If you want to play with this and you're looking for a decent sample exam note, I grabbed some pieces of standard looking notes (Physical, ROS, Hand exam) and tossed them into a gist here: https://gist.github.com/molsches/32fcec2499e95b5b23bc268800e...
Mostly impressive in how it parses the data and can find conditions and tests in the unstructured data. Handles a few things strangely, but I imagine that gets better over time as it continues to be trained.
I think for it to truly be useful it needs some layer of semantic data mapping to something to standards like IMO/SNOMED/LOINC/RxNorm etc but I could see that being where other companies build their "products" on top of AWS vs. AWS competing with other Healthcare ML vendors in the space.
Did you grab several pieces from several notes? This looks like a garbled mess to me, but if I understood how you had transformed it to get it into this state, it might make more sense.
The posted text is actually a series of "dot phrases" or "smart texts" (aka templates in EHR speak), _not_ actual notes. Mostly these are inserted into clinical notes to achieve a certain level of documentation; typically they are saved in a "all negative" format, and the relevant parts are edited to reflect the patient history and physical. These seem to have been derived from the Univ of Washington Emergency Medicine residency.
These might make OK material for some initial testing but they don't reflect a real clinical note. A good source of those might be the MIMIC database [1]
Of course. I mostly wanted to get something out quickly. Those are all a bunch of small snippets of ROS and Physical Exams which demonstrate many of the things which the article discussed, notably finding Dx, Medications and whether or not they were negative or not.
I didn't know that there was a database of notes like MIMIC! I'll have to check it out.
"It could also be an economic benefit to the Seattle-based center, Mr. Trunnell said. The center has employed about 60 people to scan and pull essential data from records on about 500,000 cancer patients. As automation does more of the work, some employees could do other tasks."
I work in the tech healthcare industry. I wonder why they only went with (or focused on?) ML/NLP text analysis to analyze data. There is a wealth of discrete data available in EMRs (pharmacy, lab, eMAR, etc.). Yes, there is plenty of diagnosis text but that is almost always associated with ICD10 codes. The only area where I believe text analysis would be useful is documentation and microbiology data, and in many cases micro is discrete as well.
The textbook answer is typically that there is exponentially more data in unstructured data, but I agree with you. I thought about the unmodified use cases for this tooling for awhile and the major use case that I could come up with was finding unknown interactions or side-effects for medication and treatment combination regimens. That's all tracked very poorly in discrete fields today, especially with the nightmare that is cross-organization problem list management. Most drugs are approved by the FDA with the understanding that there will be post-marketing research (read this: https://undark.org/article/fda-drugs-post-marketing-research...), which this could hypothetically streamline.
Ex-healthcare IT data person weighing in with a probably ridculous idea:
The focus on NLP is because in 20-30 years, discrete data won't be entered into EMRs by providers at all.
Here's the pitch: Providers hate EMRs. They begrudge the time they spend working in EMRs instead of focusing on patient care. They loathe spending their office hours working on documentation. They hate spending their time in meetings with IT staff talking about EMR workflows and form-building. They hate getting lectured by accounting staff about how their documentation affects billing. They couldn't care less about this stuff and in an ideal (or even just a slightly more efficient) world, they wouldn't have to.
Every EMR that depends on webform-like inputs that map to database fields is going to be seen as a roadblock to patient care, which is what providers see as their primary concern.
I think the focus on NLP makes a ton of sense if you frame the problem with EMRs that way. EMR software will continue to be the bane of providers' existence right up to the day that providers can simply talk to their EMR in whatever way they want and have the computer interpret their speech and reliably spit out the discrete data that the accountants and number-crunchers want.
Obviously, I'm talking about a monumentally complicated undertaking and it's a bit on the ridiculous side. But I honestly believe we're going to go there, if only because providers won't be happy until we do.
As someone who works in healthcare clinically and also is a product designer, I agree with some certain parts of your idea. It would certainly help providers. While you are right that providers might not like documentation and filling out forms, other staffs like nursing hates typing reports. Getting rid of forms would lengthen nursing documentation which would directly delay patient care more than providers having to fill out forms. Delaying nursing care would stagnate patient flow which directly impacts hospital's bottom line. Furthermore, there are quality metrics and other stats hospitals measure that the cost of getting wrong would be more than the cost of spending a little extra to get it right since it affects ratings etc. Hospitals are too risk averse so how widely they depend on NLP would be questionable. What we need is better designers to design the EHRs better to actually fit the reality of operations. A lot of times you have hospital admins making decisions on form designs without understanding usability so you end up with horrible forms.
I don't think EHRs are going away anytime soon because of their strong integration and impact on hospital operations. I would love for Amazon to partner with existing EHR companies but those companies tend not to be very open with their applications. If anything, Amazon can use this as a way to get into the EHR business and slowly take away the operational control EHRs have over hospitals.
Operational control EHRs have over hospitals is because of executive level dysfunction. I have personally seen large Epic sites where the CxOs had never actually used Epic hands on before forcing it upon the staff and spending well over $100M on implementing it. Later when they saw the 40 clicks it took just to create an encounter and heard disgruntled providers complaints, they were stunned (too late).
This is the beginning of the final chapter in the story about the last kind of private data becoming owned by advertising companies. The final stroke of pen writing away our privacy to another institution in the name of a better future.
19 comments
[ 3.0 ms ] story [ 64.8 ms ] threadMostly impressive in how it parses the data and can find conditions and tests in the unstructured data. Handles a few things strangely, but I imagine that gets better over time as it continues to be trained.
I think for it to truly be useful it needs some layer of semantic data mapping to something to standards like IMO/SNOMED/LOINC/RxNorm etc but I could see that being where other companies build their "products" on top of AWS vs. AWS competing with other Healthcare ML vendors in the space.
These might make OK material for some initial testing but they don't reflect a real clinical note. A good source of those might be the MIMIC database [1]
[1] https://mimic.physionet.org/
I didn't know that there was a database of notes like MIMIC! I'll have to check it out.
[0] https://aws.amazon.com/blogs/aws/amazon-comprehend-medical-n...
or be fired
The focus on NLP is because in 20-30 years, discrete data won't be entered into EMRs by providers at all.
Here's the pitch: Providers hate EMRs. They begrudge the time they spend working in EMRs instead of focusing on patient care. They loathe spending their office hours working on documentation. They hate spending their time in meetings with IT staff talking about EMR workflows and form-building. They hate getting lectured by accounting staff about how their documentation affects billing. They couldn't care less about this stuff and in an ideal (or even just a slightly more efficient) world, they wouldn't have to.
Every EMR that depends on webform-like inputs that map to database fields is going to be seen as a roadblock to patient care, which is what providers see as their primary concern.
I think the focus on NLP makes a ton of sense if you frame the problem with EMRs that way. EMR software will continue to be the bane of providers' existence right up to the day that providers can simply talk to their EMR in whatever way they want and have the computer interpret their speech and reliably spit out the discrete data that the accountants and number-crunchers want.
Obviously, I'm talking about a monumentally complicated undertaking and it's a bit on the ridiculous side. But I honestly believe we're going to go there, if only because providers won't be happy until we do.
:)
I don't think EHRs are going away anytime soon because of their strong integration and impact on hospital operations. I would love for Amazon to partner with existing EHR companies but those companies tend not to be very open with their applications. If anything, Amazon can use this as a way to get into the EHR business and slowly take away the operational control EHRs have over hospitals.