to generate music with rhythmic interest and having long-term structures. We would learn the long-term structure partly from existing rhythms, rhythms of music that already existed. By using those rhythms and altering the notes, we inherit quite a lot of innate temporal structure.³

So that’s taking rhythms that exist in a score. When you listen to someone play, and you wish to transcribe that into some score representation, that’s a difficult problem. People work on automated transcription, and they write computer programs to do it. That transcription process allows us to access and represent the rhythm of the arrhythmia.⁴ To make it more fun or palatable, and perhaps to incorporate other information, we can layer on the notes and other structures and create music based on the arrhythmia.⁵

What is important about that is that when we see a report of an arrhythmia, we get only the highlights, the important highlights that a physiologist had picked out. And those are typically only 10 seconds long, not very long. So, there’s very little context.

The thing with music is it keeps evolving over time. We can listen to it and hear multi-scale relationships. When you hear the music that is based on the arrhythmia, you hear how wrong it is. You really feel it. You feel it at a level that is much more than simply looking at a graph on a piece of paper. When we hear rhythms, musical rhythms, they mimic our heartbeats, and when they are wrong, we feel it in our gut. So, it’s a great way to translate the information into a sonic form. Because the information is translatable, it means that the techniques that we have in music technology also can be applied to cardiovascular signals and vice versa.

The second example is based on expressive structures in music. In music, there’s the time structures and the frequency structures, there’s vertical structures like chords and harmonies, and horizontal structures like melodies and episodes that unfold over time. And there’s expressive structures, the expressive details, the ornaments. We have also studied expressive structures like vibratos. When a singer sings, they don’t just sing an unwavering note without expression, unless it’s music by someone like Philip Glass. Most operatic singers will have this lovely vibrato. If you’re a violinist, you’ll make a vibrato with your hands. Or if you’re a cellist, you wobble your hands to give this lovely, variable sound. This variable sound is hard to analyze, and we have developed techniques to do that for these waveforms which are between 4 and 8 hertz.

When someone has atrial fibrillation, there are these lovely waveforms in between their heartbeats, and they’re just like vibratos and between 3 and 12 hertz. So, the technique that can analyze vibratos can be used to analyze atrial fibrillation (AF), and we were able to use that for AF stratification. So based on whether someone had paroxysmal AF, early persistent AF, or long-standing persistent AF, only by characterizing these “vibratos” in between (they’re called fibrillatory waves in AF), we were able to detect which category of AF it was with nearly 60% accuracy; chance is 33%. So, with only this one feature, it had already improved the odds quite a bit.⁶ There are many other examples like this in music that would also work for cardiovascular signals.

In what ways might music-based digital tools contribute to diagnosis, treatment planning, or given even therapeutic intentions in arrhythmia care?

This is another branch of our work. Before, I talked about using what is similar in music to arrhythmias to capture and represent arrhythmia patterns; so that’s taking arrhythmia patterns and mapping it to music.

Now we also look at music and its structures and how these get transferred to people’s bodily response to music. We entrain to aspects of music when we listen to it. The entrainment is not absolute because our heart has to beat, and we need to breathe no matter what the music is doing. But we can see in our data that the entrainment is significantly greater than chance, so this tells us that we can actually detect patterns in people’s reactions to music.⁷

What we found in the data we’ve collected – and what is important here is we collect data while people are listening to music, this concurrent recording of music and physiology is important because it allows us to trace exactly what’s happening at the same time between the music and the body – so what we have learned is that a person’s reaction is based on their baseline autonomic profile. If someone predominantly has a high sympathetic tone and low parasympathetic, versus someone who is high parasympathetic and low sympathetic, they have a different response pattern profile.⁸ This tells us that if we want to have tailored music therapy, and this is now moving towards precision medicine, if we know someone’s innate autonomic tendencies, we can predict what their reactions to that music might be, and this can guide us to select or alter music to achieve the desired effect.

So that’s one outcome. The others I will highlight has to do with hypertension. Hypertension is important for many reasons: heart disease is still the number one killer, and hypertension is the number one risk factor for heart disease. Half of the people who have hypertension are not aware they have it, and of people who are aware and put on pharmacological treatment, within a year, 50% are no longer taking it. So, it is important to have early detection of hypertension and to find alternate treatments that people will adhere to. Music is a good candidate.

Most people think of music as a therapeutic, but what we have discovered is that it can also be a diagnostic tool. Most people think of music as being relaxing. And yes, some music is relaxing, but most music is not relaxing – it engages us, and when something engages us, it raises the heart rate, and it raises blood pressure, but only during the music; afterwards, they generally recover and even return to a lower level. So, it does have an effect that goes beyond the listening engagement session.

What we have found is that people with hypertension and people without hypertension react differently to music. People who have normal blood pressure, their heart rates and their blood pressures can go higher above their baseline in reaction to the music (and also come back down); they have a greater variability in their reaction to music.⁹ People with hypertension, noticeably, have less of a reaction to music, and what this means is that there’s a difference between people with and without hypertension.

In another study, we used artificial intelligence (AI) tools – deep learning – to predict if someone has hypertension based on their other physiological signals like their ECG or their respiration or both. What we find is that the accuracy is so-so when they’re not listening to music, but the accuracy of the diagnostic is much higher when they are listening to music. When they are engaged with music, their body is reacting to the music: people who have hypertension are less able to react, people who have normal blood pressure have a greater variability. There is a greater difference between them when there is music. So, if we capture these signals during music engagement, the AI model is better able to detect their hypertension, improving detection on all counts: accuracy, specificity, recall.¹⁰ Our initial tests showed an improvement of 10%, but more recently with a larger cohort, it was more like 25-30% improvement with music.

One might say, well, why would you want to measure blood pressure with music rather than just put on a cuff? The thing is, most people are not putting on a cuff every day to check their blood pressure, but on average, people listen to 2.6 hours of music per day, and if we are able to harness that, perhaps with wearables and with their music playlist, we could use that to give an early warning that they should go see their GP if they are found to most likely be hypertensive. This could happen, say, in a concert setting. You check into the concert, you have a lovely time, and when you check out, you have a reading of whether you should go see your GP about raised blood pressure. So, this is a great way to reach thousands of people quickly. And it doesn’t have to be classical music; it could be any kind of music. What we are evaluating here are the acoustic features of the music like loudness and tempo. Such expressive features are universal across all musics. We can use these measures regardless of what kind of music people are listening to.