Context A friend of mine has a very unstable form of diabetics, not sure how much insulin to take is his daily struggle. I once got him reasonable stable using manual fuzzy logic calculations (as in your thermostat), but the hospital advised us not to keep him around 4.0 our goal should have been 6.0, ( the 4.0 calculation was extremely complex and partly guessed but worked) for goal 6.0 I hope maybe the community can help using a neural net. Content The data currently contains a period of about 4 weeks and is nearly complete. There is some missing data, I would not recommend filling this missing data, for the moment. Since he's not so predictable, I think a neural net might predict well in those cases where the insulin worked well Maybe add columns for good behavior day, and perhaps you could find some patterns. There is a more detailed description in the text file. Acknowledgements We hope to achieve some calculated advice here, you might train 3 separate neural networks for each main insulin moment (morning midday evening), or create an LSTM perhaps. I put the data into a CC0 license in return, we politely ask you to comment us back upon result/advice. ea provide your kernel with some explanation. Inspiration What I would like to create is a lookup table for him so he could simply and without depending digital devices. To simply have a good indicator of advisable insulin intake to use, in any situation (even if your offgrid without a phone).. Be aware his blood glucose measurement should not drop below 3, while getting above 10 isn't good either. In rare situations, he might get over 10 for short period of the day, it's not as risky as going below 3. Real low values 1 and 2 are potentially deadly, because they can go in shock. In the past there was a lot of such data online, these days most of that data has been removed from the internet because of privacy concerns. We don't believe that's good. However I have studied the old data (i can do such things) but then I also noticed the old research data was of a very poor quality. My friend's data is much more complete as he has much more daily measurements, in which the effectiveness of a dose can be seen. My friend during workdays (not in the weekend) lives by a very strict lifestyle, and we can assume he does do the same work (using about the same amount of energy on a daily workday). This numeric problem should be solvable without keeping track carbohydrates usage as he eats the same meals at work, and in the evening the same amounts as well. Kind of food isn't a huge impact he's not overweighted either. The dose values are "additional" manual correction doses on top of an automated dose (and so since we will not change the automated dose we can do without the automatic data as well. *Finding a strategy could help others as well !* Ranking There is no ranking here, its science in real live progress. We will judge derived advice from the data, and if we think it's safe we will test it. (for a week at first) Eventually, though it is up to my friend to choose an insulin dose, (though if he differs a dose that will generate data too). Data updates This is a work in progress, as told earlier, during this research I will put more data online (monthly or weekly), when it comes availble (and no good solution is available) I will try to update project information if people have additional questions. Ideas, hints I think the morning data (plus maybe the previous evening) is a good start to base models upon. Since morning data is the least disturbed by work activity (ea work energy usage, in the form of burning sugars). Whatever happened to your fuzzy logic, and what you do in the meantime ?. Well currently we try out advice tables and adjust those fuzzy, but despite the slowly improving results i'm not convinced that a single table is the best solution for this. (can be backup for offgrid) but eventually i think the data can act like a neurel net based model advice. I also hope my limited time on kaggle can attribute to getting more insight into the data