Given the volume and depth of data available in benefits plans, as well as the complexity of problems to be solved, machine learning and broader artificial intelligence are unquestionably the future of plan and member health management.
That said, there are some important challenges for plan sponsors to consider as they look at these new management tools. The key is to initially focus on using machine learning to optimize the plan before wading into applications focused on optimizing individual member health.
Due to privacy concerns, it’s still early days for using machine learning to identify individuals at risk. While the technology and capability exist, areas such as consent and appropriate boundaries in interacting with individuals need to be explored more deeply before its rolled out extensively.
Read: Government looks to AI to speed up CPP disability benefits
In the meantime, some machine learning uses plan sponsors can leverage to better manage their plans include:
- Predictive analytics: Quantifying where a plan’s spending will be in 12-, 24- and 36-months’ time to facilitate strategic planning and plan design review;
- Valuing trusts and post-employment benefit liabilities: Using plan-specific discount rates in the near term that factor in the current plan design, demographic and disease state profile of a plan’s population and impact of new therapies and disease progression moving forward. Traditional actuarial assumptions about growth of health benefits costs in the near term may not be as relevant for certain plans;
- Population health: Quantifying the disease state profile of a plan’s population and changes over the next 36-60 months to prioritize investments in health and wellness, and establish baseline metrics that can be used to measure the broader impact of chosen initiatives moving forward;
- Disability management: Using integrated data to understand real drivers of disability and proactively identify risk factors for disability claims using enriched data in order to optimize disability management;
- Risk management optimization; and
- Management of complex specialty drug claims.
Let’s take a step back for context. AI is a general term for hardware or software that exhibit behaviour that appears intelligent. A subset of AI, machine learning is focused on algorithms and statistical models that computers use to perform a specific task, without explicit instruction, by relying instead on patterns and inference.
Read: Using data analytics, AI technology to curb benefits fraud
The growth in the data available to plan sponsors and industry stakeholders, as well as the consequences of suboptimal plan management, requires more sophisticated tools for optimizing this management. In 2019 and beyond, large plans can’t be managed using pivot tables and attractive graphs and dashboards based on high-level data.
All of AI’s different subsets are focused on the general problems of simulating or creating intelligence. Some of the common subsets include:
- Reasoning and problem solving;
- Natural language processing;
- Perception;
- Motion and manipulation;
- Social and general intelligence; and
- Learning, the major area of opportunity in managing benefits plans.
There are many reasons why machine learning isn’t more prevalent in benefits management today. These include: it’s profoundly complicated; it requires a robust analytics infrastructure that includes meaningful data enrichment; the required data sits in silos (often with multiple vendors); and most important of all, the base quality of the data is poor. It requires extensive cleaning and augmentation.
In addition, health benefits management using machine learning requires both domain expertise in health combined with data science expertise.
Read: New digital HR assistant uses AI to answer employee questions
The current AI focus in benefits is on natural language processing and related AI applications, like virtual assistants. The most advanced example in the marketplace today is Sun Life’s Ella, which is referred to as an interactive digital coach. It’s focused on member engagement and experience, as well as the effective use of benefits by members. There’s absolutely a need for this kind of member-centric innovation, and it’s likely to be the face of member contact in the years ahead for every benefits plan provider.
However, plan-centric innovation that’s focused on solving the challenges of managing benefits plans and responsibly containing costs requires subsets of AI that are much different than virtual assistants based in natural language processing. Plan management requires machine learning that uses data cleaning and validation, integration and enrichment.
Virtual assistants that focus on the member are excellent, and there’s huge potential for these innovations, but a plan needs to focus on leveraging large quantities of enriched data and machine learning to optimize design and claims management. Essentially, there’s a need for both member- and plan-focused AI. Today, most of the focus is on the member, but that’s changing.
The challenge for the broader industry, structurally, is that a machine learning infrastructure can’t be put on top of a basic reporting platform. There’s a layer missing. Reporting is simply organizing data into summaries in order to monitor how plans are performing. The vast majority of stakeholders in group insurance and employee benefits are stuck in this domain today because of the lack of meaningful investments in raw data cleaning, integration and enrichment at the claimant-level using analytics.
Read: Manulife rolls out group benefits option through Amazon’s Alexa service
Machine learning needs to sit on an analytics platform. This widespread lack of analytics infrastructures, along with a lack of industry-wide data standards, will be the most acute challenges to the broader use of machine learning in benefits plan management.
Plan sponsors (and other stakeholders) that aren’t interested in waiting for the broader industry to support machine learning should consider the following challenges to overcome so they can responsibly leverage the wealth of data they sit on without drawing erroneous conclusions.
- Data quality is the biggest barrier to successful machine learning in employee benefits. Without clean data — and robust and accurate enrichment of that data — machine learning will fail.
- There’s a need for a common classification system to properly integrate and enrich raw data from disparate benefits lines and/or sources at an encrypted certificate level.
- Lack of domain expertise. Health benefits are profoundly complicated, so guiding the data science behind the scene requires domain expertise. Data scientists can’t tell you what disease state a claimant is treating in claims data with a specific drug that can have four or five different uses. If someone can’t determine the specific disease state profile of a population of claimants, how can they meaningfully complete predictive analytics? This is where data enrichment and domain expertise are critical, because they provide the specific jet fuel that data scientists need for their supersonic aircraft, and why a vendor can’t simply attach machine learning tools to a reporting platform.
Read: Does artificial intelligence have a place in human resources?
There’s enormous possibility here, and plan sponsors successfully using machine learning today are testament to that. However, with big opportunities comes equal risks. Plan sponsors and industry stakeholders need to ensure they have strategies to overcome challenges in this space to ensure they’re making the best use of the robust data assets at their disposal.