A Physician’s Perspective: How Machine Learning Will Improve Medical Care
The ASCVD risk calculator is a bear of a mathematical formula. I know this all too well. This fall, I spent several long nights building it into the code our Virta Health physicians use to analyze our patients lab results automatically. It requires seven specific input variables — all converted into their natural logs — and eight complex multivariable interaction terms…and all that just makes up the first part of the formula. There is still a subtle, difficult probability calculation needed to obtain the final risk estimate. Even then, to make a clinical decision, that estimate, along with even more patient characteristics, needs to be plotted on a Byzantine flow chart to determine, finally, a clinical recommendation for the patient.
Doing that process with a pencil and log tables would have taken a good hour per patient. A scientific calculator alone would bring it down to 15 minutes or so. Fortunately, there are web-based calculators that we used until recently that make it faster still, but the simple tasking of digging up and inputting each data point still consumed a full 10 minutes per patient. Now that we have the whole thing built into our software, we can analyze a patient in milliseconds.
Compared to what physicians could be doing with machine learning, all of this is child’s play.
Complex as that risk calculator seems to a human physician, it is an abacus next to the artificial intelligence (“AI”, although “deep learning” is the more precise term) algorithms behind today’s voice assistants and self-driving cars.
Remember the Roomba(R)? The adorable, if slightly ominous, crawling saucer robot that vacuums your floors automatically? It first came out in 2002, so at 17 years old, it’s hardly cutting edge technology today. Next to the Roomba, however, the ASCVD risk calculator is as sophisticated as June Cleaver’s vacuum cleaner on the original Leave It to Beaver (off the air since 1963).
Thus, it is all the more disheartening that the ASCVD risk calculator is the cutting edge of calculations in modern medicine. Most of the time, we physicians are nowhere near the level of a 1963 vacuum cleaner. On the Roomba-scale, we are somewhere between broom and picking-things-off-the-floor-with-our-hands.
I am looking forward to the coming transformation.
Combining data in complex and subtle ways to determine the best course of action is not the only thing we do in medicine, but it sure is important. Deep learning is great at exactly that and is far, far more powerful than the mathematical and statistical methods behind things like the ASCVD risk calculator. The sooner our field learns to embrace this and learn to apply it wisely, the sooner we will be able to deliver far better care for our patients.
Peruse a few top tier medical journals, though, and you would never guess this transformation is coming. Though we may use Alexa, Google Maps or even the Roomba everyday, sadly few in our profession understand deep learning well enough to apply it in our work. I hope we will learn soon. The ASCVD risk calculator and formulas like it are terribly constrained in how they can describe the world. Embracing deep learning formulas will open our profession up to far more powerful means of analyzing data and predicting the best treatments for our patients. Innovators in medicine, including our team here at Virta, are already exploring this. The sooner our field — academic centers, training programs and journal editors — catches up, the better so that we can collaborate and discover the best possible solutions as a community.
Powerful and inevitable as it is, deep learning is no panacea. The judgment, discretion and compassion of our human physicians and researchers will be irreplaceable for a long time to come. My vacuum cleaner is definitely not washing my dishes or making my coffee yet, but it sure is a lot better than picking dust up with my hands.
