George Foreman — boxer, clergyman and entrepreneur — named his five sons after himself. So when the Nationwide Health Information Network (NHIN) is up and running, how will a doctor find the records for the right George Foreman?
Accurately matching patients with their electronic records is at the heart of the proposed network. But what if doctors search NHIN and find no records for anyone named George Foreman? If few matches are found, users will soon pronounce the network a waste of time and money, and they’ll abandon it.
However, if too many George Foreman records are found, the network could seem equally useless. Just imagine the number of records created over the years for the boxer’s sons and others with the same name who are not related to the more famous Foremans.
In that case, a doctor might be unable to determine which of the many records relate to his or her patient. If the doctor guesses wrong, the patient could end up with treatment that’s ineffective or even harmful. What’s worse in the eyes of many people is that the doctor’s employees could see the records of someone else’s patients.
Alternatively, someone from the doctor’s office could call the patient and ask questions such as, Did you ever live on Maple Street? Did you seek treatment for a broken leg in Grand Rapids? What was your maiden name? But that approach is labor-intensive and hardly seems to fit with the notion of a 21st-century information network. It also isn’t likely to provide enough value in return for the billions of dollars it will cost to create the network.
Those are the issues confronting NHIN’s planners and architects, including the four systems integrators that assembled prototype networks under contracts with the Office of the National Coordinator for Health Information Technology: Accenture, Computer Sciences Corp., IBM and Northrop Grumman.
First lessons The four contractors demonstrated their systems in January, locating patients’ records in systems across state lines and reviewing them in real time before audiences in Washington, D.C. No doubt they breathed sighs of relief because, until then, no one was 100 percent sure the systems would work.
Still, questions remain about whether records matching will work fast enough and accurately enough in a network with at least several thousand users. And although few skeptics are willing to go on the record, some have expressed concerns about scalability and response times.
The technique the prototype systems used to match patients with their records is called probabilistic matching. It involves searching for several pieces of information — often the patient’s name, address, date of birth and gender — and retrieving records that match.
Such searches generally look only at the demographic data in a database, called the master patient (or person) index (MPI). When the match meets the desired degree of accuracy, the network retrieves the clinical record from a separate data system.
The NHIN architecture calls for this two-step process — matching records then retrieving them — partly to avoid inappropriate searches for patients’ personal data and partly to ensure faster retrieval times because the system gathers only the most likely records.
Although the MPI can reside at various locations on a network, a regional health information organization (RHIO) or other health information exchange will probably maintain the MPI as one of the services it offers.
The primary, or enterprise, MPI in three of the four NHIN prototypes was Initiate Identity Hub software from Initiate Systems, a data integration company based in Chicago. The fourth prototype, developed by Northrop Grumman, primarily used the Elysium software suite from Axolotl, a Silicon Valley company that specializes in health information networks. Other MPIs were at work in the RHIOs that exchanged information through the prototype networks.
Accurate re
cords matching is important to meeting the expectations of communities and patients, said Lorraine Fernandes, senior vice president of Initiate Systems’ health care practice. It will increase confidence in a health information exchange, and confidence is crucial for ensuring the participation of health care providers and their patients.
“Certainly, you don’t want everyone to categorically opt out or you’ve defeated your business objectives,” Fernandes said.
The deployments of the Initiate Systems MPI in the NHIN prototypes were not particularly complex, said Scott Schumacher, the company’s chief scientist and senior vice president.
Each of the networks was a little different from the others, and the company had to resolve some connectivity issues, he said, but “the basic matching/linking technology was pretty much out of the box.” What won’t work What about scalability and response times? The prototypes did not involve large numbers of records and transactions, so they did not demonstrate the feasibility of a system with thousands of users and millions of records. But, Schumacher said, the company has deployed systems with billions of records that provide responses to queries almost immediately. He declined to elaborate, but Initiate Systems is working with intelligence agencies that use its matching capabilities.
“What won’t work is if we have 500 separate [enterprise MPIs] or RHIOs or whatever, and they’re all trying to communicate with each other,” Schumacher said. “I think you’ve got to keep that number around 50 or below 100 maybe.” With about 50 nodes, one-second response times are possible, he said.
Not coincidentally, that number coincides with the number of states. With the Health and Human Services Department putting more money into developing state-level health information exchanges, a consensus seems to be emerging that states will be important building blocks for NHIN. However, other kinds of health organizations are also expected to serve as nodes — for example, the Veterans Affairs Department or electronic prescription networks. At the prototype demonstrations, all the prime contractors seemed to share the belief that NHIN would have 100 to 200 nodes within a few years.
When asked about that, Schumacher said he thought 150 would be feasible — and more desirable — than a centralized service for the whole country. “Most health care is local, so you’d get most of your lift at the RHIO level and…the next one at the state level,” he said.
“There are a lot of different ways to architect it,” he added. “I don’t see anything, though, in terms of scalability that scares me from either approach. The one that does scare me is if we say we’re going to take all of these little RHIOs — and there are a thousand of them — and we’re going to try to link them together in a peer-to-peer set. I don’t think that works.”
When asked about scalability, Don Grodecki, founder and president of Browsersoft, cited the Internet as an example of a distributed architecture that is highly scalable. His company markets an open-source MPI used at one of the prototype node sites.
National identifier: Pros and cons Probabilistic matching is an alternative to an approach that seems, on the surface, to have the advantage of simplicity: the use of a national health identification number, much like a Social Security number. Most people have ID numbers issued by insurers, hospitals and medical practices. A national ID system would assign each person one such number. If the number appeared on every electronic record associated with that individual, linking records nationwide would be much easier.
But Congress has responded to voters’ concerns about privacy and bureaucracy by barring HHS from issuing national health ID numbers. The outlook for a change in that law is not promising. Some people have discussed ways around the ban — for example, establishing a nonprofit number-issuing organization that would charge fees to cover its costs. But assigning national
ealth ID numbers would take years, and no one is beginning the task.
Without a national ID number, there will always be uncertainty about whether the George Foreman with records in one system is the same George Foreman with records in another system.
The consequences of a mismatch could be dire, and mismatches will happen from time to time. But NHIN proponents say the information on the network could be so valuable to health care providers and their patients that it’s worth the risk of occasionally getting the wrong records.
Adjusting the accuracy settings Fortunately, matching systems can be adjusted to improve accuracy and control the outcome of the matching process. Administrators can set the algorithms used to determine matches so that a result would almost certainly be correct. Or they can tweak the algorithms to return a batch of possibilities that includes less certain candidates.
The NHIN prototypes used both approaches. The Northrop Grumman Health Solutions version was set so that there was practically no chance that a requester would get the wrong record, generating a false positive. However, that meant that some records belonging to the patient would not show up. In that case, the system might report falsely that no records were found, known as a false negative.
IBM’s approach, on the other hand, presented a list of possible matches, ranked in order of likelihood. But that strategy raises the possibility that the doctor’s office will see more than just the records of its patients.
“There were questions that were asked a lot this year but were unanswered,” said Robert Cothren, chief scientist at Northrop Grumman Health Solutions. “Is it worse to provide the wrong information, thinking that you got a positive match on this patient, but you matched the wrong one? Or is it worse to play it safe and not provide very much information when you’re pretty sure you had it right, but you weren’t positive?”
The NHIN MPIs “are going to set their thresholds such that there is an extremely high probability that those records that are presented to the authorized query” belong to the same person, Fernandes said.
Administrators can also tweak matching algorithms to meet local circumstances. For example, if a community has an ethnic population with an unusually high incidence of the same names, administrators could adjust the MPI to make clearer distinctions among the names.
“There is a lot of tuning going on in these things,” said Richard Hillestad, who leads the Management Sciences Group at Rand. He is working on a study of matching technology that will be published later this year. He would not reveal the conclusions of his research but did say that probabilistic matching appears to work. Dirty data Another concern often voiced about health information exchanges and records matching is the poor quality of the records in many health providers’ offices. “We know that the information out there is not perfect,” Cothren said. “There have been studies that said only 98 percent of the time is your gender right on your medical record. Simple things like that.”
If the records have misspellings and other errors, how can an automated system find matches? How will a computer determine that George Foreman and George Forman are the same person? Probabilistic matching can tolerate data variations, but its flexibility has limits.
“Sharing data is kind of like having a party at your house,” Schumacher said. “You’re opening your doors and letting people in, and I think that the individual owners of the data are going to want to paint things up a little bit and make sure things are a little cleaner. I would [expect to] see an increase in data stewardship” among health care providers.
That raises the question of how much human intervention will be needed to maintain NHIN once it’s up and running. Will its operation consist primarily of computers sending queries and responses, or will people have to field the requests and fulfill them? Dr. Brian Kelly, executive director of the health and life sciences practice at Accenture, said that with the typical probabilistic matching specification, “there’s a 98 percent chance that these are the same patient and a 2 percent chance that they’re not.” The 2 percent requires human intervention to determine whether the match is accurate. Although that sounds like a small number, Kelly said, “if you start talking about national data exchanges, even a couple of percentage points of manual intervention is a lot of work.”
The health information exchange in Mendocino County, Calif., uses Browsersoft’s MPI, and record location “is sort of a semiautomated process” there, Grodecki said. But whether that situation will be the norm nationally is unknown.
Fully automated or not, the matching must work. “Patients need to be confident that we’re matching them properly,” said Maria Vargas, a project manager at Ciracet, a health IT consulting firm in San Juan, Puerto Rico, at a recent conference.
At another conference, Dave Webster, an executive enterprise architect at IBM Global Business Services, said much the same thing. False matches “mean we’re going to lose the trust of our patients and our physicians,” he said.
“If you can’t match the patient, we’re not exchanging anything,” Vargas said.
From the battlefield to the home front: Managing medical data
Government Health IT presents Col. Claude Hines Jr., program manager for the Defense Health Information Management System, in this recent InSight eSeminar. Col. Hines discusses the health information technology and tactical challenges faced by the military medical community in Iraq, Afghanistan and other areas of conflict. In doing so, he describes the current information technology solutions for transferring clinical data between battlefield care givers to health care personnel at military treatment facilities worldwide.
