Show HN: GlycemicGPT – Open-source AI-powered diabetes management

The risk to benefits ratio of introducing a language model to interpret so clear signals is nowhere near justified.

Monitoring and analytics is important, but it is a solved problem. A language model will only be able to hallucinate about the relationship between meals and glycemic response. At best it does no harm, at worst it can directly misinform.

Interesting. I can see the utility if you're going to see a nurse practitioner. But if your physician doesn't pull the actual charts for your device and visually inspect them.... try finding someone else.

Does it prompt logging? For example, when I was trying to monitor my BG after diagnosis, I tried to log my meals to correlate later, but 1) would forget and 2) wouldn’t have the energy to time align the stats. So a tool that even saw changes in BG and shot me a text or message (did you eat/exercise do something @ [time]?) and used the LLM or something else to capture and enrich the metadata. Paired with boring things like med reminders (I just realized I forgot my metformin while typing this) and giving me an easy visualizer with these meta points would be useful. If I’m tracking sleep on a device etc.

As others have said, the analysis might be risky. I don’t want to trust interpretation to anyone but myself (bear my own risk) or my clinician. But just remembering to capture the data and making it easily time alignable and possible augmentable in the future would be useful.

I'm a T1D who has an insulin pump looping with AndroidAPS and NightScout, what does this give you that Nightscout and Autotune doesn't give you?

And how do you deal with AI hallucinations?

I don't think that LLMs are trustworthy companions in managing a complex metabolic disease like diabetes - especially if you deviate (ever so slightly) from the norm (very lean, very active, strict diet, etc.)!

I'm a T1D myself and like to experiment with ChatGPT (or Opus). My experiences are mixed

LLMs are overly cautious when it comes to correcting with insulin. They regularly advise against correcting before going to bed, even if this means that my blood glucose remains above 140 mg/dl for the whole night.

I am following a low to medium carb diet (<100g a day). ChatGPT always nudges me to consume more carbohydrates, even though I have a TIR of 90% (70-150 mg/dl). Why would I change my diet if it currently works very well for me? Still, most LLMs seem to favor carbs for some reason.

I am using Fiasp as my fast acting insulin. Typically, I inject around 1 to 4 IUs of Fiasp. Its glucose-lowering effect typically lasts for roughly 2-3 hours. Therefore, I know that it is safe to re-inject after three hours without risking insulin stacking. But ChatGPT regularly advises against that and wants me to wait another 1-2 hours.

I am not against automating diabetes management. In fact, I really appreciate projects that help with that. But I don't consider LLMs to be helpful in this regard. Their combination of training data bias, liability aversion, lack of context, and one-size-fits-all thinking disqualifies them from such tasks.

I'm a T1D and tbh it's not that hard to manage, I just wouldn't need that. But for kids or the elderly, I see a use case.

The hardest to learn was that an unhealthy lifestyle resulted in a diabetes that was harder to manage. Too much carbs, not enough exercise, etc. After adjusting my lifestyle, it became quite easy.

The most pain, in my experience, comes from the discrepancy between the CGM - measured value and the prick-test value, even when accounting for time lag. I've used several CGMs and they've all been wildly off sometimes. I have a few T1D acquaintances who relied on their CGM alone and have significantly improved their HbA1c after accounting for that.

Maybe that information is useful to you.

This is quite possibly a horrible idea. Personal anecdote: ChatGPT once read a blood work report value as 40, when the actual report said 4.

As a urologist who built and runs his own clinic management software, I'd encourage thinking about this question early: what does the system do when the LLM refuses to answer, returns malformed JSON, or hallucinates a glycemic value? In medical contexts, a 'silent failure' (system continues despite bad data) is much worse than a noisy failure (system stops and asks the user). The 'happy path' for an LLM-powered medical tool is usually well-designed. The failure paths are where the project lives or dies. Curious how you handle that.

Really nice of you to share this, well done!

About the risks, managing type 1 diabetes is exhausting, and most people will still sanitycheck the output alongside the hundreds of treatment decisions they make every day. That doesn’t change the fact that tools like this can nudge you to notice and look into patterns or things that needs attention.

You know that current AI systems are not reliable and produce errors?

How do you protect your life and the life of others using your software against potential lethal errors?

"This will all end in tears, I just know it"

Marvin

The alerts system and sharing with caregivers is a solved problem already (e.g. Dexcom's Follow, Abbot's LibreLinkUp).

Do you find the analytics actually helps? I.e. a lot of this will depend on what you ate and whether or not you logged it?

Looks interesting, being a Whoop user for the last few years, I have seen for myself that their AI Coach/AI based suggestions are a hit or miss 3 out of 10 times, slightly concerned about how accurate this will. Not a diabetic patient, but I do monitor my levels with a CGM from time to time, will definitely check it out!

What’s the limit on badges in a README

Went through pregnancy with the mother having recently-diagnosed T1 diabetes – just barely not killed by grave neglect on behalf of healthcare due to how badly they missed the diagnosis to begin with.

On your work:

this is legit

it is appreciated

Hats off, I salute this, thank you

I mean this in the nicest way possible.

But if someone dies because this thing hallucinates their reporting - would you feel any sense of culpability?

“GPL says no warranty”

“People need to double check LLM output”

“You’re holding it wrong”

I really don’t know if we, collectively as a civilization, should be willing to accept this kind of hand-waving when it comes to creating things like this. Sure, tools make mistakes or people misinterpret reports without the help of LLMs - but LLMs are just on a whole other level where the mistakes are just part of how these things work from a fundamental level.

I don’t even trust AI scribes at my doctors office to transcribe my appointment due to errors. There is no way in hell I would ever use something like this that could just straight up lie about something that kills me if I get it wrong.

This is THE ONE domain where you would want to use classical machine learning and not unreliable LLMs. Unless you want to kill yourself, that is.

I've done this with the Libre 2 sensor. I added Gemini to it. It gets like 2 weeks of readings at once, and the user can "chat to their data". I added a meals tool as well, where the user can photo their meal, and the ai estimates the impact on the readings.

It's so helpful to offload some the thinking about the condition to ai, all these people moaning about 'muh safety' don't get it. T1D suffers have to think about it all day all the time. A person doesn't have their own blood glucose data in their head.

Life imitates comedy...

I'm just happy to see a GPL project.

So, I'm in the medical field building an EMR and LLMs have obviously been a really important topic in the industry the last few years. We're still not even sure that giving LLM-assisted suggestions TO ACTUAL DOCTORS AND CLINICIANS will be helpful let alone to the patient themselves.

It's breaking the golden rule of these tools which is to have someone with enough knowledge to verify the accuracy of the data it spits out. Patient's famously don't. Hell, even the actual staff don't really understand or know how these tools work (or the ways in which you can/can't trust them).

FDA approved?

[flagged]

[flagged]