FDA Using Medicare Data Linked to NCI Databases to Look at Cancer Signals

FDA is looking at linking Medicare data to National Cancer Institute registries to enhance the information from medical claims data and gain a better understanding of disease-cancer and drug-cancer relationships. A familiar figure from FDA epidemiological work is keenly involved: safety officer David Graham.

FDA Office of Pharmacovigilance & Epidemiology Associate Director for Science and Medicine David Graham is digging back into the Medicare database to check for relationships between drug use and the incidence of cancer.

The starting point for his investigation into new linked database sources is familiar: diabetes and pancreatic cancer. Graham was very publicly associated with an examination of Medicare data for evidence of a different set of side effects (cardiovascular) associated with the GlaxoSmithKline PLC drug Avandia in 2010. (Also see "FDA's Brilliant Management of Avandia Re-Review Provides Flexibility on Final Decision" - Pink Sheet, 1 Sep, 2010.) He also received unusual prominence for an FDA safety officer from his role in the Vioxx safety challenge.

The current projects are big database coordination efforts and may actually be keeping Graham out of the limelight and deep into the weeds of putting together databases and finding ways to create new search tools. He describes the project publicly as “mountains of work” and looks to timelines ten years in the future when the length of data from different streams will be much larger. However, the fact that he is digging into the data and is pursuing familiar fields represents a form of early warning on topics that could develop into higher profile issues.

At a September 10-11 FDA public workshop on “Methodological Considerations in Evaluation of Cancer as an Adverse Outcome Associated With Use of Non-Oncological Drugs and Biological Products in the Postapproval Setting.” Graham explained his recent efforts to combine information from the FDA/CMS SafeRx Integrated Medicare Database and the National Cancer Institute’s SEER (Surveillance, Epidemiology, and End Results Program) Database into a resource called SIMS – SafeRx Integrated Medicare-SEER database.

SafeRx is an ongoing collaboration with the Centers for Medicare & Medicaid Services that allows FDA access to data back to 1990 for all 50 states. CMS and FDA work together to determine research topics; FDA designs studies, and CMS contractor [Acumen Medical Inc.] does the studies under FDA’s direction.

FDA and CMS decided to try using the Medicare data for drug-cancer associations several years ago. FDA has previously used SafeRx to look at REMS compliance and impact; patterns of drug use and persistency; in-depth exposure-outcome studies; and drug effects in the dialysis population.

SEER is an NCI-sponsored network of regional and state cancer registries, with rigorous quality control. For the SIMS linkage, each registry provides the patient identifiers to CMS for patients who are 65 or older. CMS links those SEER patients to the Medicare enrollment files and then returns the file to NCI. That creates a file for FDA use which has the complete Medicare claims history for each Medicare beneficiary cancer patient in SEER.

Graham said the linkage is complete for about 93% of SEER patients 65 or older and is updated every 2 years, with a 3-4 year lag in data. In December 2014, FDA is anticipating an update with SEER data through 2011.

SIMS consists of all SEER cases in Medicare and the corresponding geographic-based population for those cases.

For SIMS, FDA obtained the NCI SEER Medicare data for 2000-2009 and identified Medicare beneficiaries who live within SEER-associated zip codes for 1990-present. That creates a database from which a denominator can be built from the number of all Medicare beneficiaries and the “numerator is cancer cases that are in Medicare and SEER as well.” In ten years, the SIMS database will expand to include 34 years of hospital and outpatient claims data, 18 years of drug data, 20 years of SEER cancer data. “The hope is that we’re sort of on the cusp of having what we think will be a very valuable resource for studying drug-cancer associations in older Americans,” Graham said.

FDA has flagged several pilot projects to apply to SIMS. One example: the agency is currently using it to look at disease-cancer and drug-cancer associations in pancreatic cancer, as a precursor to look at drug exposures for diabetes and their relationship with pancreatic cancers. Graham said incretins are at the top of the list to examine for pancreatic cancer.

FDA designed an etiologic study to identify all incident cases of pancreatic cancer between 2000 and 2009 and to look at exposures of interest: type 2 diabetes with a known duration; and antecedent pancreatitis. Both are thought to be risk factors for pancreatic cancer. In the database, FDA has 27,000 cases of pancreatic cancer and 260,000 match controls.

Going Beyond SEER

The SEER-Medicare database covers about 26 percent of the Medicare population and only has data up to 2009. FDA believes it can expand and update its examination of Medicare data by creating an algorithm with which FDA could search the full Medicare claims database up to June 2014. “It’s a way of really increasing the sample size and leveraging the amount of exposure that we get to work with, as well as follow up time to account for latency and the like.” Graham explains.

To that end, FDA is working on using the SIMS database for algorithm development and so far have developed cancer algorithms for seven cancers with sensitivity and positive predictive value >80%.

FDA’s first foray into algorithm development was for esophageal cancer. FDA imposed prior history screens and inclusion/exclusion criteria for the entire database and then stratified the database by SEER registry. Within each SEER registry, FDA partitioned the database into a random sample for training purposes to develop the algorithm. FDA develops candidate algorithms in the training database and do performance characteristics in the testing database.

Graham said that algorithm is performing well with high agreement between the Medicare diagnosis data using the algorithm in the SEER data.

A further goal is to apply an algorithm – such as the esophageal cancer algorithm –to the entire Medicare database.

Graham described a case-control study, not yet analyzed, where FDA identified all cases of esophageal and cardia cancer using, looking at the exposure of oral bisphosphonates.

A current limitation of a drug study is the length of maximum exposure data – now at six years. “But ten years from now when this database is much more mature, you’ll have a huge amount of follow up time and will be able to account for longer durations of use, if there’s a threshold effect, and longer latencies of use as well,” Graham said.

One issue will be how well algorithms perform in the jump from SIMS regions to the entire country. “We don’t know yet what the tradeoff will be,” Graham said. “One of the things we will have to do is go after a sample of medical records in the non-SEER areas to see how well the algorithm performs.”

He noted that in pharmacoepidemiology, a positive predictive value for an outcome of 85 or 86 percent is pretty good. It “brings us a lot closer to answering public safety health concerns than being limited to SEER, where we’ve got four year lag in the cancer data and we’re limited in terms of exposure information that we have.” But, he noted, those limitations may become less of a problem down the road when there are more years of SEER-Medicare data.

The possibility of too much misclassification is real, but “the misclassification introduced by positive predictive value of 85% -- that’s better than most algorithms we use for most outcomes except maybe myocardial infarction, so it would be in line with the limitations of general pharmacoepidemiology.

If, however, a poorly performing algorithm has a positive predictive value of 40% “it’s probably hopeless.”

Graham acknowledged applying algorithms to the entire Medicare population is not “by any means the end-all and the answer” and there are a lot of restrictions and limitations. “But we feel it’s a step forward.”

CDER Lead for the Sentinel Initiative Marsha Reichman suggested a potential side benefit would be that algorithms with high predictive values could perhaps be used in other non-Medicare claims-based data sources. There was also discussion of the possibility of doing validation studies with other non-SEER cancer registries, such as working with the North American Association of Central Cancer Registries.

Next steps for FDA are to complete initial algorithm work and etiologic studies. Graham said FDA has spent the last several years just trying to put the database together and are “only beginning to see the mountain of work that needs to be done.” It’s now a matter of time for the agency to complete work and publish algorithms so others can use them.

“You’re looking at the entire FDA resource,” Graham said. “It’s me. And I’m the rate-limiting step.”

Exhibit 1

Cancer Counts in SafeRx Integrated Medicare –SEER Databse

Below is a prevalence ranking of the number of Medicare beneficiarieis treated in fee-for-service settings by type of cancer, from Office of Pharmacovigilance & Epidemiology Associate Director for Science and Medicine David Graham’s presentation on the SafeRx Integrated Medicare-SEER database.

Category	# of Beneficiaries	% of Medicare FFS SEER Cases
Prostate	286,952	16.0
Lung and bronchus	262,339	14.6
Breast	248,165	13.8
Colorectal	201,713	11.2
Bladder	90,761	5.0
Melanoma – skin	89.831	5.0
Kidney and renal pelvis	48,047	2.7
Non-Hodgkin lymphoma – nodal	47,889	2.7
Pancreas	46,420	2.6
Uterine corpus	39,445	2.2

Source: David Graham slide presentation “The SafeRx Integrated Medicare-SEER database (SIMS): a resource to study drug-cancer associations” September 10, 2014