4 ms·

Why don't they just deidentify the data set? This not-new concept would allow the researchers to provide the data without compromising patient confidentiality.

Why don't they just deidentify the data set? This not-new concept would allow the researchers to provide the data without compromising patient confidentiality. As for the argument re: follow-up study, seems that the scientific community should agree on a reasonable amount of time during which a group can keep and reanalyze their own published data after which the data must be made public (30 days? 60 days?).

Paul-ish 11y ago
It isn't easy to deidentidy a dataset. It isn't as simple as removing the names. You don't know in advanced what features in a dataset, when combined with a second dataset, will allow an adversary to reidentify data points. It isn't as simple as adding noise either, because reidentification can be treated like a machine learning problem.
rflrob 11y ago
Getting the scientific community to agree to almost anything makes herding cats look like a military parade. Even within a field, there are wildly varying opinions, and some fields tend to have higher or lower means. And thus far, I'm only talking about non-human where the primary concern is the ability to benefit from your own hard work without being "scooped on the analysis". Patient data adds a whole new dimension---with just genome sequence data, it's possible to infer someone's last name [1]. So deidentification is not as trivial as you make it sound. You can remove the obvious factors, but there's likely to be other things in there that you can't or don't know to remove that will inadvertently identify people. None of this is to excuse the original authors or KCL, but it is worth remembering that science is huge, and no single policy will work for everyone. What I think. PLoS could do is, as an editorial position, require anyone publishing with them to have some "close to raw" data publicly available BEFORE the manuscript can be published. Would at be controversial? Sure, but I think it's also in line with PLOS's mission of making science accessible to the public. [1] https://www.sciencemag.org/content/339/6117/321 https://www.sciencemag.org/content/339/6117/321

Paul-ish 11y ago

It isn't easy to deidentidy a dataset. It isn't as simple as removing the names. You don't know in advanced what features in a dataset, when combined with a second dataset, will allow an adversary to reidentify data points. It isn't as simple as adding noise either, because reidentification can be treated like a machine learning problem.

rflrob 11y ago

Getting the scientific community to agree to almost anything makes herding cats look like a military parade. Even within a field, there are wildly varying opinions, and some fields tend to have higher or lower means. And thus far, I'm only talking about non-human where the primary concern is the ability to benefit from your own hard work without being "scooped on the analysis". Patient data adds a whole new dimension---with just genome sequence data, it's possible to infer someone's last name [1]. So deidentification is not as trivial as you make it sound. You can remove the obvious factors, but there's likely to be other things in there that you can't or don't know to remove that will inadvertently identify people. None of this is to excuse the original authors or KCL, but it is worth remembering that science is huge, and no single policy will work for everyone. What I think. PLoS could do is, as an editorial position, require anyone publishing with them to have some "close to raw" data publicly available BEFORE the manuscript can be published. Would at be controversial? Sure, but I think it's also in line with PLOS's mission of making science accessible to the public. [1] https://www.sciencemag.org/content/339/6117/321 https://www.sciencemag.org/content/339/6117/321