Artificial Intelligence and the Human Heart: Of Mice and Men and Machines

Patterns we loved and the shadows they covered the ground with

Tapestries, mystical, faint in the breathless air.

- F. Scott Fitzgerald (This Side of Paradise)

For some time, it was thought the combination of the human eye and brain could determine the context of images – i.e., what they represent - presented to it in as little as 100 millisecond increments. In other words, if that was all we did, we would be able to determine what was being represented in 864,000 images every 24 hours. However, as ridiculously fast as this seems - it appears to be an underestimation. Investigators at MIT have recently determined we can do this every 13 milliseconds - so move that number up to 6,646,237 or so images per day. However, it isn’t the ability to quickly recognize context, as it is our ability to perceive patterns that is truly remarkable. No other living thing can process, combine and interpret them as we can, or imagine new patterns based on those seen before and share these with others - the latter in part responsible for our singular ability to use language and create highly advanced and far-flung social structures. Evolution of certain regions of the human brain and more sophisticated information processing are the substrate for our “superior pattern processing” abilities.

When a field mouse sees a tiger, it recognizes the immediate context of danger, and runs. When a bushman living in the wild sees a tiger – he may run as well, but not until he has processed the “pattern” in front of him - in no more than milliseconds… Does the tiger see me? Am I downwind of the tiger? Is it the time of day when tigers are likely to be hungry? Have I encountered this particular animal before? Are my children nearby? Unbounded imagination, improbable social skill and insatiable curiosity are the abilities that differentiate Homo Sapiens, and some neuroscientists believe most, if not all of these characteristics emanate from the superior pattern processing capabilities of the human brain.

Healthcare is nothing if not our most challenging pattern recognition exercise – each patient bringing an external appearance and internal pathology, as well as thoughts and words to the provider - organized into their own patterns to communicate concerns, complaints and symptoms which then allow us to categorize (i.e., diagnose) and treat disease. We create even more patterns for processing when we create images of their internal organs and then extract information from their bodily fluids and tissues. 

Up until recently, we have traditionally used radiographic images in medicine more like the field mouse. We have focused largely on rapidly determining the context in front of us… Is the left ventricle thickened? Is there a tumor in the cortex? Is there a stone in the gallbladder? What we have not done is discern the more nuanced patterns we see (or we actually may not see) in images, and combine this information with other data - as the human brain does so effortlessly - to make better predictions about disease, and more accurate decisions regarding treatment and diagnosis, but we are on the threshold of doing so - and in a big way.    

Depending on the techniques used and number of views obtained, there are between a few hundred megabytes and a handful of gigabytes of data in an MRI study of the heart, but it is the the patterns that these “data” represent – the spectrum of what is light and dark and these contours and shapes these shades create… what is moving in the image and what is not… etc., that are interesting. When combined with other information - other patterns - we have always collected (and other information we are just now developing the ability to collect - via advances areas such as genomics and physiologic monitoring), the science of complexity tells us insights will come to light and more comprehensive patterns will be recognized we often would not, or even could not anticipate or expect. 

A recent study published in the European Heart Journal provides a glimpse of what is to come by the applying intelligent computing in cardiovascular diagnosis and treatment.  More than 10,000 patients were followed for five years following coronary computed tomography (CT) angiogram and sixty-nine data variables were evaluated - 25 related to patient clinical information (age, sex, standard cardiovascular risk factors, etc.), and 44 gathered from the CT angiogram x-ray findings (degree of coronary vessel narrowing, amount of calcium present in the vessel wall, etc.). When machine learning was used to combine and process these data, predictions regarding the patients’ prognosis for risk of future cardiac events and death from cardiac or other causes was significantly more accurate than when using CT angiogram readings or traditional cardiac risk scores alone.  This is literally the tip of the iceberg – imagine doing this study again, but this time adding - for the sake of argument - not only CT scan but also cardiac ultrasound data, genomic analysis and ambulatory home cardiac monitoring data (rate, rhythm and electrocardiogram patterns)? Beyond this, what if we did this type of study on 10 million patients rather than 10,000? Would the results be more impressive, and perhaps even surprising? Would it both better predict events and prognosis, and also help us understand how to tailor therapy for each individual? There is no way to know for certain without doing the experiment, but my very strong assumption is an unequivocal “yes”. When you approach artificial intelligence in this way, what you are talking about is augmenting human knowledge and decision-making rather than replacing human involvement. Interestingly, if you believe what makes us unique as humans is our ability to process patterns more quickly and more comprehensively than all other living things – what we are actually trying to do here is simply teach our machines how to do the same thing - what we do - on a larger scale.

The more sophisticated evaluation of images in addition to clinical and other data could be - and will be - impactful for cardiac disease in other areas as well, such as identifying patients at risk for the development of congestive heart failure. One could envision screening them with a portable handheld ultrasound device and then using genomic, clinical and other data to determine who should be followed more closely, or even treated during the formative stages of disease. While we talk a great deal about “precision medicine”, and “individualized medicine”, for cardiovascular disease these approaches will be the true price of admission in the future. We will better understand which drugs are likely to be effective in which patients, rather than using a few historical data points from studies involving a few hundred patients. Perhaps we will also put to rest the controversies that have roiled over the past few decades regarding whom should undergo surgical vs. percutaneous coronary interventions for coronary vascular disease - a perspective at times perhaps more colored by whether one held a scalpel or catheter in her hand while training, rather than scientific evidence.

As we approach World Heart Day, we remember the heart is a mechanical as well as biological organ, so discerning both structure and function via the creation and interpretation of images we can see has obviously revolutionized cardiovascular care. However, we now know that simply understanding the immediate context of what we see in these images is not sufficient. We place patients in our machines to lift the heavy cover of the body's codex and then study the pages within - seeking representations of illness. These representations will be even more meaningful in the future. A picture may be worth a thousand words, but the near future value to humanity of a picture combined intelligently with other data in the pursuit of better cardiac disease outcomes has not yet been calculated.