Anthony Liu, Head of Data Science at the American Medical Association, discussed what prescriptive analytics is and does and how to develop it using AI.
At the outset of his keynote presentation at the 2017 Leadership in Big Data and Analytics Forum, held on December 5 in Chicago, Liu noted that healthcare industry spending exceeded $3.2 trillion a year in 2015, equivalent to $9,990 per person (source: CMS). This spending increases about 5% annually, he said. “What can we do to lower this cost? There’s a need to increase efficiency and effectiveness of health information management, including physician information.”
Liu outlined some of the specific, data-science practices that the AMA provides:
• Comprehensive and accurate information about physicians in the U.S.
• Effective engagement among physicians, patients, payers, pharma companies, hospitals, and other parties.
• Data support for decision making and policy recommendations.
“Physician data include critical pieces of information that continue to evolve and need to be updated, including contact information, credentialing, specialty, etc. Additional insights are needed about behavioral patterns of physicians,” said Liu. For example:
• Are physicians in some states more mobile than in others?
• How many physicians change their specialties after medical school?
• Do some physicians conduct more innovative treatments than others?
“Physician data include critical pieces of information that continue to evolve and need to be updated, including contact information, credentialing, specialty, etc.”
“Our data-science group, the AMA Health Solutions team, comprises people with hard skills—diversified skillsets such as computer science, statistics, and information systems—and soft skills—diversified experience such as within the AMA, health insurance, pharmacy benefit management, financial services, food services, and retail arenas,” he explained.
To deliver the best support to its stakeholders, Liu thinks in terms of three paradigms or patterns for delivering analytics:
• Descriptive analytics, which reflects the historic and current states of the business. “Most companies are in this generation of analytics.”
• Predictive analytics, which leverages historic data to predict future behavior of the business and customers. “Some companies are here.”
• Prescriptive analytics, which prescribes solutions to meet business needs and provide consultation and guidance to business operations. “In this third-generation analytics, the data-science team participates in decision making by recommending options.”
For the second half of his presentation, Liu talked about the use of AI as it relates to data about physicians. His team wanted to gather behavioral information—such as, what is a cardiologist?—on the estimated 1.2 million physicians practicing in the U.S. “We don’t have the budget to survey the estimated 30,000 cardiologists out there. AI can help with this.”
“What is a cardiologist? We don’t have the budget to survey the estimated 30,000 cardiologists out there. AI can help with this.”
Liu came up with the following questions to build the foundation of AI for physicians:
• How do we segment physicians in the same specialty?
• One cardiologist treats 20 patients a year, while another treats 200. Do we segment them the same way?
• Two cardiologists both treat 200 patients a year but perform different procedures. Do we segment them the same way?
• An internal medicine physician performs a certain number of cardiology-specific procedures. Do we segment him/her as a cardiologist?
“To answer these questions, we gather data from medical claims and other sources. From this data, we build machine-learning programs to create predictive models that answer questions like, ‘What is a family doctor? What is a cardiologist? What is a surgeon? What is a vascular surgeon?’ Different models generate different results, but we [humans] can tell which ones have the better results and provide feedback into this algorithm. Eventually, the algorithm is able to select which one is better,” he explained. “With the answers to these questions, we can use a machine-learning program to assign scores for all cardiologists—high, medium, or low. Interestingly, we found that some family doctors earned a medium cardiology AI score, so we think of those family doctors as the same as a cardiologist.”
“Different models generate different results, but we [humans] can tell which ones have the better results and provide feedback into this algorithm.”
In summary, Liu offered the following observations about data-science teams:
• Building hard skills can lead to a good team, but it takes the right mindset to build a great team.
• No single skill is sufficient to build a strong data-science team. It takes a village of diversified backgrounds and experiences to build a great data-science team.
• Building the right predictive models brings better business results.
• Starting from the end goal, striving to provide prescriptive analytics brings great value to the business.