I'm curious how this problem might be approached. In a nutshell, and simplifying things a bit: for diabetics glucose levels have to be controlled manually. Eating carbohydrate / sugar raises blood glucose levels, and injecting insulin lowers. So if blood glucose is a 5.6mmol/l and 50g of carbs are eaten, 5u of insulin might be taken to balance out the carbs.
In practice, the ratios between carbs and insulin is different at different times of day, and drifts over time. Diabetics have to discover these ratios, and when the effects of doses change, have to review and revise them.
Also as well as doses with meals, there's a long acting (24 hour) daily injection of insulin taken that provides a background. There's a rule of thumb that this should be 60% of the total daily doses. This prevents levels from changing too much at meals - a small miscalculation can have a large effect, and if blood sugar goes too high or low, there can be difficult consequences.
So, I'm curious how this might be modelled. Ideally, a model would provide a way to recommend optimal ratios and detect changes in ratios, also it'd suggest when to check blood sugar - tests are done manually, typically between 4 and 8 times a day.
Finally, a disclaimer, because I feel I should say this. I'm just curious about how the machine learning community would look at this. I'm not seeking medical advice, and will not treat any information or thoughts shared as medical advice.
[link][4 comments]