MedCity Influencers

Why machine learning is a lot like Sriracha sauce

As exciting as machine learning's potential is, there is a temptation to apply it way too generously on everything healthcare-related. But there are some important limitations to consider for this technology.

I like to think of machine learning as Sriracha sauce.  When you add it to something, that something immediately becomes more interesting and more appealing. It’s also kind of addictive (as Sriracha lovers will tell you) and there is a strong temptation to put it on everything.

When it comes to healthcare innovation, this certainly appears to be the mantra.  Today you cannot be a successful health innovator without putting “machine learning” in your pitch deck.

While the touted ROI and disruptive nature of machine learning differs across companies, the underlying promise is the same. Machine learning algorithms are going to take the multitude of data we generate from the healthcare system, analyze it and allow us to see things we could have never seen before.  Everything from better diagnostic tools to ER logistics will be transformed.  But, perhaps the most touted and highest value prize in healthcare is the promise that machine learning will be able to help us predict human behavior.

It’s tempting to believe this. In a world where computers can beat GO champions, algorithms can predict whether humans will shake hands or hug and even drive a man in life threatening condition to the hospital you could be forgiven for thinking that their potential is limitless.

There is, however, one underlying fundamental construct here that makes all of these seemingly disparate applications of machine learning possible. Structure. In each case outlined above, the algorithm was able to generate a reliable response because of the structure of the environment itself. Whether it be the rules of GO, the artificial environment of TV entertainment (the algorithm used The Office as source and test material), or the consistent road infrastructure and traffic laws, the algorithm was able to perform because random variation was limited. In effect, these accomplishments were achieved in a controlled system — one where the sources of variability and error are known and can be reasonably and mathematically controlled.

And herein lies the issue for machine learning and human behavior in healthcare: the system is highly complex and humans are inherently irrational beings.

That is why machine learning is not healthcare’s messiah. The largest problems in healthcare stem from inherently irrational human behavior. For example, the Big Three chronic diseases (hypertension, high cholesterol and Type II diabetes) consume by far the largest proportion of healthcare dollars. Ongoing medical research has given us excellent disease prevention strategies and highly effective treatments for affected individuals. However, without fail most people have difficulty managing their conditions.  Treatment plans are not followed, diets fall off and exercise regimes are put off until “after the holidays”.

This exemplifies a problem that the healthcare system calls “patient engagement,” a seemingly laissez-faire approach people take to managing their own health.  The current medical rhetoric shifts the blame for this lack of engagement onto the patient themselves (while I don’t personally agree with this stance that is a topic for a future discussion) and suggests that machine learning is the solution to finally understand why patients don’t do as they’re told.  If they could connect enough data points and have large enough computers, they could finally understand what to do about it.

Unfortunately, their intentions are well placed but their perspective is flawed. With all the possible factors at play with patient engagement (motivation, cost, disease knowledge, medication literacy, just to name a few), we would need petabytes of data from millions of people before we got close enough to a model that would have even a slight chance of predicting behavior.  Furthermore, this approach creates a model by design that regresses to the average and it has been well established that there is no average person. People are individuals and our empirical analysis of behavior can only tell us about the population’s probability of action.  Add in the high degree of variance in causal factors and the human irrational decision-making engine and the limits of this approach become clear.

I believe we would benefit from a different approach. One that embraces the irrationality and uniqueness of the individual but also leverages the fact that a person will tend to behave the same in similar situations.  What if servers and algorithms were not interested in 1 model to predict the action of 100 000 people, but 100 000 models, to predict the actions of each individual person. What if your data was analyzed over time so that machine learning could give you insights into your own self? In this way, information would be completely personalized. You would begin to know your own individual barriers to a healthy lifestyle and therefore would have absolute power to change them.  This could be truly empowering for people living with these conditions as well as health professionals on the care team who could target interventions accordingly.

I’m not advocating for an abandonment of big data-driven machine learning projects. They will continue to be useful in helping us understand the mechanisms to disease. But, if we could shift the patient engagement focus from a population to a personal view, we’d all be better off for it.

Photo: bhofack2, Getty Images


Avatar photo
Avatar photo

Bill Simpson

Bill Simpson has an extensive background in psychology, mental health and clinical data analytics. He began academic life as an undergraduate researcher before becoming a research assistant and finally completing his Ph.D in Neurosciences at McMaster University. He has managed data collection and analysis for industry sponsored clinical trials and large international cohort studies. He has 17 peer-reviewed publications covering a range of topics and speaks regularly at Canadian and other international conferences on the digital health revolution, machine learning and apps for mental health. He is currently the Director of Data Science at MEMOTEXT where he spends most of his time working on new algorithms and analytics strategies. He is also a Senior Research Associate at McMaster University where he works on clinical trials for ADHD, understanding if internet addiction is a thing and looking at how marijuana can impact mental health.

This post appears through the MedCity Influencers program. Anyone can publish their perspective on business and innovation in healthcare on MedCity News through MedCity Influencers. Click here to find out how.

Shares0
Shares0