Humana says it's leading the health benefits industry into a new world in which the focus will shift from employers, doctors and hospitals, where it has been for decades, to consumers. The company says it can better serve its members by giving them greater choice and greater control over their health and health benefits.
But giving consumers choices -- and pricing those choices optimally -- requires analytic tools of unprecedented sophistication to make sense of terabytes of health care data. Humana is developing such tools and eventually hopes to patent them. Its tools use algorithms developed jointly by epidemiologists, engineers, economists, mathematicians and -- literally -- rocket scientists.
If consumers are at the top of Humana's strategy pyramid and analytic models are in the middle, then computer technology forms the base. The US$13 billion company has put together an IT infrastructure that serves up data to analytic modelers and recycles the results of those models back into a 3.5TB data warehouse. The data store supports a vast array of users, including claims processing and billing personnel; patients' employers, doctors and hospitals; the rocket scientists; personal nurses working from home; and Humana's 6.8 million consumers of medical insurance and medical care.
"The purpose of the industry in the past was generally unlinked to the end user," says Dr. Jack Lord, a physician and Humana's chief innovation officer. "It tended to focus on itself, and on employers acting in sort of a benefactor role to employees. The result was a consumer and public push-back."
It was a simpler world then, Lord says, one in which health insurance companies managed costs by "supply-side interventions" with employers, doctors and hospitals. Traditional actuarial tools were quite adequate. "You'd say, 'I'm going to forecast tomorrow's weather based on yesterday's.' It was always a historic look," he says. "If you live in that space, you never want to move; but if you stand still, you can never influence the net cost of health care."
Now the name of the game is "choose and use," says Carol McCall, director of the Center for Health Metrics in Humana's Innovation Center. Humana has models to formulate and price health insurance plans. These predict who will choose a given program, how they will use it and what it will cost the company. Humana even has a model that predicts who will become catastrophically ill so it can intervene and try to head off those illnesses. Although Humana has not yet published the results, it says the new model has already shown costs savings for the company while saving money and improving health for the at-risk members.
This year, Humana plans to develop an ¸bermodel, which draws on these predictive and analytic tools, that could send the company in directions it can't at present anticipate. For example, the ¸bermodel could discover a major customer grouping that has been overlooked.
The models do more than simply extrapolate from the past using classical statistical methods such as regression, according to McCall. "There need to be new disciplines around predictive and behavior sciences," she says. Indeed, McCall's group is using complexity theory, agent-based modeling, genetic programming and other "new science" esoterica. It also uses Markov models, Bayesian learning networks and pattern recognition techniques borrowed from military and space programs, in which torrents of data are sluiced for tiny nuggets of information that may be good predictors.
McCall's group has developed four predictive and analytical models that it calls insight engines. This year, it's working on a fifth model, called SimHealth, that will combine results from the other models. Rather than making static predictions, SimHealth will produce scenarios that evolve during the simulation.
"It's one of those things where nobody knows the answer, but they'll know it when they see it," McCall says. "So you want to use what's called interactive evolution. You have a model -- it's like SimCity -- and you run scenarios. You say, 'I like that scenario,' and you press the big button and evolve it."
Bruce J. Goodman, senior vice president and chief service and information officer, says one of the challenges he faced when he came to Humana five years ago was figuring out how to pull together information from many different stovepipe systems, integrate it and position it for use by a number of constituencies. "We had multiple claims systems and multiple administrative systems, so one of the challenges was, how do you provide a single view for each of the stakeholders?" he says.
The answer was two huge, integrated data stores, one to feed a Web portal and one for the modeling community. An elaborate extract, transform and load (ETL) process developed to feed the data repositories. "We decided which data elements we needed for the (data stores) and pulled those systems together so we could promote the common view, even though we have disparate systems under the covers," Goodman says. "We were able to make transparent the true underlying complexity of our systems environment."
The operational data store (ODS), an e-business data mart, drives Humana's Web site, a single portal with separate, secure entrances for members, corporate customers, providers, agents, business partners and employees. The MVS-based ODS holds 24 months of data -- 1.8TB or 180 million DB2 database rows -- about providers, employers, members and their medical and pharmacy claims.
While the ODS is just for Web users, the real information engine at Humana is the AIX- and Oracle-based enterprise data warehouse (EDW), "a complete set of data assets used to run the business," according to Bruce Sterpka, a vice president for corporate information management at Humana. The EDW holds some 3.5TB of data, and the largest of its 432 tables, the table of medical claims, has 430 million rows.
The central IT function at Humana is claims processing, where members seek reimbursement for millions of medical and pharmacy outlays each month. Claims byproducts, which the IT people and modelers call data "exhaust," include diagnostic codes, severity codes and other information that the modelers extract and use to predict illnesses, benefits-plan usage, costs and other variables.
A Cobol job periodically extracts the exhaust data from the EDW for the modelers in Humana's Center for Health Metrics. The models run on two four-processor Windows 2000 Server machines in the center. Results are stored on the modelers' own network-attached storage system before being sent back, via file transfer protocol, to the EDW for recycling into other models and to the ODS for Web access.
Modelers code and test their models using custom C and C++ code and the MATLAB development tools from The MathWorks. The models then go to IT. "Our key step is to take what they've developed and industrialize it, to make it bulletproof and scale it so we can run large amounts of claims information through it," Goodman says.
IT will rewrite the models in Java for production runs, says Ramu Kannan, a director in corporate information management. That will make them more modular and will also make them capable of providing real-time visualization of model output on the Web, he says.
IT has invested $1 million on the modeling work so far and has eight to 10 people supporting it full time, Goodman says. "IT is so well aligned with the business," he says. "We anticipated what we had to do to make the data accessible . . . to enable the business to really take advantage of the technology and move forward."
Analytic Engines Deliver Insights
Humana's insight engines apply analytic models to 3.5TB of customer, claims and other data to identify markets, enhance products and predict costs. The following are the four engines Humana has completed, plus a fifth, SimHealth, that's in development.
SmartStart Plus Goal: Predict the consumer's choice of benefit plan; explore benefit/contribution strategies.
Approach: Models consumers as "rational agents" that evaluate plans and trade off costs, benefits and risks to pick the best plan.
Predictive Modeling Goal: Predict future high-cost (illness-prone) members; improve customer relations.
Approach: Combines medical knowledge, engineering methods (asynchronous signal processing, nonlinear dynamic time series) and computer science (learning algorithms, advanced visualization).
Impact Tool Goal: Evaluate effectiveness of programs; analyze consumer behavior.
Approach: Creates control and test groups on the fly for dynamic analysis of clinical and financial results.
Insight Tool Goal: Enhance pricing and underwriting competitiveness; early detection of trends.
Approach: Uses historical data and predictions of individuals' future health to identify patterns and drivers of health care costs, including early trend and anomaly detection at the employer, market and provider levels.
SimHealth Goal: Simulate consumer choice and behavior via self-evolving models.
Approach: In development now, SimHealth uses "rules of the game" (weighted consumer objectives) to evaluate different benefits-plan/consumer scenarios. Evolves using the results of other models, genetic algorithms and agent-based modeling.