Healthcare may be among the most debated and controversial issues on the national agenda in the US, but one insurer is taking the initiative by utilizing predictive analytics and machine learning to help make patients' journeys more cost effective and efficient though the use of social network analysis.
Michael Xiao, divisional vice president of enterprise analytics at Blue Cross Blue Shield of IL, NM, MT, OK & TX, explains how social network analysis seeks out stronger relationships between physicians and patients, as well as how his team manages the swaths of data at their disposal.
Xiao, who will be speaking at the Predictive Analytics Innovation Summit in Chicago, also discusses the talent gap within the predictive analytics field and the growing use of AI in making vital business decisions, as well as offering his advice on setting up an effective analytics team.
Innovation Enterprise: During your upcoming presentation at the Predictive Analytics Innovation Summit in Chicago, you are planning to speak about social network analysis. Can you explain what this entails and how it can benefit the healthcare insurance sector?
Michael Xiao: The US healthcare industry is extremely complicated, with some patients visiting multiple doctors to have their conditions treated. Patients may see five, six, seven or more physicians during their treatment; often, it's not the primary care physician driving the overall cost, but who the primary care physician refers or recommends their patient to visit next. Even for something very common, such as a knee or hip replacement, patients have an initial appointment, which might cost $100 or $200, but everything that comes afterwards, including follow-up care, has a huge variance in the cost and that could range from $20,000 to $100,000.
Social network analysis seeks out the relationships between physicians, especially the relationships between primary care physicians and specialists, and the relationships between specialists and healthcare facilities. The results are tied to recommendations made during office visits and we use social networks to understand the relationship between the patient and their physician, and then to see how we steer the member (somebody covered by Blue Cross Blue Shield insurance) to go to a better specialist once they've visited their primary care physician.
Our members have an incentive to visit doctors that can deliver high-quality care at the lowest cost possible and we use our algorithms to help them to make the best choices.
IE: What challenges does the large amount of data you collect present and what tools do you use to analyze it?
MX: One of the biggest challenges is actually getting the infrastructure in place to manage the data. There are a lot of open source tools available and many solutions today are based on Hadoop, using distributed computing to manage and process the data. There are also technologies like Spark to run distributed SQL queries and distributed machine-learning models.
For this specific initiative we use Python and we structure the data in a way that establishes the social network relationships via metrics such as network centrality.
There's also the visualization side of things which I'll be talking about during the Summit. Through visualization, we can create graphics that help tremendously with stakeholder buy-in. If you give your stakeholders a bunch of numbers, most of their eyes will glaze over, but if you show an executive a really interesting visualization, it helps them to understand what it is that you're doing with the data.
IE: Are there any risks associated with using AI to make decisions that affect human healthcare needs?
MX: I think there are two dangers, with the first focused on defining what AI actually is. It's a term that gets thrown around loosely; when some people say AI, they mean machine learning and predictive analytics, but they are not thinking about what the lay person associates with AI, such as a robot perhaps taking their job.
The second danger is that an automated process making a lot of decisions is not fully vetted to ensure all of its calls are made correctly or ethically. There is a lot in the news about how automated machines incorporate the bias of humans, which we've seen with Facebook, Google and other large organizations' algorithms.
In a lot of models today people are taking a gazillion variables and throwing them into this giant machine, which they switch on and tell to "do this, do that". That could prove quite dangerous because, if you don’t understand that data or the lineage of that data, things could go terribly wrong. For example, take a variable that was OK when you built the model, but in which there's been a change in your systems or the economic environment; now it has technical problems that can make your model completely invalid and you're still running this machine behind the scenes where its making all these decisions. All of a sudden, it's gone from performing really well to performing terribly.
You have to make sure there is always someone who actually understands the data to monitor your models and ensure that they still make business sense.
Visit Innovation Enterprise's Predictive Analytics Innovation Summit in Chicago on October 30–31, 2018
IE: What are the main problems with adopting predictive analytics in the healthcare industry?
MX: There are essentially two types of predictive analytics. There is automated predictive analytics, where you have a model making very low-level decisions, and there's another form of predictive analytics where you have humans trying to make decisions based on information a model is providing.
In healthcare you have both types, but the human issues are a little different because you have to get a human to understand what it is that you're delivering to them; you have to simplify your solution and get a human to understand how to interpret the output and ultimately use it in a complicated system. For example, say your audience is a doctor or a nurse employed by your company – what is their workflow and how do they make decisions? How do you design it so it's seamless and integrates into their daily workflow? If it doesn't quite fit, then it leads to resistance to using what it is you've built. Starting from the endpoint and building backwards helps to avoid a lot of the issues that pop up.
IE: How much of an issue is talent retention within the predictive analytics field?
MX: I think it has more to do with management than individuals entering or moving around the industry than people think. You can't just hire people who are technically strong and ask them to lead a team – they also need to be a good manager who understands the strengths of their teams.
It's also not just about pay, although it can be an issue. While you have to be competitive, some think that if they pay more than everyone else, they can get their employees to stay around, but it's all part of a bigger equation.
You have to differentiate yourself, so you have to bring in talent that is interested in the types of problems you are solving, which other companies are not. If you can find that talent interested in your specific business problems, it can be a huge differentiator.
IE: Is there an evident talent gap within the predictive analytics field and is there a solution to the issue?
MX: It's a really interesting time for predictive analytics, machine learning and AI, as there are a lot of new undergraduate and Master's programs springing up focused on data science and analytics. That will lead to a relative oversupply of people at the entry level and tremendous undersupply of people at the senior level.
You need to have a long-term strategy for employing entry-level employees, starting at the undergraduate level, instead of just focusing only on PHD-level graduates. Applied data science isn't rocket science.
A lot of these people enter the industry and don't really know how to work in a big corporation or navigate their first real job, so you need to be able to manage that. You do need to have senior talent, but the type of senior talent you want are people that can also develop that junior talent, which will really help with filling the talent gap.
IE: Do you think we will see the widespread adoption of AI solutions within the healthcare insurance space in 2019?
MX: In terms of healthcare, in the US at least, we've always lagged a little behind, so we're still catching up.
However, we're starting to deploy something closer to "real AI" solutions. There's a been a lot of hype around AI, but I think it will start showing up more in 2019. I don't think we're going to fully reach the state where AI will be commonplace in healthcare, but I do think there will be more and more of it.
Moving forward, as with any other maturing field, the focus will be a little less on the technology and a little more on solving real business problems, which is all part of the technology's maturity curve.
IE: Finally, what advice would you give to someone setting up a new analytics team?
MX: Start by focusing on the business value of what you're doing, rather than the technology. There's a tendency for people who are rather new to this area to be dazzled by the coolest and latest toys. And, often, they think everyone else is as well. Start with the business proposition and work out the challenge you have, then consider the latest machine learning algorithms or data science advanced analytics tools (in some cases AI) that can solve your problem, and you'll soon be able to turn that $50m problem into a $10m dollar problem.
Michael Xiao, divisional vice president of enterprise analytics at Blue Cross Blue Shield of IL, NM, MT, OK & TX, will be speaking at Innovation Enterprise's Predictive Analytics Innovation Summit in Chicago on October 30–31, 2018.
Blue Cross Blue Shield of IL, NM, MT, OK & TX is an independent licensee of Blue Cross Blue Shield Association, a federation of 36 US health insurance organizations and companies providing health insurance to more than 106 million people.