3 Ways Payers Can Employ Machine Learning, Advanced Analytics
Machine learning and advanced analytics have various uses in the health insurance industry, including condensing medical records, hypertargeting, and supporting risk adjustment.
Source: Getty Images
- MVP Health Care (MVP) has been piloting machine learning and advanced analytics solutions and, in the process, explored three key uses for these types of tools.
Machine learning falls under the umbrella of artificial intelligence tools. It consists of alorithms that rely on historical data to create future-oriented outputs. Advanced analytics employs machine learning algorithms and other models to predict trends.
Patrick Roohan, vice president of quality and clinical analytics at MVP, highlighted three primary uses for machine learning and advanced analytics that MVP has explored:
- Condensing medical record data
- Hypertargeting
- Assessing risk adjustment in Medicare Advantage
While this is by no means a comprehensive list of these tools’ capabilities, increased efficiency in these areas can impact various payer processes.
Uses of machine learning, advanced analytics
Condensing medical record data
Machine learning’s basic role in healthcare is to condense information. When used effectively, the tool can free the provider and payer workforce to focus on interventions.
The digitization of medical records was critical to making healthcare processes more efficient. However, the transition to electronic health records (EHR) did not shrink the amount of information that a medical record might contain. An EHR can contain up to hundreds of pages’ worth of patient data.
“How do we find and glean out information that we think is important for a person that we have evidence they have diabetes and have hypertension? We're beginning to use that, both working with our IT team and our analytics team, to really target and try to figure out ways we can cut down a review of that record,” Roohan told HealthPayerIntelligence.
“This machine learning pilot we're doing is tremendously helping our nurse reviewers to try to minimize the time it takes to find the information we're looking for,” Roohan added. “There is a vision that we spend less effort gathering medical data and information on people and actually do more with it.”
There is a vision that we spend less effort gathering medical data and information on people and actually do more with it.
When MVP applied machine learning and advanced analytics to the medical records of Medicare beneficiaries who do not appear to have reportable diagnoses, the payer identified that eight to ten percent of these individuals had a conditions that were overlooked.
Through machine learning, MVP condensed the EHR data and uncovered that patients had undiagnosed conditions. This information can then be used to hypertarget individuals for interventions and to perform more accurate risk adjustment.
“Even when there's no claim for a diagnosis, what we're finding is people do have more conditions,” Roohan said.
On the other hand, for members with chronic conditions, machine learning can efficiently condense information for predictive functions.
“Machine learning takes data from medical charts and searches for specific conditions that can inform us of a customer’s quality needs. It can also rule out certain words to identify potential conditions. As an example, if a patient had hypertension and diabetes in 2022, we can anticipate they will have it in 2023,” Roohan explained.
Hypertargeting
MVP used machine learning and advanced analytics to address care gaps through hypertargeting. Hypertargeting is a marketing term, but the principles of hypertargeting can apply to segmenting a population for the sake of health interventions just as effectively as they apply to targeted advertising.
First, MVP builds personas using a technique called clustering. The health plan leverages machine learning to organize members with shared characteristics into populations.
“To build valuable personas, MVP Health Care relies on data to group people with similar aspects and needs,” Roohan said. “Personas are a way to categorize customers in groups to improve their care.”
Then, MVP uses the personas to create hypertargeted interventions for members.
Roohan offered the example of hypertargeting women over 40 who need mammography. Out of the population of women over 40 who have not received a mammogram, MVP can develop three personas.
The first persona would be a woman who routinely gets her mammogram and does not need a reminder. This persona does not require a hypertargeted intervention.
The opposite, secondary persona would be a woman in the appropriate age range who refuses to get a mammogram. Many factors could contribute to her resistance, including that she may not believe in undergoing mammography. This persona is extremely difficult to engage, Roohan said.
The third persona may be the most likely to benefit from hypertargeted interventions. The woman who embodies the third persona may have an inconsistent history of mammography. Machine learning can inform MVP of the woman’s population based on her medical history. Once MVP has identified a member’s persona, the health plan can focus on assessing her barriers to accessing care.
To build valuable personas, MVP Health Care relies on data to group people with similar aspects and needs.
Importantly, hypertargeting for health insurance can leverage both healthcare data and social determinants of health data to inform its output.
For example, a woman over 40 who needs a mammography and who falls into the third persona might face social determinants of health challenges such as transportation barriers or lack of access to the appropriate facilities. If MVP has access to these social determinants of health data points, the payer can more effectively address the woman’s needs to ensure she has access to a mammogram.
Medicare Advantage risk adjustment
One of the primary motivations behind MVP’s expansion into machine learning was related to Medicare Advantage risk adjustment. The payer wanted to see its Medicare Advantage plans crest the five-star rating.
In 2021, the payer’s Medicare Advantage plans received four stars. The following year, the payer received a 4.5 star rating overall. For the 2023 plan year, its health maintenance organization-point of service (HMO-POS) and its HMO dual special needs plans (D-SNP) ranked five out of five stars, but its preferred provider organization (PPO) plans lagged with four out of five stars.
To achieve five-star quality across all of its Medicare Advantage plans, MVP employed machine learning and advanced analytics and assessed its care gaps. Using this technique, MVP was able to evaluate its own performance.
Roohan indicated that CMS will have to account for machine learning in the future because the tool will impact Medicare Advantage risk adjustment and similar processes. He expected that by 2027 or 2028 EHR data would be the core data for evaluating healthcare—equaling and eventually surpassing claims data in five to ten years.
“Will this be the way to go? I think it has to be the way to go in the future to use this data at some level,” Roohan affirmed.
However, he also acknowledged the challenges that this could pose for federal agencies as they reimburse payers for claims.
“The problem is the more tools you have, the more diagnoses you can find,” Roohan said. “They’ve got to define the rules a little bit better, obviously, and audit a little bit better.”
As payers like MVP continue to expand their use of machine learning and advanced analytics, regulatory agencies and other healthcare stakeholders may have to adjust their processes as well.
Challenges of machine learning, advanced analytics
While machine learning and advanced analytics can improve payer processes, health plans may encounter barriers to using these tools. One of the major challenges is knowing who owns the data, particularly when leveraging a network for data access.
For example, State Health Information Network of New York (SHIN-NY) is a group of six health information exchanges that MVP and other stakeholders fund. MVP uses the network to share data statewide. This model has advantages—namely, it allows for better and broader data-sharing which can improve quality of care.
SHIN-NY has partnered with every hospital in the state of New York and more than 100,000 providers to allow electronic health data transfers with patient consent, according to the organization’s website. Millions of patients’ data live in health information networks that SHIN-NY connects.
However, having so many stakeholders involved—including EHRs, hospitals, providers, and the patients themselves—makes defining and maintaining proper ownership over the data a challenge. The state of New York gives permission to use the data. Patients give consent. EHR vendors also want to have a say in where the data goes.
“Who owns that data is a classic question, right?” Roohan said. “Ultimately the real answer, the best answer, is what we call the customer. The member should own that data, but it is complicated.”
If data ownership or another barrier gets in the way of data access, the repercussions for payer operations and patients’ quality of care are significant.
“Limited access to data from providers can inhibit our ability to predict patient needs,” Roohan explained. “With more data availability and access, we have more opportunities to positively influence care by designing our services to close the care gaps and meet customers’ needs.”
SHIN-NY’s impact exemplifies the influence of data access. New York eHealth Collaborative (NYeC), the nonprofit that runs SHIN-NY, reported that the network halved hospital readmissions in the state, decreased repeat imaging procedures by 35 percent, and reduced emergency department visits by 26 percent.
Without the appropriate data, machine learning models and advanced analytics models are disadvantaged. But when payers have access to data, machine learning and advanced analytics can support a variety of tasks, including condensing medical record data, hypertargeting, and Medicare Advantage risk adjustment.
