[HN Gopher] GPT-4 will hunt for trends in medical records thanks...
___________________________________________________________________
GPT-4 will hunt for trends in medical records thanks to Microsoft
and Epic
Author : samizdis
Score : 33 points
Date : 2023-04-18 20:27 UTC (2 hours ago)
(HTM) web link (arstechnica.com)
(TXT) w3m dump (arstechnica.com)
| just_steve_h wrote:
| Gee, I'm so glad that I "opted in" to have my private medical
| data used for corporate research!
|
| How did I "opt in"? My healthcare provider migrated to Epic for
| all medical records. I had the "choice" of... not receiving
| healthcare, I guess? But even then, all my past records were
| imported to Epic anyway.
|
| And before you "market choice" folks start squawking: your
| provider changing their records system is not a "qualifying
| event," so you don't get to shop for another provider until the
| next enrollment period. Also, ALL of the providers available to
| me use Epic.
| avereveard wrote:
| Poor IBM Watson, killed by not being hip
| unsupp0rted wrote:
| Having GPT-4 play Jeopardy today would be no fun at all
| Arrath wrote:
| You might be able to make a drinking game on when it forgets
| to phrase the response as a question.
| civilized wrote:
| If Watson had been good at anything but Jeopardy, IBM would
| have spent billions making sure we all knew.
| unsupp0rted wrote:
| Is anybody currently feeding a country's worth (e.g. Israel) of
| medical records into an AI and asking it to come up with novel
| off-label uses of medicine?
|
| E.g. "TDAP Tetanus vaccine = 42% lower risk of dementia in older
| adults"... more findings like this would be great.
|
| https://pubmed.ncbi.nlm.nih.gov/33856020/
| sigmar wrote:
| >Several vaccine types are linked to decreased dementia risk,
| suggesting that these associations are due to nonspecific
| effects on inflammation rather than vaccine-induced pathogen-
| specific protective effects.
|
| Doesn't this make it sound like there's a confounding factor?
| Such as: vaccines are more common in people that are wealthier
| or people that choose to align their lifestyles with
| recommendations from doctors?
| kjs3 wrote:
| In my dark little heart of hearts, I know there's some US
| insurance company executives sitting around thinking "how can I
| use machine learning to automate the process of denying
| coverage based on 'trends' or whatever that I won't have to
| justify because I'm not obligated to make the LLMs decision
| process transparent to anyone".
| andrewstuart wrote:
| I've been very much against regulation of AI at all.
|
| But now that you're talking drawing inferences from medical
| data...... it's making me pause to think.
| worldsayshi wrote:
| At some point there needs to be a comprehensive way to allow
| individuals to make decisions about how their data should be
| handled. Allow me to decide how my medical records should be
| handled in the same way that I can manage whether my location
| data should be shared with the app I'm currently using. Allow
| me to revoke consent and set up general rules. Don't make me
| repeat myself unless there's a good reason. Make a system I can
| trust.
|
| I think it could be done in the EU at least. If we up the it
| competence level of our decision makers a couple of notches.
| Semino wrote:
| I hope not.
|
| It's time we start finding cures for illnesses like mine
| which doesn't has medication.
|
| It's time to use big data and ml on as much data possible to
| start helping much more people faster.
|
| Every CRT every diagnosis helps.
|
| Feel free to advocate for proper anonmysation if the data but
| not for the use of it.
| kazinator wrote:
| GPT-4 or its successors, will kill medical records as such.
|
| And it will be like shooting fish in a barrel.
|
| It is said that doctors spend a lot of their time on bullshit
| computer procedures, involving garbage software, when they should
| be seeing patients or learning something.
| debesyla wrote:
| I doubt that anything can kill BS medical records. Two reasons
| that I see:
|
| 1) privacy and other laws have high demand of "structured" and
| "secure" information keeping ("it's a law that I must fill this
| information into this BS table and it's 100 BS fields");
|
| 2) governments love to spend the money on "standalone" software
| that has 1000 BS requirements... And contact companies do love
| to make that software as BS as they can because the worse
| system it is, the more additional contracts do they get to
| solve the problems they made.
|
| Basically, I am sceptical that anything could help medical
| staff. (I am pessimistic because of what I see in my country of
| Lithuania. Barely working proprietary systems ("e-sveikata" and
| "e-receptai") that costs tax payers hundreds of millions
| Euros.)
| worldsayshi wrote:
| > will kill medical records as such
|
| Why do you think so?
|
| I see a huge thorny ball of legal issues that needs to be
| resolved before gpt-like services can really become useful...
|
| I personally feel very held back by legal (and infosec)
| uncertainty. I would use chatgpt a lot more of it didn't feel
| so legally sticky (?).
| Spivak wrote:
| From Microsoft's pov there's basically no infosec issues.
| They're running their own copies of GPT inside their airlock.
| It's no different than any other analysis they do on behalf
| of their customers.
|
| To everyone else it's turning data over to a 3rd party but
| for them it's like running scapy.
|
| They can't train on one customer's data and use it on another
| but that's less of an issue than you'd think.
| astrange wrote:
| Hmm, how does a language model "search for trends"?
| yeck wrote:
| Doesn't understanding language require that a model understands
| trends in the features of text embeddings inputted into it? And
| haven't we found that the more text and transformers the
| complex the trends it can understand in that text? If you have
| enough features present in medical records, why wouldn't it be
| able to find trends?
| astrange wrote:
| Because 1. it has a limited context window, so it has no
| permanent memory 2. you can't fine-tune ChatGPT/GPT-4 to get
| around this, and of course 3. language models don't do
| anything by themselves unless you input text into them.
| og_kalu wrote:
| By inferring connections in data pasted into its context
| window. Except its context can be as much as 50 pages at a time
| and semantic search means you can paste relevant context
| related to any query buried in mountains of data into context
| automatically.
|
| Pages and pages of text that can't be read by any one person
| can be parsed in seconds by a machine with enough knowledge and
| understanding to make something meaningful of it.
|
| The same idea is already being applied for law -
| https://twitter.com/ai__pub/status/1644735555752853504
| Semino wrote:
| Good.
|
| Can't wait for more people getting help.
|
| Alone a system which knows all patients with my sickness (RLS)
| could potentially show something when clustering 100ktens of
| diagnosises
___________________________________________________________________
(page generated 2023-04-18 23:02 UTC)