The following is a guest blog post by Andy Oram, writer and editor at O’Reilly Media.
One couldn’t come away with more enthusiasm for open data than at this month’s Health Datapalooza, the largest conference focused on using data in health care. The whole 2000-strong conference unfolds from the simple concept that releasing data publicly can lead to wonderful things, like discovering new cancer drugs or intervening with patients before they have to go to the emergency room.
But look more closely at the health care field, and open data is far from the norm. The demonstrated benefits of open data sets in other fields–they permit innovation from any corner and are easy to combine or “mash up” to uncover new relationships–may turn into risks in health care. There may be better ways to share data.
Let’s momentarily leave the heady atmosphere of the Datapalooza and take a subway a few stops downtown to the Health Privacy Summit, where fine points of patient consent, deidentification, and the data map of health information exchange were discussed the following day. Participants here agree that highly sensitive information is traveling far and wide for marketing purposes, and perhaps even for more nefarious uses to uncover patient secrets and discriminate against them.
In addition to outright breaches–which seem to be reported at least once a week now, and can involve thousands of patients in one fell swoop–data is shared in many ways that arguably should be up to patients to decide. It flows from hospitals, doctors, and pharmacies to health information exchanges, researchers in both academia and business, marketers, and others.
Debate has raged for years between those who trust deidentification and those who claim that reidentification is too easy. This is not an arcane technicality–the whole industry of analytics represented at the Datapalooza rests on the result. Those who defend deidentification tend to be researchers in health care and the institutions who use their results. In contrast, many computer scientists outside the health care field cite instances where people have been reidentified, usually by combining data from various public sources.
Latanya Sweeney of Harvard and MIT, who won a privacy award this year at the summit, can be credited both with a historic reidentification of the records of Massachusetts Governor William Weld in 1997 and a more recent exposé of state practices. The first research led to the current HIPAA regime for deidentification, while the second showed that states had not learned the lessons of anonymization. No successful reidentifications have been reported against data sets that use recommended deidentification techniques.
I am somewhat perplexed by the disagreement, but have concluded that it cannot be resolved on technical grounds. Those who look at the current state of reidentification are satisfied that health data can be secured. Those who look toward an unspecified future with improved algorithms find reasons to worry. In a summit lunchtime keynote, Adam Tanner reported his own efforts as a non-expert to identify people online–a fascinating and sometimes amusing tale he has written up in a new book, What Stays in Vegas. So deidentification is like encryption–we all use encryption even though we expect that future computers will be able to break current techniques.
But another approach has flown up from the ashes of the “privacy is dead” nay-sayers: regulating the use of data instead of its collection and dissemination. This has been around for years, most recently in a federal PCAST report on big data privacy. One of the authors of that report, Craig Mundie of Microsoft, also published an article with that argument in the March/April issue of Foreign Affairs.
A simple application of this doctrine in health care is the Genetic Information Nondiscrimination Act of 2008. A more nuanced interpretation of the doctrine could let each individual determine who gets to use his or her data, and for what purpose.
Several proposals have been aired to make it easier for patients to grant blanket permission for certain data uses, one proposal being “patient privacy bundles” in a recent report commissioned by AHRQ. Many people look forward to economies of data, where patients can make money by selling data (how much is my blood pressure reading worth to you)?
Medyear treats personal health data like Twitter feeds, letting you control the dissemination of individual data fields through hash tags. You could choose to share certain data with your family, some with your professional care team, and some with members of your patient advocacy network. This offers an alternative to using services such as PatientsLikeMe, which use participants’ data behind the scenes.
Open data can be simulated by semi-open data sets that researchers can use under license, as with the Genetic Association Information Network that controls the Database of Genotypes and Phenotypes (dbGaP). Many CMS data sets are actually not totally open, but require a license to use.
And many data owners create relationships with third-party developers that allow them access to data. Thus, the More Disruption Please program run by athenahealth allows third-party developers to write apps accessing patient data through an API, once the developers sign a nondisclosure agreement and a Code of Conduct promising to use the data for legitimate purposes and respect privacy. These apps can then be offered to athenahealth’s clinician clients to extend the system’s capabilities.
Some speakers went even farther at the Datapalooza, asking whether raw data needs to be shared at all. Adriana Lukas of London Quantified Self and Stephen Friend of Sage Bionetworks suggested that patients hold on to all their data and share just “meanings” or “methods” they’ve found useful. The future of health analytics, it seems to me, will use relatively few open data sets, and lots of data obtained through patient consent or under license.