Yeah. LLMs are helpful IF you now what they are, what they can do, etc. and that you still have to check their solution.

For some things they are good and safe time.

Example: We currently work with a company who has understood this verx well: They provide doctors reports and also validate manually written ones. Their workflow basically is: The doctor orders a doctors report for patient A. The system checks witht the doctor that it has understood the major bullet points from the clinical information system and the order given by the doctor. (Basically a “yes” “no” answer system like “Patient was admited for chestpain?”) and then creates the report with a color coding in the version the doctor needs to check. (E.g. "white background for “we are 100% sure about the data as it comes directly from the clinical information system” aka names, dates, lab results, green for “very likely to be correct”, yellow for less than 99% certainity, orange for less than 90%,etc.) The system also has a mandatory “minimum verification time”, e.g. for a longer letter you will need more time to read it and can’t simply click okay without reading it. Nevertheless the doctor needs around 10-15% of their time for a report and overall satisfaction by the facilities receiving them has massively improved as they are more “standardized” in it’s format and they find the same information in the same parts every time.

For quality assurance a cetain degree of reports will be send “upwards” and the QA manager of the facility reads them as well, additionally some departments have implemented that if doctor A writes a report e.g. 5% goes to doctor B to check and vise versa.

Additionally the system also validates manually written reports (for training reasons med students and interns are not allowed to use it), e.g. “are you sure this patient had hypotension when admitted? He was given large amounts of a anti-hypertension drug within 5 minutes after being admitted”. This has significantly reduced errors in manual reports (we check a four digit number by now).

Whole thing runs locally (*) and does not need any outside connection at all and the model is open to the clients (they are actually encouraged to train it with their own data and to let their data security officers check it).

(*: This is actually an issue,sadly. As doctors reports are basically written all at the same time during the day and of course the model needs a fair share of ressources it needs a somewhat beefy appliance for it. While that is not an issue for a larger hospital it is not feasible for a smaller community hospital who barely manages it’s own IT infrastructure or even a small doctors office - but we found these to be the ones who would benefit the most from it. At the moment we don’t have a good answer for that beside hosting it elsewhere which would defeat the purpose and make it a privacy nightmare. We will see if these guys solve that before they fully introduce it into the market)

That saves time, money (even if it’s only for the larger facilities atm) and benefits patient care directly.