NEW YORK – As researchers and health regulators in the US increasingly rely on real-world clinical and genomic data to understand the efficacy, safety, and impact of interventions on cancer patients, the lack of structured data from molecular testing labs remain a barrier for large data repositories trying to amass this information and advance precision oncology.
At a public workshop on real-world evidence, hosted by the US Food and Drug Administration and the American Association for Cancer Research, an expert leading one such data repository asked molecular diagnostics labs to step up and provide structured genomics data when reporting results.
"We make a collective call for molecular diagnostic labs to provide structured data as part of routine reporting," said Wendy Rubinstein, deputy medical director of the Association for Clinical Oncology's CancerLinQ.
While many labs aren't doing this, Rubinstein pointed out that they have responsibility to enable the integration of genomic information in electronic health records under the 21st Century Cures Act, which aims to facilitate secure exchange of electronic health information and discourages information blocking.
ASCO launched CancerLinQ in 2015 aiming to address the challenge of interoperability among oncology practices. The real-world data analytics system collects data through multiple healthcare systems and EHR vendors and identifies trends that doctors can use to improve the quality of cancer care and ultimately, deliver personalized care. The resource now includes data on the clinical encounters of more than 1 million patients with a primary cancer diagnosis. CancerLinQ has also supported EHR implementation programs with 10 vendors and boasts around 100 participating organizations.
While CancerLinQ has been ramping up, real-world data has also gained traction among researchers and regulators as a window into how cancer patients are being treated outside of the well-controlled parameters of traditional clinical trials. The FDA defines real-world data as information about a patient's health status or on healthcare delivery that is routinely collected from sources other than a clinical trial, such as claims databases, EHRs, registries, and mobile devices, or data generated by patients at home. Real-world evidence is what can be gleaned about the benefits and risks of healthcare products from real-world data.
There has been particular interest in using real-world data to evaluate precision oncology approaches since the increased use of genomic testing often identifies rare cancer indications and treatments that may not be approved for that particular use. While there is a need to assess how these molecularly-informed strategies are impacting patient outcomes and costs, a traditional clinical trial may not always be possible.
The current challenge for real-world evidence generation, however, is that most big data sets in oncology either have a lot of clinical data and little genomic data, or vice versa, and Rubinstein acknowledged that CancerLinQ is stronger on the clinical data. Although CancerLinQ is working on growing its cache of genomic data, the lack of structured genomic data has made things difficult. For example, most next-generation sequencing test reports are sent to providers as PDF files or as faxes, and therefore aren't structured and computable for integration into EHRs.
Rubinstein cited data from CancerLinQ to illustrate the scope of the problem. For example, less than 2 percent of around 32,300 records from advanced non-small cell lung cancer patients within the database had innate structured data on EGFR testing, which is widely recommended and readily performed in this setting. After CancerLinQ curated approximately 8,000 records from advanced NSCLC patients, however, 85 percent had structured EGFR data.
For two years, CancerLinQ has tried to grow the genomic data in the system largely by curating test reports. According to Rubinstein, CancerLinQ is evaluating other strategies, such as extracting genomic data by scanning reports in standardized formats, as well as using data aggregators.
The best solution, however, would be if molecular testing labs shared genomic test reports in a structured format to begin with, but labs don't want to give away this data for free. In an editorial published in the Journal of Oncology Practice earlier this year, Rubinstein wrote that labs commonly sequester structured genomic data out of "a desire to leverage the financial value of the structured format."
In an interview this week, she elaborated that although structured genomic data exists within labs and they are able to provide it to organizations like CancerLinQ, most want an incentive to do so. Moreover, data exchange agreements have to be inked with individual labs, and the terms can differ from lab to lab. "It's not as easy as just calling up the lab and saying, 'Hey, can you send me this [structured] file," Rubinstein said.
While labs may not want to freely share structured genomic data, they certainly generate it and use it in within standard genomic testing and analysis workflows. Molecular testing labs generate variant call files containing structured data following NGS testing that is used to perform bioinformatics analysis, for example. But the information that goes into patients' reports is subsequently restructured in a way that impedes its extraction into EHRs.
This practice not only hinders the aims of projects like CancerLingQ to use real-world evidence to improve cancer care, but it also restricts oncologists' ability to practice evidence-based medicine. "Genomics reports are reimbursed through government and public insurers but the format itself is impeding things like quality measures, which are part of required reporting, so the utility to clinicians is impeded," Rubinstein said at the workshop.
The lack of structured genomic data is also hindering the delivery of precision oncology, she said during the interview. Cancer centers and hospitals that are investing in expertise and tools to evaluate NGS test reports, and are using the results to inform patient treatment, will need to use real-world data to determine whether these approaches are improving patient outcomes. Without linking structured genomic data to clinical data, such healthcare quality analysis will not be possible.
Groups like CancerLinQ are spending enormous resources to find technological workarounds to this problem, but they wouldn't have to, Rubinstein observed, if labs just reported results in structured format.
Some labs already involved in big data initiatives, such as Foundation Medicine, may be better prepared to provide structured data than others, she acknowledged. For example, Foundation Medicine and Flatiron Health, both subsidiaries of Roche, have built a clinico-genomics database that's been a leading resource used for real-world evidence generation. Several months before Roche brought these companies under its aegis, it inked a real-world data collaboration with Syapse.
Syapse is a company that aims to facilitate research and precision care through its data sharing network and software platform. At last week's real-world data workshop, Syapse CEO Jonathan Hirsch highlighted that Syapse's learning healthcare network includes more than 440 hospitals across multiple healthcare systems.
"These organizations have all agreed to share deidentified clinical and molecular treatment and outcomes data with each other in order to power clinical care as well as outcomes research," he said, highlighting that they have come together despite significant legal, political, and technical challenges in creating this type of data sharing network. "Many of the problems associated with creating one of these networks is how you align the incentives of participants … and one of the most powerful ways to do that is to put the aggregated learnings back in their hands."
For example, using the clinical and molecular data amassed within Syapse's network, researchers looked at the concordance between the treatment approach suggested in molecular test reports and the therapy prescribed by the physician, and reported an average compliance of around 18 percent across three healthcare systems. "This may reflect different definitions of 'actionable' between molecular testing companies and clinicians as well as patient performance status changes over time, insurance coverage for off-label use, or referral to a clinical trial," researchers concluded.
On the other hand, another 2017 study in collaboration with the Swedish Cancer Institute showed that when physicians followed the recommendations in a molecular test report, reimbursement tended to be lower. "The more actionable mutations a report had, the less likely an insurance company was to reimburse and the lower rate the insurance company paid," he said. "The insurance company doesn't know the results of the test, so there's no causation there, but it's certainly curious."
Although Syapse has been steadily generating real-world evidence for its healthcare system partners using its network and database, integrating molecular data is a "notorious" impediment to scaling up any clinical-molecular database, Hirsch acknowledged. The difficulties are further exacerbated by the growth of genomic testing in cancer care in recent years, as well as the diversity of tests on the market.
Syapse has brought genomic data into its platform by integrating directly with molecular testing labs, including Caris Life Sciences, NeoGenomics, Foundation Medicine, Guardant Health, Paradigm, Tempus, and CellNetix. Although each lab reports genomic results differently, the direct integration makes it easier to automatically harmonize the data and interpret it.
"Most labs don't have the data structures behind the scenes that would allow for easy interoperability," Hirsch explained. Syapse works with these labs to either help them create structured data report files or the labs provide the structured files they have. Syapse then uses an internally developed process to map the genomic markers in these reports to a standardize format and bring that data into its platform.
Hirsch gave the example of EGFR exon 19 deletions, which labs reported 43 distinct ways, according to data tracked by Syapse, with one lab reporting this variant using 16 different labels over four years. Syapse's process would extract this information from structured reports and normalize it to a standard representation.
While deals like the ones Syapse has inked with its partner labs are happening, financial incentives are likely driving them at the moment. Molecular diagnostics labs, under constant threat of having their test reimbursement cut, undoubtedly see additional revenue opportunities in the genomic data they've amassed.
"In this space labs are under financial pressure to show clinical utility in order to get paid for their test," Rubinstein said. "They're trying to figure out ways to combine structured genomic data with clinical data with organizations that have it. Therefore, there is a value to it … They'd rather get some kind of concrete value for it, rather than give it away."
Given these tensions in the field, it's perhaps not surprising that molecular diagnostics labs have largely remained quiet in response to Rubinstein's request that they share structured genomic reports. She said that neither her editorial in the Journal of Oncology Practice, nor her statements at the workshop have spurred much of a reaction.
"It's sort of a beginning of a call to action, and we're figuring out ways to amplify it," Rubinstein admitted, reiterating that molecular diagnostics labs have a responsibility to share structured data under the 21st Century Cures Act. But she also acknowledged that right now, "it's ultimately up to [the labs] to fall under those parameters."