Master Patient Index: If You Build It, Will the Data Come?
 | |
| HealthInfonet's MPI ensures that providers like Dr. Jay Naliboff Franklin, of Franklin Community Health Network in Farmington, Maine, have access to accurate patient information. Source: HealthInfoNet |
With patient care on the move, MPIs are increasingly important to ensure that patient data travels as well. This is why HIEs are only as good as their MPIs.
Demographic matching is the foundational function of every MPI—without a positive match, data can potentially be lost or associated with the wrong patient. The problem is, standards for positively matching patients to data aren't here yet.
Patient demographic matching was on the roster for the Office of the National Coordinator for Health IT (ONC) Health IT Standards Committee (HITSC) and Health IT Policy Committee (HITPC) "Summer Camp for Standards," according to John Halamka, MD, CIO, professor at Harvard Medical School in Boston, and co-chair of the HITSC. Halamka made this announcement in May at the Massachusetts Governor's 2011 Health IT Conference.
"Our focus is really narrow. We are trying to figure out what standards we might employ for the representation of patient data that is going to be used for patient matching," says J. Marc Overhage, MD, PhD, CMIO at Siemens Healthcare and former CEO of Indiana Health Information Exchange (IHIE). Overhage heads the patient demographic power team under the HITPC Quality Measures Workgroup Patient Safety Tiger Team.
The Health IT Policy Committee's patient-matching effort focuses on how the governing body of the exchange delivers data during a relatively simple encounter where a clinician requests a patient's record, Overhage says.
A second task on the agenda for the HITPC Power Team, with respect to matching for simple encounters, is to determine what kinds of information that clinicians must receive when a patient is matched in order to feel comfortable that the information will be used in accordance with the clinician's intentions, he adds.
For example, for data matched to patients for public health surveillance, a lower level of confidence in a match might be more easily tolerated than in a patient care scenario, because patients needn't be identified by name—just by location, age or some other non-identifying demographic information, says Overhage.
"How do we get that data lined up in a way with the right level of confidence and how are [those data] transmitted?" are questions that need to be asked and answered, says Overhage. Recommendations from the power team are expected by the end of the summer.
The godfathers of MPI
Bringing together data for one patient from disparate locations is the raison d'être for HIEs, and an effective MPI is critical for that effort. As the ONC's Health IT standards and policy committees hammer out patient demographic standards, healthcare organizations are grappling with various exchanges' different workflows and confidence intervals for patient demographics."For patient matching … one size doesn't not necessarily fit all, so understanding what strategies work for which circumstance for various requirements is important," says Shaun J. Grannis, MD, MS, research scientist at the Regenstrief Institute and associate professor at the department of family medicine at the Indiana University School of Medicine in Indianapolis.
"Matching for patient care purposes by pulling together disparate clinical data to make medical decisions typically has a low tolerance for false matches. For population-level aggregate reporting, where one might be more interested in making sure [to] identify all of the known cases for a particular condition, you might want an aggressive matching algorithm."
The Indiana Network for Patient Care (INPC), the HIE operated by Indiana Health Information Exchange, was established in the 1990s and leverages an internally developed MPI. Currently, the index includes information on more than 10 million patients across 90 hospitals, long-term care facilities, rehabilitation centers, community health clinics and providers in Indiana.
IHIE leverages the technology that Regenstrief develops to provide HIE services within the INPC. Some of the patients' data goes back to 1972, Overhage says, and some identifiers have changed during the past four decades, which limits their utility. For example, "If you want a longitudinal patient record, you can't rely on ZIP Codes," says Overhage. "If a patient sees a physician for eight years and then switches [to a provider] in another ZIP Code, it might hard to link up that data five years later."
Overhage, speaking as the former CEO of IHIE, says that when building an MPI, a big win is having a highly specific matching approach—a matching system that would never or very rarely match patient data inappropriately. He also adds that an MPI, at scale, needs to be able to function without human intervention. "In the traditional hospital system-based MPI, it would be very common to have someone work on the MPI and resolve questions, [but] that is simply not practical at a community scale," says Overhage. In a traditional enterprise MPI, the goal is to provide one common enterprise-wide ID for a patient. In a community or HIE MPI, "you accept that there will be many IDs for the same patient that need to be matched. This is a somewhat subtle but important distinction."
"Everyone tends to focus on the algorithm, or method for matching, but that can be a red herring [deliberate attempt to divert attention]," says Grannis. "Even more important is understanding the quality of data being used for the matching. Understanding the nature, characteristics and quality of data will determine the maximum accuracy and performance characteristics."
"Once you understand the quality of data, then an algorithm and strategies can be developed to match the data."
Regenstrief developed a deterministic rule-based system where approximately 30 rules guide demographic data to a patient match, according to Grannis. The Regenstrief MPI has been around for nearly 20 years, and the HIE has continuously validated, reviewed and updated those rules as data characteristics change over time, he adds.
Where as well as who
Creating a longitudinal record of a patient healthcare summary that shows where the patient received care is just as important as being able to obtain that information. IHIE uses a patient matching algorithm to grab all unique patient encounters for one patient identity. "Each participant in the information exchange gets an entry in the MPI to define the identity of a particular person," says Grannis.The INPC contains more than 20 million registration events. Because each stakeholder is the curator for patient identity, as each unique patient enters different care environments within INPC, a patient encounter has the potential to be added to a patient group. The INPC system "takes the separate encounters/identities and creates a 'patient group' to aggregate patient encounters across the care spectrum to one unique patient," says Grannis.
Ideally, the information that all stakeholders in the HIE enter is quality data so the information can be pulled together and recognize separate registration events as the same patient. "If one stakeholder, for example, enters an erroneous date of birth, we may not include her in a patient group."
A Regenstrief analysis suggests that the INPC's specificity, the true negative rate, is higher than 99.5 percent, meaning there's less than a 0.5 percent false positive rate, Grannis says. "Our sensitivity, which is our true positive rate, is at least 93 percent, approaching 95 percent." To date, the INPC's functionality (including the MPI) has adequately supported the use cases desired by the stakeholders.
However, there's always trade-off in patient matching, Grannis notes. "If you want more matches, you will do so at the expense of false positives. Conversely, if you want to avoid all false positives, you will do so at the expense of missing true matches. We are consistently monitoring the system to ensure it's performing as well as possible."
| Three traits of successful MPI matches |
Effective matching systems require ongoing care and feeding. Three vital functions include:
Source: Shaun J. Grannis, MD, MS, Research Scientist at the Regenstrief Institute |
To create a longitudinal record for patient's personal health information, names, addresses, phone numbers and date of birth are matched to patients who opt into the HIE, says Gary Christensen, COO and CIO at RIQI. "The more information we have, the more we can do the matching," says Christensen. Currently, there are 155,000 patients enrolled in currentcare.
If data are presented to the HIE participation gateway and the demographics are considered a match above a certain probability, then the record is considered a match and goes through the consented patient gateway. If a patient has opted in, those data will go into currentcare. If the patient is not consented or if the demographic probability is ambiguous or negative, the data are deleted, although RIQI keeps a record of when demographic data are ambiguous.
In instances where an ambiguous demographic lands in the gray area range for match probability, data managers in RIQI's Operations department review the ambiguous data to link demographics to records that the system might not have picked up. This often occurs in transcription cases—for example, if a patient lives on "St. James Street" and an organization indicates he or she lives on "Saint James St." or "St. James Street." Following the data intervention, "we can say, for future reference, 'if you encounter this data again, it's the same person,' " notes Christensen.
Reliability is top priority
"If you don't have a reliable MPI, you can't have an effective exchange," says Devore Culver, executive director of HealthInfoNet, an independent nonprofit organization based in Portland, Maine, that built and operates Maine's statewide HIE.The MPI tool used by HealthInfoNet, similar to RIQI, uses a series of rules to run against patient demographic data when considering whether patient information matches with a record. Those rules look at the probability of like data or near-like data being the same and assign it a score. Culver explains that the system determines whether to match the two records based on that score. "For example, if we are higher than 90 percent certain there's a match, generally the demographics will automatically match to the record. If the demographics fall below a 90 percent threshold, then it falls into a work queue for a human to look over," says Culver.
"Our strategy is to be more conservative than not," says Culver.
HealthInfoNet uses an IBM Initiate MPI tool that uses last name, address, date of birth, sex, ZIP Code, street address and phone numbers to match demographics for separate records in the exchange. Of the 1.3 million Maine residents across 34 health service areas, approximately 850,000 have a record in the exchange. In a two-year demonstration effort of the HIE, 15 hospitals statewide connected under HealthInfoNet's HIE services, yet none were in the same community, except one crossover hospital.
According to Culver, the rate of crossover for patients seeking care between unaligned organizations reached just under 20 percent, meaning 20 percent had at least one visit to two unaligned organizations during that two-year period.
Culver explains that when bringing in new organizations to the HIE, there are sometimes incidences of multiple patients with very similar demographic information. "Those are the cases that would be identified by the MPI to be reviewed by HealthInfoNet staff prior to a match," says Culver.
The challenge with managing patient demographics is achieving balance, says Culver. On one hand, an organization could run the risk of having two records on the same patient that remain separated because there was not a strong enough match for the patient demographics. This could create the perception that when a provider is looking at a patient record in the HIE, he or she is looking for a complete record.
"Conversely, automatically merging people who aren't the same creates a very serious liability to patient care. That's a more dangerous outcome in some respect. Those are two extremes of risk that we try to avoid," Culver says.
CMIOs should weigh their organizational needs before building an MPI or connecting to one through an HIE. But demand for quality and accurate patient data is only going to increase, and an effective MPI can help both individual facilities and HIEs confidently exchange data within set parameters.