Using new technology to identify high-risk patients | Behavioral Healthcare Executive Skip to content Skip to navigation

Using new technology to identify high-risk patients

July 20, 2009
by Andy Sekel, PhD
| Reprints
Neural networks can help care systems provide holistic treatment

Government-sponsored programs account for more than 45% of all money spent on healthcare in the United States.1 Timely prevention and early intervention remain the keys to maximizing limited resources, helping consumers to recover from illness and empowering them to live fuller lives.

OptumHealth is enhancing its ability to find the highest-risk consumers and improve their health status before their condition deteriorates to the point that they need hospitalization. This is being accomplished through the deployment of neural network predictive models, a technology with a record of success in complex settings, such as pricing real estate, evaluating credit-worthiness, and predicting the stock market. A neural network is a type of modeling, a way of processing a lot of complex information in a manner that simulates brain functioning in which variables, like brain neurons and synapses, interact in nonlinear ways to solve problems. Neural networks can derive meaning from complicated or imprecise data sets; they "learn" patterns that are not easily observable without complex computing.

OptumHealth and its sister company AmeriChoice have piloted a set of neural network predictive models created with data from more than 135,000 Tennessee Medicaid beneficiaries. The models successfully identified a subset of high-risk patients who represented only 3% of treatment recipients but accounted for 31% of the system’s psychiatric costs in the subsequent 6 months. A youth model (under age 18) successfully identified a subset of 3% of the population that generated 59% of subsequent psychiatric costs.

Because of the success in Tennessee, OptumHealth will begin using neural network risk modeling to support its work as the statewide behavioral health contractor for New Mexico starting this month. This is the first time neural network predictive modeling will be used for this purpose.

The neural network predictive models that will be deployed in New Mexico are both proprietary and fully customizable. As a result, they will be used to define how large numbers of complex variables interact and will support the prediction of future outcomes based on this knowledge. This approach has been used successfully in complex financial markets and in real estate. In these settings, neural networks can help analysts identify market trends, segment populations of interest, and support decision making.

The introduction of highly customized neural network technology to the behavioral healthcare industry is a significant step toward treating the whole person. Current healthcare predictive models use claims data to identify high-risk consumers. While this can be effective for cost estimations, it also is inherently limited because:

  1. a consumer needs to use medical services to be identified; and
  2. information that may be critical to the assessment of risk—such as a person’s family medical history, weight, exercise patterns, or smoking habits—cannot be obtained from claims.

In sum, traditional healthcare predictive models can miss many key parameters that affect medical and psychiatric outcomes.

With customized neural network modeling, data from social service organizations could be combined with claims information to create a more complete portrait of each individual and more accurately predict his/her healthcare risks. Data from social service sources might include a person’s level of engagement with psychotherapy, the presence of children in foster care, episodes of legal problems or incarceration, and housing issues. A customized neural network model can combine claims and social data to generate risk predictions that can be used by a care manager to initiate interventions that can stabilize high-risk patients.

The specific benefits of customized neural network models for healthcare include the ability to:

  • Clearly define the specific medical and behavioral risks for Medicaid recipients as individuals or as a subpopulation (e.g., all children under 6—categorical eligibility)
  • Identify specific sources of risks (gaps in medical care, gaps in behavioral care, unaddressed psychosocial needs) at both population and recipient levels
  • Help package and communicate timely, relevant health and well-being risk information across and between care systems to facilitate proactive efforts to find and help the people most likely to benefit from medical management processes and programs

Consider the following scenario. A traditional predictive modeling tool, through analysis of medical claims, identifies a member with a longer-than-average inpatient stay for acute pancreatitis as moderately risky. The traditional predictive modeling tool, which analyzes only medical claims, would miss the presence of behavioral health claims for the treatment of schizophrenia, as well as social service information that indicates a chronic pattern of medication noncompliance.

Even with the availability of behavioral claims, traditional predictive modeling tools would miss critical intake information that may, for example, document a history of alcohol abuse as well as the presence of periods of homelessness. The individual who has been evaluated with limited information may not receive timely clinical interventions, resulting in possible deterioration and hospitalization. In many such cases, additional information could have facilitated outpatient stabilization, leading to more effective and less costly treatment.