The Case for Machine Learning in Prescribing

If you ever think your doctor has it easy, consider the following.  A patient comes in and is diagnosed with high blood pressure, high cholesterol, and diabetes.  Unfortunately a common trio that often travel together.  If the patient is on a statin for cholesterol, insulin and one other medication for diabetes, and needs three mediations for hypertension (a diuretic, ace inhibitor, and beta blocker), the physician may have quite the task in front of them.

In this all too common situation, there are 1,329,442,410,240 possible combinations of treatments.  That’s over 1.3 trillion. And this is not even considering other common disorders such as arthritis and depression.

Its also not considering that there are often a brand name and generic version of the drug from which to choose.  Six different medications may seem high, however, 20% of Americans report taking 5 or more drugs per day according the Mayo Clinic.

This astronomical figure is well beyond any person’s ability to manage.  Most would get overwhelmed with hundreds of choices, making trillions almost comical.  It would be funnier if lives did not depend on these decisions, however.  One way that people deal with any enormous space of choices, is by using heuristics or biases.  Now, biases are usually used in a negative context.  When someone is “biased”, we immediately think of them as being unfair in some way, such as the hometown referee or judge who has been bribed in some way. As human beings, bias is unavoidable, but, it is also sometimes necessary.  Without some sort of bias, we would be crippled with choices that we have no way to evaluate analytically in any real time fashion.  According to a Cornell study, we make 225+ decisions per day based on food alone.  Bias allows us to quickly choose something, removing the time necessary to evaluate every option.  For people living in a dynamic and rapidly changing environment, these biases can even be argued to be vital to our survival.  We don’t want to evaluate every possible escape option in immediately life threatening situations (the anecdotal animal attack), if we did so, we probably would not survive.  For everyday life, not only are biases part of us as human beings, they can actually be positive, if not necessary part of our lives.  Prescribing medications is a different animal, if you will.  There is no doubt that different medications will help or potentially even hurt patients in different ways.  It is also known that everybody is different, our genders are different, our genetics are different, our financial resources are different, and the other medications we are currently on is often different as well.

When prescribing medications, if possible, we would want to make the best possible decision given the options and information we have at hand, in this removing any bias from the equation.

It is unreasonable to ask physicians to handle this task alone.  One exciting option is to use the data being collected on patients in their electronic health record (EHR) and how they responded to specific treatments in order to guide physician prescribing.  in other words, we currently have access to data that tells us how people of different genders, weights, ages and so on actually responded to specific treatments.  Why not use this information to help inform doctors as to what treatment path is most likely to succeed?  Indeed, this is exactly the type of problem that the tools of predictive analytics, data mining, and machine learning excel at handling.  A skeptical point of view could be that this field is just too complex to use these approaches.  However, this complexity is precisely why you need these approaches!  Using predictive models to augment human decision making and information has a large history of success stories.  Some of these models have become so integrated into our lives, it’s hard to imagine people ever doing better than their predictive model counterparts.  Would you trust your meteorologist at all if they didn’t have computer models to predict the weather?  Would you rather have your local stock trader or a high speed trading algorithm?

Interestingly, researchers at Indiana University have shown that using predictive models for diagnosis can lead to 50% reduction in overall costs and 40% improvement in patient outcomes… (

Doctors are amazing at what they do, but it’s time to give them a hand with the task of prescribing medications.