The sounds of your heart beating and your lungs filling and exhaling are something everyone is familiar with.  Your heartbeat is the sound of your heart valves opening and closing.  In between those familiar sounds is sound energy given off by the of blood rushing through your heart.  In most cases, these sounds can represent normal healthy operation; where valves close properly and at the right time, and blood flows in the right direction.  Abnormal functioning of the heart will produce abnormal sounds, for instance valves closing  insufficiently and  with irregular rhythm, blood moving in the wrong direction or wrong velocity.  There are dozens of conditions that can be detected by analyzing the kinetic mechanical energy radiating from the heart as is functions.  These sound waves carry with them a vast amount  diagnostic information about the structure and condition of the heart, but separating these sounds from the background is tricky.

The most medically interesting sounds are also some of the hardest to isolate and hear. Like trying to separate out a single wave in an ocean. Important symptoms, such as gallop tones S3 and S4, mitral stenosis murmur, Still’s murmur and Austin Flint murmur exist and extremely low-frequency which are thousands of times harder for humans to hear than everyday sounds we are used to. Some diseases like infectious endocarditis can be detected at very early stages through sound, but it requires that we detect by the presence of a (very low power) quiet cardiac noise.

The use of machine learning is going to change things dramatically. It’s possible to train AI to detect and assess these sounds – especially those that exist on the edge of human hearing ability. This will significantly augment the diagnostic capacity, early detection, and screening available to people. Of course the real work today is in making this audio data machine readable. Bioacoustics is what they call unstructured data. It needs to be annotated and organized and its key features identified well before AI can work. If you want the AI to work well (and we do) you need a huge amount of this data. Daunting as it may seem, it’s an enormously worthy pursuit and will lay the foundation for years of innovation in the detection of dozens of diseases of the heart. lungs and possibly other organs.

Left – A wave off the coast of New Zealand. Right – a spectrographic visualization of a cardiac murmur as a result of aortic regurgitation