This article is based on a presentation at IBA’s Spring 2023 Analytics Conference on Natural Language Processing and Generation by Greg Hayworth, Associate Vice President of Enterprise Data Science at Humana.
Greg Hayworth started his presentation at the Spring 2023 Analytics Conference on Natural Language Processing and Generation with a quote: “No amount of fancy math can salvage a project which is solving the wrong problem.”
Hayworth, the Associate Vice President of Enterprise Data Science at the health insurance company Humana, leads teams that use machine learning and artificial intelligence (AI) to build analytics products for Humana. In his presentation “Data Science on the Wild,” he spoke about his experience developing analytic solutions in a corporate environment, or “the wild,” and shared tips learned during his nearly 30 years of work in this area.
A lot of analytics projects “die on the vine” in the corporate world, Hayworth said, so he offered three principles for ensuring that these projects add value in a new marketplace, starting with solving the “right” problem. He then discussed how to talk about analytics projects as products and why it’s important to deliver the simplest method that actually works.
Defining the Problem
As a way to avoid solving the “wrong problem,” Hayworth offered the “CoNVO” model: Context, Need, Vision, and Outcome.
Gathering context involves defining the business problem that the organization wants to solve, big picture goals, how much time can be saved, and how much revenue is being generated.
“A lot of times, the people who you’re in contact with for your analytic projects are not the end users,” Hayworth said. “Make sure you’re asking the right questions to figure out who really uses the programs day-to-day, who’s going to be on the IT team that will implement this, and who’s going to train people once a solution is in place.”
The next question to ask, Hayworth said, is if the project can be formulated in terms of what information is needed. If an analytics project can answer who might respond to an advertising campaign, that’s an information need that can be filled, Hayworth said.
“If you struggle to formulate a project description in terms of information that’s needed, it might not be an analytics project,” Hayworth said. “Sometimes, killing a project before you spend a lot of time writing code and gathering data is a win in the business setting.”
Developing a vision includes describing what outcomes and implementation might look like for the project, Hayworth said. Drawing these marked conclusions in the first couple of meetings means that users can ensure the project meets a business need before time and money are spent on writing code or building models.
Outcome describes who’s going to be responsible for next steps of the project and how the creators will be evaluated on whether or not the project is doing what the business actually wants it to do. While this step might sound like common sense, Hayworth said, it’s essential to make sure that the people who will execute the reporting on a day-to-day basis can reproduce the required analytics feed and daily, weekly, and monthly reports.
“You’re going to be deploying data scientists to solve analytic problems,” Hayworth said. “These are some of the most highly trained, valuable people in some of your organizations, and you should still go through the checklist to make sure you’ve covered all the bases.”
Creating Products
While discussing creating “products,” Hayworth gave another warning: don’t let the work done on an analytics project die on a PowerPoint. With a lot of projects, Hayworth said, great work is done, models are built, and a presentation is given, and after leaving the boardroom, nothing happens.
“It might be the absolute best model in terms of model performance, but if I didn’t take those model result and put them into the hands of my users exactly when they’re going to make a decision, then I’m really, really losing a lot of the value of completing that analytic project,” Hayworth said. “We can bridge the gap between building models and delivering value through product thinking.”
Members of an analytics department should collaborate with software engineers from the beginning of a project to think about products that are actually able to be used, Hayworth said.
“Your analytics teams thinks, ‘how do I solve this problem at all?’” Hayworth said. “Your software engineers are thinking, ‘how do I get this to work a million times a day?’ They’re both useful mindsets, but you really need to think about it in a symbiotic way and make sure that your projects involve both the software engineers that are going to make this happen a million times a day, and the data scientists who are going to figure out new ways to do things at all.”
Keeping it Simple
Hayworth also offered a lesson for newly minted Masters of Business Administration (MBAs) and data scientists PhDs: credibility and trust don’t come from credentials, but from consistency and reliability.
“The key to consistently delivering results isn’t in how cool your algorithm is,” Hayworth said. “The key is to determine how people will measure whether or not this is delivering value and to measure the project in exactly that way, then use the least complicated method that actually delivers on that value.”
NLP in Action at Humana
Humana has used NLP to improve the process of reviewing the millions of faxes of medical records and health records that the company receives every year, Hayworth said, including reviewing past medical history to automatically highlight points of interest that nurses commonly care about.
Before a nurse does a clinical review, hundreds of pages of medical records can be scanned using NLP to identify answers to key questions that they are asked. Nurses can get the answers before they even get to the documents, and even view hyperlinks that go directly to the context of that key information.
“We use NLP to gradually add up capabilities, but deliver value all along the way,” Hayworth said. “Stacking capabilities is the kind of thing that we need to do with more and more of our analytic projects, whether they’re NLP or not, as we escape from the walls of academia into the wild.”
Leave a Reply