Artificial Intelligence May Drive More Personalized Treatment Protocols
Q: How do you define AI in the spectrum of technology? (And how can it work in the healthcare system?)
John Edwards, vice president, Healthcare Solutions Consulting, SoftServe: We talk about artificial intelligence (AI). It’s still a pipe dream that a robot would be thinking and feeling and being able to replace what a human brain does. But what AI was originally intended to do was to automate some set of activities that a person would typically do that you can create rules using computers to be able to replace that and do it more efficiently. And so, it is evolved to be a set of techniques that have their roots in statistics.
I actually have a master’s in statistics from Ohio State [University]. And we studied the underlying methods of how you would create these types of learning capabilities, and to be able to derive insights from them. And it was a different type of statistics than probably what you may have studied if you if you took a general course in college, because it became more theoretical. As big data became more real—that we could process more data and we had more data to be able to have— research capabilities with different techniques of AI began to evolve. And really, we often talk about the underlying learning models that are part of AI, whether they’re supervised or unsupervised learning models. Sometimes people get frustrated with AI, because the computer sees patterns and connections between data through these learning models that are hard to explain.
Traditional regression models always had people making decisions about variables that report and trying to predict with some type of linear model, you know, a progression through a line. You know, AI goes well beyond a linear prediction of capabilities is truly multivariate multiple things happening at the same time. And using these new statistical techniques to derive new connections between data that one would hope to have predictable capabilities. Now, the hope is measured with scores like what you may have had that in your statistics course of alpha scores and reliability. They’re called something different, but they really are used to test how reliable the prediction coming from the model is.
And so it’s a whole new area of science that has really hit on every industry, and healthcare specifically. It’s been a focus of a lot of researchers, a lot of great papers written out there and published around very specific examples of applying this technology in the space of healthcare. And it’s demonstrated again that it can work that it can create predictable results.
We’ve seen, in the last few years, a real explosion of the using of this information— both to drive drug discovery opportunities but also other places in the lifecycle of the pharmaceutical companies— from better recruitment of patients for clinical trials to better understanding of what happens with your medication after it’s released in the real world. And all of those use this new set of underlying techniques called AI.
I believe I saw an estimate recently that suggested that it was almost a $30 billion industry last year, of AI and healthcare. So it’s not an insignificant amount of work or activity that’s occurring. But as is in with every new technology, it’s often siloed and still not reaching its optimal potential of what it could do. But it’s absolutely a growing trend within healthcare, with pharma being a little bit of head ahead of how providers or payers are using that data. But providers are catching up quickly with thinking about how to use you know, these techniques to be able to create better services for themselves. But I go on.
Q: Can you elaborate on predictive modeling as it relates to healthcare?
John Edwards: There’s been a lot of studies that have been published about the value of using AI to predict how often a woman should get a mammogram. Today there are standards that are applied for preventative medicine across the board. If you’re a certain age, you should get them in a certain frequency. If you have a family history, maybe it’s a different frequency. If you’ve had some situations where you’ve had cancer, you’re in a different pattern but still it’s not precision medicine— it’s not taking everything that we know about that woman, giving her great advice about how often she should be going through the procedure of a mammogram and exposing herself to radiation through the technique that’s used, and then receiving advice about her care. We can do a better job.
If we create a model that predicts the likelihood beyond just the age of a woman—about whether they should be getting a breast cancer screening— by taking other data that’s available about the women that have received screenings [or] those that have ended up with cancer diagnosis is a different stages, we can create a more complete model that would allow us to take available data about the whole person and create a protocol that makes sense for that individual, about how frequently they should be getting mammograms… moving on from “everybody should get one at a certain age” to “get one this often”.
And because of your risk factors that have been predicted—based upon the data and the body of growing evidence we have by looking across lots of people in the real world and what happened when we apply this diagnostic procedure to this real problem—you can identify people that should be getting more or less frequently than what general protocols give us.
Predictive models have to be trained on a set of data. And if the data is narrow, it’s difficult to know whether the prediction that’s coming from the AI model makes sense. You know, only about 3% of the population participate in clinical trials. And by its nature clinical trials have a small group of people for the drugs [that] have been tested. And yet, when we put drugs out into the market, they get treated by everybody right there. The whole world starts to use the drug. But only 3% of that group had characteristics that led them to allow themselves to be part of the clinical trial.
Real world evidence that can come from examining this data and creating AI models driven from the data allows us to use more data from a broader set of people, perhaps more representative of all of us, and create better advice about what’s happening with diagnostic procedures—and their use or treatment protocols can be used to develop a greater expectation about how the treatments and the diagnostic procedures that we were using bring about results that makes sense for the individuals.