Important Points Highlighted by the Individual Speakers
- Both academic health centers and community centers are working to incorporate genomic information into their systems, but the efforts are largely separate. Establishing data standards and common ways of representing outcomes would facilitate the scalability of efforts and the translation of genomic information into clinical care. (Moss)
- The most practical way of integrating genomic data into the clinic is to provide it through clinical decision support, but that means the community would need to agree upon common allele and test code nomenclature so that the guidance is scalable and interoperable. (Chute)
- Cultivating a “data donor” culture in which data sharing is commonplace and encouraged because it would help the greater population could be achieved by ensuring the privacy of personal information. (Chute)
Establishing standards for data will facilitate the incorporation of genomic information into the EHR, allow for the interoperability of data flow among systems, and increase the ease with which big data can be shared and managed. Still, even once those data standards have been established, additional challenges to a genomics-enabled EHR will remain, including deciding how and what information will be shared, ensuring equity of access to the information, developing useful clinical decision support and providing clinicians with the knowledge to use it, and providing insurance coverage for genetic tests that have been demonstrated to have clinical value.
The developers of EHRs are working hard to incorporate genomic information into the clinical record and to rapidly translate new discoveries into clinical care, said Moss. The push to incorporate genomic information in the EHR is coming from the consumer demand, he said. It began with the academic medical centers, but now community health centers are also very interested. Much of these efforts are carried out in separate siloes, and because of that, efforts are not consistent. Providing consistency would reduce the need for rework in this area, Moss said. Organizations are using genomic information in different ways—some of them to drive alerts, for instance, while other groups are using EHRs in genome-wide association studies.
The disparate approaches, lack of standardization, and limited sharing of approaches all present barriers to developing a scalable genomics-enabled learning health care system. “People can learn best practices from work that others have done,” Moss said, “but at a technical level there’s no sharing of what’s been done to make it easier, especially for the non-academic medical centers that are trying to do this.”
Moss pointed to three areas in particular in which changes need to be made in order to make learning health systems a reality. The most important barrier he discussed is the lack of data standards for genomic information. Various standards exist, but they are young and have not been tested and tailored to meet the needs of the genomics community. Getting useful feedback from the genomics community and moving toward standardized data models and exchange formats would help lessen the burden on local efforts to integrate genomic data into care, he said. The DIGITizE Action Collaborative of the Roundtable on Translating Genomic-Based Research for Health (see Chapter 5 for more information), which Epic participates in, is working to solve this issue by assembling a framework for integrating genomic data into the EHR.
A more standardized approach to represent knowledge is also needed, Moss said. For example, standard representations of outcomes data would help make the value proposition for payment models, additional funding, and new research. In addition, the standardization of genomic knowledge in a shareable and scalable way would help speed the translation of discoveries into clinical care. As an example of a system that works well today, Moss cited the system for checking drug–drug interactions, which has been quickly translated into clinical care and is scalable. “This model could work great for something like drug–gene interactions,” he said.
Not every system is going to do things in the same way or use the same resources, Moss acknowledged. Already, many different models have been developed to incorporate genomic data into health care, and standards will need to be flexible to support these. Developing those standards will require a collaborative process that crosses many stakeholder groups. The demand for genomic information in EHRs is only growing, he said.
A learning health system aligns science and informatics, develops strong patient–clinician partnerships, provides incentives for innovation, and creates a culture of continuous improvement to produce the best care at the lowest cost, said Steve Leffler, the chief medical officer at The University of Vermont Medical Center and a professor of surgery at the University of Vermont College of Medicine. Each of these actions, in the context of genomic medicine, can be used with EHRs to advance health care.
EHRs will need to be optimized to use genomic information effectively, Leffler said. Given that patient charts can become overloaded with extra data, standardizing EHR displays can ensure that health care providers see important information. But, he added, too many alerts can be detrimental because they are eventually ignored.
To personalize care for patients, genomics needs to be incorporated seamlessly into the EHR, Leffler said, and the genomics information needs to be accurate if it is to be useful. Genomic data will be most useful in the background, where they will help providers make good decisions without distracting them from their jobs. Eventually, all health care providers will need to know how to use genomic information, but for now primary care providers who are not comfortable with genomic information can work with genetic counselors, geneticists, pathologists, and others who understand the the test results in order to make genetics-informed health care decisions, he said. Physicians who do not want to learn from computer screens while they are practicing are likely to see this type of expert advice–based learning as a welcome alternative, said Peterson.
Although specialists will be the more likely point of interaction with regard to genomics, Jarvik agreed that “every physician is going to need to become literate in genomic medicine. But we have a long way to go right now.” She cited the example of a patient informed of a warfarin
sensitivity variant in a research study who was switched to a different drug by a physician. “This drug did not need to be changed as far as we know. We are interviewing the patient and the physician about this experience, and maybe there was some valid reason, but I’m concerned that there wasn’t.” Health care providers need training to be able to do phenotyping, to see the benefits of genomic information to patients, and to use the information in clinics, said Fowler. “There is a real dearth of skills in this particular area, and for us that’s a particular challenge.”
There are several barriers to the integration of genomics into the EHR, Leffler said. For example, determining a way to identify who will benefit in the initial stages of integration when not everyone can be included in such a system is a challenge, he said. “Are you going to focus on people who already have a disease, on their family members? Who is going to make those decisions before it’s universal?” Informed consent is another issue. Will everyone have to opt in or opt out? Will incidental findings be conveyed to family members who might be affected? Will patients be able to see all their genomic information, including incidental findings? If they are concerned about having a disease, will they be tested for that disease every time? Are providers expected to review every piece of available information? How will the use of genomic data by providers be monitored? Another consideration is that emergencies need to be dealt with quickly, so EHRs cannot slow down responses, Leffler added. “How we’re going to deal with incidental findings has to be well understood and planned out ahead of time,” he said.
Shared decision making in the age of genomics will also generate challenges. For example, a person in his or her 20s who lacks markers associated with a predisposition to lung cancer may misinterpret the results as meaning that there is protection from the disease and that smoking would be safe. Other patients may not want to know that they are at risk for a disease and would consider such information to be an intrusion into their lives. Another possibility is that patients will be overtested when genomic information is available. The result will be “rich discussions,” Leffler said, “but it’s going to take a lot of time. You’re going to need to have knowledgeable providers who understand that genomics is probabilistic, not deterministic, so these markers can make you more likely to have something, but it’s not an absolute.” Ultimately, genomics will make possible shared decision making, allow new research, drive improvements in population health, optimize care, and prevent complications, thus driving down the cost of health care and improving value.
The infrastructure for genomic medicine is lacking in critical areas, Jarvik said. For example, not all of the variant annotations are getting into central databases where they can be widely used by academic laboratories and companies. Furthermore, EHRs are not standardized nationally. Information is entered into systems in different ways, the systems do not communicate with each other, and they are not currently standardized to accept genomic information. Institutions often have difficulty communicating with each other because much of their laboratory genetic data are in the form of PDF files, Peterson said, which is “the lowest common denominator to exchange genomic data at this point, and that clearly needs to change.”
One thing that would be very helpful, Jarvik said, would be if EHRs automatically pushed these variants to the relevant databases, such as ClinVar. In addition, providing access to all the information from a genetic test, not just the information that goes into a report, could create new opportunities for discovery. Genetic test providers are competing for work, so an incentive could be created for them to adhere to a shared format for complete results, and this incentive could be reinforced with policy.
The transition from the International Classification of Diseases1 version 9 (ICD-9) to ICD-10 creates a problem for genomic research. Though few codes are fundamentally different, the transition creates discontinuities. For example, a single ICD-9 code can correspond to many ICD-10 codes, and vice versa. “That’s intractable in terms of having trivial table lookups,” said Chute. Though ICD-11 promises to improve the situation, the current systems are problematic—a point that was reiterated by several other presenters. EHRs need to make it easy for health care providers to do the right thing and hard to make errors, Leffler said. Genomics will add huge amounts of new information to EHRs, and how this information is incorporated and viewed will be critical to how useful it will be, he said.
Clinical Decision Support
Providers need to be knowledgeable about using genomic information and about discussing what the information means with their patients, Leffler said. Providers need better information, not necessarily more data, he continued. If adding genomic information to the EHR only
adds data, its usefulness will not be maximized. The key will be to integrate the information in a way that makes sense to providers and adds value to the provider–patient encounter.
The most pragmatic way of taking genomic data and integrating it into the clinical process is through clinical decision support. Computational tools and infrastructure must be available to inform physicians of relevant findings, rather than expecting them to look the finding up or know them off the top of their heads, Chute said. However, he added, the challenge to enabling clinical decision support is that the nomenclature for alleles is collapsing. The number of alleles that must be distinguished is rapidly exceeding the capability of the current system, and the designation of variants is not always consistent. Genetics laboratories are creating their own names and codes for genetic tests, which works against the consistency and comparability of laboratory results. “It’s simply not usable for clinical decision support,” Chute said.
Decision support tools tend to be binary, yielding yes or no choices, whereas genomics is probabilistic, Leffler said. He also added that for clinical decision support tools to work, the problem list needs to be correct, which is often not the case today.
The classification of variants is also a problem, Jarvik said. The University of Washington has a Return of Results Committee, which has taken on the difficult task of figuring out how to classify challenging variants and deciding what incidental findings should be returned to patients, she said. The committee has identified 112 gene–disease pairs that it considers returnable, along with reporting formats and clinical decision support (Dorschner et al., 2014). When six different genomics laboratories, all of which were certified through the Clinical Laboratory Improvement Amendments, or CLIA, classified the same variants using new guidelines developed by the American College of Medical Genetics, the resulting classifications were the same across labs for only one of the six variants (Amendola et al., 2015). “We’re going to have to come up with a system we all agree on,” Jarvik said.
Peterson said that health care providers do not always follow the advice provided by clinical decision support, and the reason is often that they have additional information about a patient that factors into their decisions. “We would like our rates of following advice to probably go a little higher than they are,” he said, “but it’s never going to be 100 percent and probably shouldn’t be.” The information generated by not following program advice goes back into the EHR, and this information could be used to do comparative effectiveness studies of the value of advice.
Genomic medicine is a big data problem, said Ketan Paranjape, the worldwide director of health and life sciences in the Health Strategy and Solutions Group at Intel Corporation. Large portions of genomic data are accessible, but they are gathered, stored, and disseminated differently in health care than in other industries such as financial services or manufacturing. Furthermore, a variety of types of data exist—not just genomic data, but also clinical trial data, various forms of bioinformatics, and even payer and reimbursement information. Genomic data are being generated not just for individuals but for pathogens, tissues, and other biological entities. Even patients are generating data of various types that could be incorporated into genomic medicine through such means as personal genomic tests and wearable monitors.
Today, the Broad Institute2 produces amounts of data that are on par with the big cloud producers such as Microsoft, Facebook, and Amazon. Even more data—more than 300 petabytes—is expected to be produced by the Broad Institute in 2015, Paranjape said. And other organizations in the United States and abroad are producing even more data.
Paranjape said that various problems with data generation, management, and interpretation pose barriers to genomic medicine. As a single example, he pointed to the problem of storing genomic data for long periods of time. “Have you thought about keeping the data in your hard disk forever?”
Several projects are intended to overcome these data-related barriers. One is a project of the Charité hospital system in Berlin, which is performing real-time cancer analysis to match patients with the proper therapies. The system uses structured and unstructured data to collect and analyze up to 3.5 million data points per patient, completing in seconds a process that used to take 2 days, Paranjape said. The result has been improved medical care received by patients and provided by doctors and hospitals. Additionally, the system has generated higher-quality information that is usable for research with on-the-fly analysis using medical records, PubMed references, pharmaceutical databases, and survival curve statistics.
The Regional Health Information Network in Jinzhou, China, is another example of an organization that is successfully managing big data.
2The challenges of analyzing hundreds of thousands of genomes, http://www.broadinstitute.org/~carneiro/talks/20140612-qatar_genomics_conference.pdf (accessed February 25, 2015).
In response to problems involving scalability, performance, maintenance, and data storage, the network developed EHR systems and health care service that run on a distributed computing system in order both to address these issues and to significantly reduce storage costs.
Paranjape cited a system that connects HCC (hierarchical condition category) codes with ICD-9, -10, and -11 codes. The goal is to identify relevant features and patterns behind diseases to more accurately identify suspected conditions in patients.
In addition to developing processors to handle big data, Intel supports several training programs in genomics and technology. Specifically, the company has a team that works with clinicians with the goal of understanding how the clinicians use genomic data. These programs have helped bioinformaticists, life scientists, computer scientists, clinicians, and other professionals work more effectively to create the personalized medicine of the future, Paranjape said.
A lack of substantial evidence for the clinical utility of genetic information has led insurers to be reluctant to pay for these tests, Jarvik said. Because of the significant amount of time allocated to consenting patients for testing and then interpreting and explaining the results, covering the costs would make it possible for genetic testing to be implemented in a practical way. The policies of insurers are an obstacle to genomic medicine, Jarvik said. For example, a large insurer in Washington State recently declared that any genetic panel is investigational, including the cystic fibrosis 32-mutation panel. “How does this single-gene test get involved here?” Jarvik asked. “The word ‘panel.’” The policy was interpreted in such a way that it did not distinguish between a single-gene and a multiple-gene test, and because the word “panel” was in the test name, it was considered investigational, she said.
Yet insurance coverage is critical as the end point of a process beginning with research and progressing through the development of an evidence base and practice guidelines, Jarvik said. “In medicine, even when we have a lot of evidence of what is best, we still need to get someone to pay for it,” she said. “So we have to think about getting practice guidelines from societies based on that evidence in order to convince insurers what is a reasonable level of care to provide for patients.” More investments are also needed in outcomes research, she said.
Regulatory changes pose another obstacle for integrating genomics into the health care system. For example, the new patient access rights that CLIA laboratories now have to grant may be interpreted to mean that raw gene variant data files are shared upon request (Evans et al., 2014). This will require that physicians explain the data to their patients, because the laboratories are not required to do so, Jarvik said. New regulations also require that FDA approve tests when variants are deemed to have clinical utility. Some people welcome that oversight, but molecular pathologists are generally not among them, Jarvik said. In general, genetic tests have had few errors, so the public health benefits of this regulation are questionable. “I have patients who have been followed for four generations before we finally solved what was wrong with them because of new technologies, and I don’t want to see that limited.”
Summary level information can be shared nationally and internationally to produce even larger patient cohorts. This type of model is becoming more common, Fowler said, and this means that procedures must be developed to share data and collaborate while protecting the privacy of patients. Making sure that the data are interoperable, so that data from many systems can be aggregated, is an important issue, Chute said.
Peterson made the case for sharing knowledge resources among institutions. In particular, the sharing of knowledge resources is very helpful to programs that are just getting started with performing genomic medicine, including phenotyping algorithms, variant calling, determining the clinical interpretation of variants, and maintaining a rule repository for clinical decision support.
Finally, Chute mentioned the idea of cultivating a “data donor culture.” Organ donation is considered popular in the sense that people are proud to tell others about their donor status. But, he said, “there’s no coolness being associated with being a data donor, and yet in terms of discovery and integration and learning health systems, nothing is more important culturally than for society to understand the importance of data sharing.”
This page intentionally left blank.