GPT healthcare startup ideas
Get Out-Of-Pocket in your email
How to Build a Healthcare Call Center
A Call For Startups
Previously, I’ve done some “what’s a startup you wish existed?” posts asking people to submit their ideas for a future newsletter. I’ve gotten some pretty good ones from here and know for a fact that a couple went on to actually get built.
It’s been a while so I thought I would try it again.
Send me a short (2 paragraphs or less) blurb about a startup you wish existed. Don’t pitch me a startup that you’re working on, have raised money for, etc. This is meant for new ideas where you’ve seen a problem and wish someone built a solution for it.
I’ll pick a few of my favorites for a future newsletter. To get you started, here’s an idea I’ve been thinking about. Your idea does not need to require a large language model to be submitted.
Hyper-Specific Search In Healthcare
I’m still trying to understand what large language models can and can’t do. But it feels like the absolute perfect piece of technology to create a shift in healthcare - an industry known for massive amounts of data, laden with jargon, and hard for patients to interface with. There are already some companies like Science.io, Abridge, etc. that I’ve written about in the past that have been leaning into these large language models (LLMs). But it feels like the technology can really be pointed in a million different directions.
One I’ve been thinking about is search. A big issue with search today is that it’s hard to specify search over a source of data. For example, Reddit’s search functionality is so bad, I assume whoever made it also created provider directories and depression. In order to search something on Reddit, the best method is Googling “______ site:reddit.com”. I use this all the time.
But what do you do if you want to search across subjects that aren’t neatly under one URL? That actually starts getting pretty difficult. And what if instead of needing to click through links, sort through the legit answer from not, the answer was neatly presented to your query?
One of the pretty amazing things about these new large language models like GPT is that the models have been trained on massive amounts of text already, so learning a new task needs very little additional data and tuning. I’ve been watching people take these AI models and tuning them on specific datasets to surface answers. For example, here's one from Humata that lets you do a Q&A of a study pdf. BioGPT seems like it’s trained specifically on PubMed abstracts. And uh…other ones…
I think there are several small projects (maybe companies) to be built that have specifically curated data sources underneath the model that give users more confidence in the results.
- ChatRFP - A structured repository of different healthcare RFPs that’s easy to query (“which RFPs have higher weights for maternal health”, “given we have these features, what kinds of RFPs should we prioritize”). Medicaid RFPs in particular are very state-by-state and quite long.
- ChatPayerProviderManual (I’m not good at naming things, whatever) - Each insurance carrier has a variety of “provider manuals” which basically tell you everything from what you need to be credentialed, how to bill for services, what you need during a prior authorization, etc. Here’s an example of Aetna, Florida Blue Cross, and NY Medicaid. Each of these are nearly 100-page PDFs that absolutely no one has time to read and are usually ctrl + f’ing for some specific thing. Payers spend a TON of money in admin just responding to clarifications that are in these docs. Layering a much easier-to-use text interface across all of the payers that a provider can query would be a huge time save.
- New search for literally any healthcare government website - I dare you to use the search function for any website the government operates without wanting to commit slight tax evasion. FDA.gov is just a mishmash of PDFs, articles, transcripts, etc. without a good way to find the right information. Searching clinicaltrials.gov will actually make you eligible for TBI trials. The national/local coverage determination search is designed to look up a specific determination vs. understand things across them. I think you could just dump all this text info into a large language model, let it do the explaining, and point you to the underlying primary source for confirmation on anything specific.
- Open enrollment assist - Digest the insanely long pdfs, explanation of benefits, formularies, etc. that all of these health insurers put out publicly to assist patients figuring out what plan they should get. Honestly lots of other entities would probably be interested in that data, too. I just want people to stop texting me, please.
Honestly, anything that’s a bunch of long-ass free-text PDFs should have a chatGPT-like layer over it that makes it easy to interrogate the underlying report. No one has time to read 100+ pages that all start with “we aim to provide the best patient experience” and then 10 pages of disclosures when they need the answer to 1 or 2 questions that are somewhere in it.
It might even incentivize people producing shorter reports to reduce the fluff. And that’s coming from someone that writes 10+ page emails.
All of the above is for publicly available data. I’m not even sure if these are standalone businesses, a wedge or feature of a different business, or just projects that anyone can build. But I think they’d be interesting pain points that large language models can solve.
The next frontier though is being able to use these large language models locally…
One issue with using large language models is that the stuff you put into them goes to the deities above us in the cloud. Basically, it’s not very privacy-preserving, which is tough when you’re dealing with personal health information or sensitive company information. I don’t want the cloud to know I'm worried my pee is blue, just my newsletter subscribers.
But what if you didn’t need to reveal sensitive data to these large language models? It seems like we’re moving in that direction:
- LangChainAI just released a github repo + contest to see other local repositories of data you could use large language models over.
- Azure seems to be pushing for OpenAI services you can use on data locally.
- Biohacker types are already using chatGPT to answer questions using data from their own body.
- ScienceIO uses large language models to redact PHI so you can use your own local data with cloud-based services.
- I’m tweeting about it, which means the opportunity to make money here has passed.
I’m interested to see what kinds of businesses get built when you can use PHI and internal company data. Idk what’s possible, I’m just riffing here but:
- Instantly create standard operating procedures (SOPs) based on historical occurrences or plain-text write-ups of an activity. Bonus points if the AI prospectively guesses what things might happen and generates those SOPs.
- Make it easy for anyone, not just the legal team, to review contracts in plain English.
- Let anyone query internal databases by writing out regular sentences that get converted to SQL mapped to your internal schemas on the back-end. (Hex is getting pretty close to this, and disclosure: investor.).
- Listen to patient screening and outreach calls and automatically conduct script optimization (maybe even real-time while with a patient?). Listen to telemedicine visits and see if physicians are deviating from protocol.
- Auto-generate passive-aggressive emails to coworkers when deadlines aren’t met that asks for updates on projects as if you don’t already know what the answer is.
Again, this is a very new area for me, but I figured I’d throw shit against the wall and see what sticks/inspires people. I think we’re still probably far-ish away from totally automating patient facing tasks, but the back-office is fair game. There seems like a bunch of new classes of startups to be built on top of these models, and it’s happening quickly.
Once again, I want to hear some of the other ideas you’ve been thinking about in healthcare. Send me a short (2 paragraphs or less) blurb about a startup you wish existed. This is meant for new ideas where you’ve seen a problem and wish someone built a solution for it.
Nikhil aka. “about to be even more unemployed somehow”
Other posts: outofpocket.health/posts
Thanks to Will Manidis and Katie Link for reading drafts
If you’re enjoying the newsletter, do me a solid and shoot this over to a friend or healthcare slack channel and tell them to sign up. The line between unemployment and founder of a startup is traction and whether your parents believe you have a job.