Next-generation predictive model uses less data to deliver more accurate, personalized and flexible results.
Data Science

Predicting high-cost claimants to control spend and improve outcomes

March 9, 2021
Next-generation predictive model uses less data and delivers more accurate, personalized and flexible results.
Principal Data Scientist


Six in ten adults have at least one chronic condition. Four in ten have two or more.

In 2019, 90 percent of the $3.8 trillion in U.S. annual healthcare spending was for people with chronic and mental health conditions. Six in 10 adults have at least one chronic disease. Four in 10 have two or more.

Despite years of investment and effort to reduce medical spend and control chronic disease prevalence and progression, a recent report from the CDC reveals chronic disease is still the leading cause of death, disability and cost. Interestingly, and perhaps counterintuitive given the host of chronic diseases, just a small percentage of the population incurs the lion’s share of costs.

Thankfully, health plans have access to innovative, continuously improving AI-driven tools at their disposal to accurately identify members most at risk, forecast future risk and intervene before health deteriorates and money is spent.

The Geneia Data Intelligence Lab

Improving health and financial outcomes is precisely the goal of the Geneia Data Intelligence Lab (GDI Lab). While one of the GDI Lab’s many established models is a high-cost claimant model that for years has been successful in helping clients identify high-cost members, the data scientists at the GDI Lab had a vision for something even better.

Unique in this new model is the ability to predict high cost with as few as two elements from claims data: date of service and cost.

The next generation: Doing more with less.

Accuracy and equity in predictive modeling typically require many different types of data. Unique in this new model is the ability to predict high cost with as few as two elements from claims data: date of service and cost.

With less data to collect, preparing data for input into the model is easier and delivers faster results to health plans and care teams. The need for fewer data elements is also useful with limited claims data and in cases where clients limit data sharing for privacy purposes.

Greater flexibility. Greater value.

Health plans need predictive models that can adapt to a variety of use cases. For instance, predicting cost for different conditions and risk levels or assessing members who exceed costs at multiple thresholds. To deliver greater flexibility in assessing populations, our new high-cost model provides outputs in dollar amounts and risk scores.

Each member is unique. So is their future risk.

What also sets this model apart is the use of 22 AI-driven temporal cost features that provide more sophisticated and personalized cost predictions. Research has shown significant correlation between temporal cost features and high-cost members. Examples of temporal features include peak months for costs and medical cost trend.

Other important variables for future cost predictions include:

  • Demographics, such as age and gender
  • Healthcare utilization, including number of emergency department, inpatient and outpatient visits
  • Chronic conditions, such as type 2 diabetes, hypertension and lower back pain
  • Medications, for example specialty drug use

Importantly, where many models input only overall cost, Geneia’s new high-cost claimant model looks at each individual cost as a distinct event.

Member comparison

As an example, let’s look at two members who each had costs of $1,200 in the prior 12 months:

  • Member A has a chronic condition and via claims data we know she spent $100 per month.
  • Member B on the other hand, had an accident and spent the full $1,200 in one month.

By using the temporal cost features to make the future cost prediction, the new model accounts for the accident being a one-time event and provides a superior result.

An input of only overall cost would result in the same future cost predication for both members, even though actual future costs are likely to be quite different. By using the temporal cost features to make the future cost prediction, the new model accounts for the accident being a one-time event and provides a superior result.