3 ms·

So, I worked at a startup in the early 00s trying to do exactly this — detect disease trends from medical chart data. We had an advantage in that all the data c

So, I worked at a startup in the early 00s trying to do exactly this — detect disease trends from medical chart data. We had an advantage in that all the data came from a nationalized health care system, so it was relatively consistent. We were using slightly different NLP techniques (and were actually a leader in the NLP space), but ultimately NLP wasn’t really our restriction: we learned that we couldn’t trust the doctors to enter the data consistently. The data came in like 8 different versions of HL7 per hospital, and we had to fix our parsers every time some department decided to upgrade their records systems. The notes contained so much jargon (that was different in every department and often using the same acronyms with different meanings in context) that we ended up pivoting the tech to another industry. Medical chart data sucks. HL7 sucks. Pretty much everything about doing deep learning on medical data sucks. Everything about working and selling in the hospital EMR space sucks. There might be something there, but have fun building a profitable company around it.

Amygaz 7y ago
I did that data curation and analysis in the mid 2000s, before it was all hip and well paid, and the CS group employed data curators because you need and experimented human being, to analyze the data in a peer-reviewed paper, and compare it to the authors’ conclusion. Half the time it doesn’t. Finding reproducible or reproduced experiments happens at an even lower rate. Then, context matter, and we know very little about each of our cell, even less about each one behaves in an ensemble. So no amount of NLP will cure cancer, currently. It’s a simple garbage-in 101 case. It will hopefully change... However, there is a good chance that traditional approaches to science will have worked well enough for most cancers, and NLP will just be a tool to validate after the fact, build comprehensive database and assist with incremental update to a treatment.
astazangasta 7y ago
The main problem with this effort is statistical. Medical databases are not accurate samples of the population; they are biased in hundreds of ways (doctor X is running a clinic for disease Y, but doesn't see many poor people because of Medicaid issues, etc.). People get excited about these because there is "so much data", but since it isn't collected prospectively, even if we could bulk extract all the data (which we can't, it's a fucking mess) any conclusions we drew would be invalid.

Amygaz 7y ago

I did that data curation and analysis in the mid 2000s, before it was all hip and well paid, and the CS group employed data curators because you need and experimented human being, to analyze the data in a peer-reviewed paper, and compare it to the authors’ conclusion. Half the time it doesn’t. Finding reproducible or reproduced experiments happens at an even lower rate. Then, context matter, and we know very little about each of our cell, even less about each one behaves in an ensemble. So no amount of NLP will cure cancer, currently. It’s a simple garbage-in 101 case. It will hopefully change... However, there is a good chance that traditional approaches to science will have worked well enough for most cancers, and NLP will just be a tool to validate after the fact, build comprehensive database and assist with incremental update to a treatment.

astazangasta 7y ago

The main problem with this effort is statistical. Medical databases are not accurate samples of the population; they are biased in hundreds of ways (doctor X is running a clinic for disease Y, but doesn't see many poor people because of Medicaid issues, etc.). People get excited about these because there is "so much data", but since it isn't collected prospectively, even if we could bulk extract all the data (which we can't, it's a fucking mess) any conclusions we drew would be invalid.