Type 1: Would you be willing to share your diabetic data for research?

Hi all,

Me and a friend are going to do our MSc. thesis next semester. We are both studying computer science and want to explore the possibilities of developing some kind of decision support system to help diabetics with their everyday medical decisions. I have the disease myself and know that I, for one, certainly would love to have my computer or smartphone tell me how to adjust my bolus and basal rates to get better glucose values. Preferably without ever having to calculate and enter how much carbohydrates I eat.

The idea is that we, with the help of machine learning, for example would be able to infer how a typical breakfast/lunch/dinner scenario might look like and suggest therapeutic changes thereafter. Many diabetics (including myself) can adjust insulin dosage with good results without the tediousness of calculating carbohydrate intake. Knowledge gathered from previous experiences usually goes a long way. For example, if I notice that my glucose levels always tend to be way off after lunch for 5 days in a row I usually have some kind of intuition of how to adjust my insulin dosage to avoid the problem the next day, even though I never do any carbohydrate calculations. Our hope is that a computer could learn this "intuition" as well, and maybe better than us. Analysing months of glucose readings is easy for a computer and not so easy for a person.

To do this, of course, we need user data. So far I only have my own Freestyle Libre CGM reading + manually entered insulin dosages for maybe 3 months back. The most important question we have is: would you be willing to share any of your (anonymised) data, to help us with our MSc. thesis?

And while we're at it: it would be interesting to know what kind of glucose meters you use, whether you record your insulin dosage throughout the day and how/if you gather all your data somewhere (in cloud service etc.).

Feel free to answer here or send me a direct PM. Our work will start in the middle of January next year and we greatly appreciate any help we can get.

Edit: My initial explanation and motivation of the research we aim to do seems to have caused a bit of confusion. We are not going to develop a consumer product out of this. In fact, we are most probably not going to develop a product at all. However, what we will do might be useful in consumer products, should we be successful.

The thesis will be research oriented with the purpose to evaluate additional machine learning methods to the ones already being explored by others, and in this process we might very well need carbohydrate intake data as well. In fact, we don't limit ourselves to any amount of data features, quite the contrary; the more features the data contains, the better.

TorqPenderloin said:

Sounds very similar to Sugar.IQ the app being developed by Medtronic and IBM Watson.

Best of luck to you.

Click to expand...

Thanks!

Yes, there are several existing decision support systems. However, this would be more focused on research about what particular machine learning methods are most effective etc., rather than building a complete consumer product. In particular, we aim to investigate the use of recurrent neural networks for predicting glucose values and possibly also detect excessive glycemic variability.

azure said:

How would it work without knowing the carbs @simonjrp ?

i agree there's a level of intuition and knowing your own body, but carb counting is crucial to good control, especially if your meal is a new one to you.

Click to expand...

You're right. For a super accurate system that basically could automate the process of insulin injections, knowing how many carbs you eat is a must. But it might be possible to develop a system that at least detects the problematic sections in your daily CGM plots and classifies them correctly for you without any knowledge of carbohydrate intake. The suggestions for insulin adjustments that such a system could make would probably just work as "indicators" rather than "do exactly this without thinking yourself first". For people that does not count carbs, it could still be an improvement.

The intention with my description of a decision support system that could help diabetics with their insulin dosage was mainly to explain possible uses for our research, without going into details. We are not going to develop a fully working consumer product. In fact, we're not even going to build a product, but rather investigate alternative methods to the ones that have already been evaluated.

For example, there is a paper https://etd.ohiolink.edu/!etd.send_file?accession=ohiou1382664092&disposition=inline that, among other things, tries a few different machine learning methods to accurately detect excessive glycemic variability. One of our objectives is to try doing the same, but with other algorithms to see if we can get better performance and new valuable insights.

tim2000s said:

It's an interesting point of view. The alternative hypothesis to this approach is that if they had counted the carbs, instead of waiting five days with high glycaemic variability and higher than necessary blood glucose levels, it would, instead, have taken a day to improve? I guess the question for me is how do you quantify "good results"?

As a user of a hybrid closed loop system that can and does automatically adjust insulin amounts based on real time data, and having a decent understanding of the algorithms involved, and the drivers for making decisions around this, I'm not sure how, without knowing some fairly crucial pieces of information, a machine learning algorithm is going to provide a reasonable adjustment dose?

Click to expand...

No, we would not be able to develop something that performs as good as such a system without knowing things like carbohydrate intake etc. However, we might be able to improve the performance of some of the smaller building blocks of the closed loop systems that are out there. And, I think it might be worth investigating how much can be done with just the "intuition", even though it by no means would replace existing systems.

tim2000s said:

Interesting. Those algorithms included meals in their data used for future blood glucose predictions.

Click to expand...

For trying to improve on the performance they got for glucose prediction in this paper, we will indeed need data about carbohydrate intake as well.

I realise now that my initial motivation might have been a bit misleading. The 'device' that I described was basically just a way of indicating the usefulness of research within the area. One of the things we might investigate is how good problem detection and/or insulin dosage suggestions one can get without knowledge about carbs, but we also aim to research other questions where we need as many extra data features we can get, including carbohydrate intake. I'm sorry for the confusion. I've updated the first post to better reflect our intentions.

nickm said:

Data mining without pre-specified hypotheses.
And perhaps you would like to comment on what happens to allegedly anonymized data at some universities.

Click to expand...

We do have fully specified hypotheses. I just haven't specified it fully here because I don't want to "give away" a complete MSc. thesis idea. The point is that we are going to do research that could help in the treatment of diabetes, and the more patients we can generalise our models to, the better. The research can't do any harm to anyone, and considering the sheer amount of people that have this disease, any kind of improvement (however small it might be) must be worth investigating. That's my opinion, and if anyone feels the same and have data they are willing to share, that's great. If not, we'll just have a less generalised, but hopefully still meaningful model.

If someone feels that they need to know a bit more about this before they give a way their data, that's very understandable. I know I would. And of course we will elaborate on the details for those who are interested, but preferably in a private conversation.

As for the handling of anonymised data I have a hard time seeing how glucose value readings that aren't possible to connect to any person whatsoever can do any harm. But maybe I'm missing something obvious. If someone has privacy concerns we could probably construct some kind of privacy statement together that both parties agree on. The question here is just "are you interested?". If the answer to that question is yes, reach out to me in a pm and we'll take it from there.

Julian_Hands said:

I'm happy to share data after extensive work I also am carrying out with the effects of sport with my diabetes.
I agree with earlier posts that those who have the access to CGM and pump technology have already the solution in their hands, those who really struggle (especially like me) are the majority who are unable to use algorithms for their basal control and are stuck in the 24hr loop once their (in my case Lantus) basal has been injected yet may have exercise patterns that are irregular and get severe hypoglycaemic effects due to the basal rate required.
I'm currently using an Accu Check Aviva Expert meter to control my bolus(Novorapid) levels with close carb counting. But there is very little public detail available for the effects of sport (in my case cycling) and the basal rates to take.
As an example. If I commute to work 4 times a week on my bike (30km hilly each way,60km per day) the the first day I need to reduce my basal level the night before by 10%, if I continue the following day I need to reduce by 20%, on the third day if I continue to commute back to back days I need to reduce my basal levels by a further 10%, now 30% total reduction, due to recovery and my body replenishing glucose to the muscle groups.
I have to test virtually hourly to ensure my levels are maintained and often end up snacking to ensure I maintain my ranges with glucose.
The NHS In the UK doesn't support CGM technology and pumps are hard to obtain, but more support is defiantly needed, especially for individuals who wish an active and unexpected activity lifestyle, yet get held back severely by their insulin regimes that they have very little ability to adjust once taken the night before.

Click to expand...

Hi Jullian, and thank you for your interest!

I agree, the ones having CGM and insulin pumps already more or less have the solution at hand. However, this solution can always be improved and with the rise of fully automatic (closed loop) systems, there's much to explore.

Since I wrote the original post, our project has been narrowed down to "just" predicting future glucose values based on earlier data. To do this accurately, of course, working on CGM measurements helps a lot. Our main focus will be on using CGM data, but it would be interesting to see what kind of predictions one can make using fingerstick data. The frequency of your fingersticks and all extra life data would probably be of a great significance here.

The effects of sports is indeed an interesting aspect of diabetes, and something I personally also would like to know more about. I definitely recognise the tedious and often very hard work that is adjusting the basal rate to physical activity (even though I have an insulin pump). As of now, we don't have any data from especially physically active patients, so your contribution could certainly help. However, since it isn't continuous, it will probably not be our main focus.