Slashdot Mirror
Data Mining Competition To Improve Drug Safety
An anonymous reader writes "The OMOP Cup is a competition to find new methods for detecting drug side effects. There have been several cases over the last few years where drugs have had issues that haven't been detected for years after they were released. The proliferation of electronic medical records and pharmacy claims provides a large and potentially powerful new data source for faster detection. The problem is that the techniques for doing this on a large scale are immature. The OMOP Cup is trying to help fix that. They've already given out $5,000 for top methods, and there's $15,000 still up for grabs."
I'm curious as to what proportion of medical records were incorrect and what proportion of pharmacy claims were inaccurate. Just how useful could a database like this be, if its source data is prone to some really wacky error? Side effect misdiagnoses alone could make for a considerable slog through noise just to pick up an effect.
"What's the use in being grown up if you can't be childish sometimes?" --Fourth Doctor, "Robot"
You parse 16 gigs, and what do you get?...
I've lost all my marbles except one & It's fun to test angular & centripetal acceleration in my skull
Dig through a huge dataset like this looking for problems and you will find them. For everything. And with the FDA and court's approach, that means that even tiny effects will end up getting drugs knocked off the market. The COX-2 inhibitors are a perfect example. Sure, for a tiny number of people taking them they increased the risk of heart attack. But for untold numbers of people taking them, they relieved pain either where nothing else would, or with far fewer side effects than other pain medications like opoids. But in today's society, we're not allowed to trade a tiny risk of death versus an enormous chance of pain relief. So even the very tiny risks one can find this way will end up getting whole classes of non-lifesaving drugs off the market. No more analgesics, no more antihistamines, no more decongestants (oh, wait, they already virtually banned those for the War on Drugs), no more cough medicines, etc.
Datamining by Kaiser Permanente helped find a previously unknown risk of heart attack for certain users of VIOXX, and it was probably right that the drug was withdrawn. It may also be a good idea to re-approve it for those who would benefit with minimal risk. But without the post-approval datamining we would never know what the risk/benefit truly was.
That also points to some limitations of our drug approval process. Trial patients are followed only for so long. In fact, some of the trial patients who were taking it long term had heart attacks, but that data was not included because they occurred past the end of the required study period.
If you believe in the science that brings you modern medicine to begin with, then more knowledge is always better.
"I still dont understand why the FDA does not allow drugs that have not been certified to be administered to terminal patients. There is no additional risk from doing it." Because then there's no reason for companies to do trials, and you'll only have "we don't know if they work, but we sell them anyway" drugs. You'll also have all kinds of snake oil salesmen, which you don't want in a life-or-death situation. Rationality tends to break down in extreme situations.
Fleur de Sel
I could focus on reducing mortality due to prescription drug side-effects and maybe get $5k? That's awesome, I can totally help out my fellow ... what's that? I get $1M if I rock NetFlix's movie rating DB? Movie ratings here I come!
This strikes me as a huuuuuge breach of medical record confidentiality. Where exactly do they plan to legally get enough medical records to mine in the first place?
As long as eighteen HIPAA identifiers are removed, the data is considered deidentified by HIPAA. Deidentified data does not need patients' consent. De-identified data-only studies only need the hospital IRB (Institutional Review Board) approval. Believe me, it's not an easy task to get the IRB approval.
Here's the list of the 18 HIPAA identifiers.