HIPAA and Artificial Intelligence | Hacker News Reader

HIPAA and Artificial Intelligence

ailawadvisor.com

61 pointsraleighm14 comments

Can we please have more openness of the (anonymized) training data, so that we can have more healthy competition in this space?

For example, I'd hate it if insurance companies built all the apps because they have all the training data.

Healthcare organizations (insurance and delivery) see data as a competitive advantage. I don't foresee this happening any time soon.

In addition to large datasets, this protectiveness extends to data specific to an individual patient. Consider this scenario: You need to see a specialist, and have the choice of providers; specialist A is within the same system as your PCP, while specialist B is with a smaller, independent practice. Specialist A therefore has access to the same EMR as your PCP, while specialist B does not. This has a direct impact on your experience as a patient: at specialist A, it feels like their office "knows" you; your visit to specialist B, however, would include filling out a bunch of forms (demographic info, insurance, medical history...) that you've filled out countless times before.

Of course, it's in the patient's interest for their data to flow freely between providers, no matter which healthcare system they're on. And many healthcare systems recognize this, and strive to share data. But they are often going against their own economic interests to do so.

> Healthcare organizations (insurance and delivery) see data as a competitive advantage.

Its not just seeing it as such, it is. Data is everything these days.

Healthcare data can't be shared the way the browser histories, cell phone location data, etc. can. It's a completely different set of a rules that people have to play by (HIPPA for example). I build machine learning systems using healthcare claims and EHR data and without the direct cooperation of several large insurance companies (and access to their data) we'd be dead in the water. Even with access to their data there are incredibly strict limits to what we can and can't do with it. You can't just go out and collect healthcare data the way you can many other types of data.

Hopefully the NIH's All of Us project, an effort to "gather data from one million or more people living in the United States to accelerate research and improve health", will help in this regard.

https://allofus.nih.gov/

What is the state of art to anonymize data for AI ? I know numerai does this for stock markets (using homomorphic encryption or a loose version of it).

What is the state of art of research on this ? Google and Microsoft are trying something around this https://www.technologyreview.com/s/601294/microsoft-and-goog...

https://synapse.ai/ or https://doc.ai or https://enigma.co/ or ...

all have (or will have) the capacity for doing this

This is why the patient's data needs to be in their possession and control so they can share and use as they see fit.

Under HIPAA you have a right to access your health information.[0] (with some minor exceptions) Just ask and you will be able to get it if you need it.

[0]: https://www.hhs.gov/hipaa/for-professionals/privacy/guidance...

Sadly it doesn't mean that when they give it to you it will be in a reasonable form, or even free (they're allowed to charge a per page processing fee). I work in this space

Correct, it is hard to get in a useful form. Then if you want to share it with a different Provider on an different EMR, it is non trivial to get that data imported into your data record.

Of interest: we built an AI tool to detect (and optionally redact) personal and sensitive data, including health data: https://pii-tools.com.

As amelius mentions in a sibling thread, collecting realistic data for training and evaluation was pretty non-trivial. Openmined.org is a very interesting effort in this direction!

I wonder if the organ donor model might apply for data collection. People could sign up to be a data donor which would make their data available anonymously for AI or other medical research purposes.

"Anonymously" is really really hard to do.

Reply on news.ycombinator.com