A Delicate Operation

In the grand scheme of things, 12 percent of anything isn’t much. It’s a dime and some pennies out of a dollar, the merest sliver of a whole pie. But when 12 percent describes a group of people dying, the figure suddenly looms ugly and huge. That’s the percentage of pneumonia patients at Norfolk, Va.-based HMO Sentara Health System who died in 1993 and 1994. Although lower than the national average of about 14 percent, the number still seemed too high to Sentara’s doctors. What’s more, a good many other pneumonia sufferers developed complications requiring expensive antibiotics and an average stay of two weeks, as opposed to the optimal three- to five-day stay of appropriately treated pneumonia patients without complications, says Robert D. Brickman, Sentara’s medical director of clinical effectiveness.

The doctors at Sentara always knew something was wrong. When they ordered sputum cultures for pneumonia patients, they wouldn’t hear back from the lab for several days—if at all. Without lab results, they could only guess what was wrong and how best to treat it; meanwhile, patients got sicker as they waited for results. Yet it wasn’t until Sentara’s quality improvement team began an exploratory mining foray into claims data in its Oracle-based data mart that the doctors’ suspicions were confirmed. Besides uncovering the high pneumonia mortality and complication rates, the team noticed that doctors were ordering sputum cultures many times for the same patient; upon inspection, they learned that the doctors did so in hopes that at least one test would yield useful, timely information. The quality team quickly devised a new system of transferring the culture from the patient to the lab, and the lab results back to the doctor within two hours. Not only did the mortality rate for pneumonia patients fall to 9 percent, but their average stay dropped to one week and the cost of managing a single pneumonia case decreased by $2,000. “Attack the quality of health care, and amazingly the costs come down as well,” says Bert Reese, Sentara’s corporate director for information systems.

The process that led Sentara to improve the quality and efficiency of care for its pneumonia patients is known in the health-care industry as “outcomes measurement.” A form of data mining—a means of searching for previously unknown, actionable information from large databases—outcomes measurement involves examining clinical encounter information, insurance claims and billing data to measure the results of past treatments and processes. By understanding what worked—or didn’t—providers can identify areas for improvement or capitalize on successful methods.

Because outcomes measurement (also known as outcomes analysis) can influence whether patients live or die, it’s no wonder the health-care industry is emerging as a leader in data mining. CIOs in other industries don’t deal with data of life-or-death importance (although sometimes it may seem that way), but they can still apply the wisdom of their health-care colleagues to their own data mining operations to cut costs, improve efficiency and ensure customer satisfaction. And since many of the issues and problems associated with outcomes measurement apply to data mining, CIOs can benefit from learning how their health-care counterparts address such perennial problems as end-user buy-in and security.

“The business of practicing medicine is so complex that it pushes the sophistication of data mining tools,” says Brad L. Armstrong, global managing director of Deloitte & Touche Consulting Group’s Health Systems Integration practice in Los Angeles. “There are thousands of services, relationships built over time, and multiple diagnoses and interactions. That complexity will continue to push progressive data mining applications that will rival those in banking or retail.”

Applications of data mining in the health-care industry are widespread. One way data mining is helping health-care providers cut costs and improve care is by showing which treatments statistically have been most effective. For example, once hospital administrators recognize that stroke patients are less likely to develop respiratory infections if they can swallow properly, they can educate their physicians and institute a standard policy to identify and provide therapy to those who have difficulty swallowing. Outcomes measurement also helps HMOs evaluate their doctors and facilities. Physicians and hospitals benefit from knowing how they compare with their peers, and the parent company saves money by getting all of its employees up to par. (Incompetent doctors are an expensive liability because high-quality preventive care is much more cost-effective than correcting a mistake later in the emergency room.) Along the same lines, outcomes measurement also lets caregivers identify people statistically at risk for certain ailments so that they can be treated before the condition escalates into something expensive and potentially fatal. Fraud detection is yet another outcomes analysis application: When administrators compare the volume of lab tests ordered by physicians, evidence that a doctor is ordering too many unnecessary tests would leap off the page. Finally, providers are using data mining to cross-sell and market new products and services. An allergy sufferer who has tried a variety of treatments without success, for instance, might be receptive to alternative herbal or acupuncture treatments.

“Information has become the most valuable commodity in health care,” says Jeffrey C. Bauer, president of Hillrose, Colo.-based consultancy The Bauer Group Inc. “In the past, studies have shown that as many as one-third of all medical interventions do not lead to an improvement in patients’ conditions. In other words, about 33 cents on the dollar is spent on services that do not demonstrably make the patient better off. But today we can afford to provide only productive interventions. Outcomes data will finally allow us to weed out the resources that aren’t making people better.”

In the case of Sentara, outcomes analysis is helping decrease variability of treatment among its five hospitals, 180 doctors, nine nursing homes, nine urgent care centers and nine Navy outpatient clinics. “The way you eliminate error is to homogenize and standardize so [that things are] done the same way every time, on time,” explains Brickman. A team of hospital administrators and doctors scours Sentara’s data mart with Orlando, Fla.-based MedAI Inc.’s decision-support software, looking for the most effective and least expensive treatments. Their recommendations are passed on to the rest of the staff. “If we can reduce variation around the best practice, we can improve quality of care and the outcomes,” says Reese.

Defining the Focus

As with any large data warehousing and mining endeavor, the degree to which an organization reaps the benefits of outcomes measurement depends on how its IS executives resolve a host of issues. For instance, one of the biggest strengths of outcomes measurement—the ability to view data in the aggregate—can be a pitfall if it discourages consideration of cases on an individual level. “Any time you make broad, generalized statements about data, you’re liable to miss specific cases,” says Armstrong. “In banking, retail or any kind of business, it’s dangerous to make assumptions about customers’ needs based solely on common characteristics they possess.” In the medical field, overemphasizing aggregate data can have dire consequences for a patient. A doctor relying on clinical practice guidelines based on statistical analysis of outcomes might never think to check for glaucoma in a 29-year-old patient, because it’s not standard procedure, explains Wayne Farnsworth, assistant professor of emergency medicine at the State University of New York’s Health Science Center at Syracuse.

Exacerbating that danger is the need of many large organizations to reduce all symptoms and descriptions to the lowest common denominator to make data mining work. Every hospital has its own system for coding and record-keeping, which can cause difficulties when they attempt to merge their data. If one hospital records patients’ gender as “M” or “F” and another uses “1” and “2,” the data can’t be matched and therefore can’t be mined. Most providers are avoiding definition differences by sticking with standard ICD-9 codes, the international categorization of outcomes that assigns a code number to every health problem. But ICD-9 codes are too shallow to be of much value, says Trevor Price, director of health-care consulting at the New York-based consultancy and software developer Software Services International Inc. (SSI). “That’s not good enough, because you have no idea what was done 10 steps before and after that ICD-9 code was classified,” says Price.

Yet for many organizations, settling for the bare minimum is preferable to trying to get everyone to agree on a common set of definitions—the scourge of IS professionals everywhere. People’s natural instinct to protect departmental turf can jeopardize the whole project, as was the case for a large insurance company that recently spent more than $1 million in 14 months on a data warehouse before killing the project because its five lines of business couldn’t agree on the definition of “product,” according to Alan Paller, director of research and education at The Data Warehousing Institute in Bethesda, Md. But there are ways around the data standardization standoff. Melville, N.Y.-based home health organization Olsten Health Services, for instance, assigned a representative from IS, a “czar of nomenclature,” to hammer out terms between IS and the physicians, says William C. Reed, Olsten’s senior vice president and chief information and administrative officer. Health Resources Group (HRG), a national chain of dialysis clinics based in Santa Ana, Calif., relies on highly detailed pop-up menus built into SSI’s Homer data mining tool. Nurses at the point of care choose the appropriate symptoms and conditions from an extensive list. The level of detail afforded by the pop-up menus makes the aggregated data all the more useful for mining.

Collecting the Right Data

The quality of the outcomes measurement analysis also depends on the type of data being examined. In health care, the information available for mining typically is financial in nature because the return on investment for collecting such data has been most obvious, says Brickman. Somewhere along the way, people realized that billing and claims data could be mined for clinical purposes. But some critics argue that outcomes analysis based on claims data is of limited value. Mining claims data might indicate that a patient had a lab test, but to see the results of the test, a doctor would have to read the patient’s paper medical file, says Brickman. “The financial data can just suggest areas to look at,” he explains. “It can measure [mortality] and whether a patient received a drug, but the further you drill down, the less reliable the information.”

As more health-care providers invest in computerized medical records that will make clinical data more accessible, outcomes analysis for clinical insight will become more reliable, says Bauer. But that shift will be slow and expensive, costing about $35 million to $50 million for a billion-dollar organization like Sentara, says Brickman. And insurance companies aren’t in any hurry. “A lot of payers say they want outcomes information, but at the bottom line, clinical outcomes [analysis capability] is secondary and may or may not be paid for,” says Reed. “Cost is still the number-one driver.”

Yet a few forward-thinking organizations, Olsten included, are moving beyond collecting financial information. Those providers tend to view data warehousing, data collection and medical equipment not as separate projects but as components of a cohesive, overarching IT strategy. HRG, for instance, has a wealth of clinical data available for data mining because its dialysis machines automatically collect such data as a patient’s blood pressure while the patient is being treated and transmit it to HRG’s Sybase data warehouse.

Balancing Privacy and Access

Any time an organization deals with consumer data, privacy and security become paramount. That is especially true in the health-care industry where the data in question—medical records—is highly sensitive. To protect themselves and their patients, some health-care organizations have developed exemplary security strategies, says Bauer. “The fact that there is no national scandal, no outraged patients, shows that the health-care industry is doing something right that the government and banks can learn from,” he says. Aside from building firewalls and encrypting data, some health-care providers safeguard patient data by limiting who has access to it. Sentara, for instance, includes patient identifiers in its data warehouse, but lets only a team of about six doctors and administrators look at the data. If other doctors have a query, they must submit it to the quality improvement team. “The doctors can’t access the information, because it has to be confidential,” says Brickman. “Every plaintiff’s attorney would love to get his hands on this detailed information [for malpractice suits]. It can be protected from discovery only if it’s kept private.”

Intermountain Health Care, an integrated health-care system in Salt Lake City, sidesteps the security issue altogether by stripping its databases of any patient identifiers. Only an account number is used to link elements of a patient’s record, stored in six Oracle data marts. Leaving the data anonymous ensures that any of Intermountain’s 20,000 or so employees can access its riches through the company’s intranet, says Data Warehousing Project Leader Ping Wang. Doctors mine Intermountain’s Oracle database themselves using Netscape or Microsoft Explorer browsers to answer what-if questions and determine the most appropriate treatments.

Removing patient identifiers from the database makes sense for organizations interested in looking at their data in the aggregate. But it also limits what can be accomplished with the data. If no names are included, providers cannot find and alert patients whose lab results or vital signs deviate from the norm, indicating that they are at risk for certain medical conditions. Because Norwalk, Conn.-based managed-care provider Oxford Health Plans’ database does include patient identifiers, Oxford’s medical analysis team can refer problems it finds to the disease management program team, which alerts the affected patients’ physicians. Omitting patient identifiers also rules out the kind of real-time decision making in practice at HRG. If a problem arises while an HRG patient is undergoing dialysis, nurses can use SSI’s Homer software to compare the patient’s real-time data with that of other HRG patients who have had similar problems to find the best solution, says the company’s president, Cynthia Jansen.

Cultivating User Buy-In

Getting buy-in from the end user can be another thorny area for IS executives implementing a data mining project. Convincing users in marketing, HR or out in the field (be it branch office employees or doctors in a hospital) to surrender peacefully their standard modes of operation for a new technology is never easy. Doctors might be even less tolerant to forced change than most users since they’re accustomed to a high degree of professional autonomy. “The doctors were nervous at first about being compared because they saw their reputations at stake,” says Brickman about Sentara’s data mining efforts. “They believe they graduated from medical school and should be allowed to make their own decisions.”

One way to foster user acceptance is to keep the data model as simple and easy to understand as possible. End users—doctors included—will never agree to change their procedures if they don’t understand how the system works. “If I make the model very complex and intricate, how do I expect a novice with no IS background to do a query?” asks Reed.

But the surest way to circumvent end-user resistance in this and all other major IS initiatives is to include users from the beginning and listen to their feedback. Reese’s strategy of proving the concept with a real-life example also can’t hurt. After seeing the value of data mining for pneumonia patients, the Sentara doctors embraced the concept. “This is not about good and bad doctors, it’s about good and bad processes,” says Reese. “The premise is that physicians want to help patients. If you can convince the doctors that something’s good for the patients, then you’re OK.”

And, as it turns out, so are the patients. In time, data mining could render obsolete numbers like the 12 percent pneumonia mortality rate at Sentara. And if it can help hospitals save more lives, just imagine what it could do for your manufacturing plant or customer service center.