[cperciva]

The sensors aren't perfectly accurate; I don't know about the newer sensors which this closed-loop system uses, but the current generation only fulfills the accuracy requirements by reporting the median of 7 consecutive every-5-minutes measurements and carrying a warning of "don't trust these values if you have reason to think that your blood glucose might be changing rapidly".

My guess is that providing notifications of exercise and food (aka. "I expect my blood glucose to go (down|up) right now") allows the model to be tuned to respond more aggressively to new data at some times and less aggressively at others.

(The "less aggressively" side is probably most important: If the sensor suddenly reports a dramatic change in blood glucose in the middle of the night, the system should probably respond with "HEY WAKE UP THE SENSOR IS BROKEN" rather than "let's kill the patient".)

[shanselman]

Exactly. Not to mention they are testing interstitial fluid and not whole blood, which introduces significant delays.

[cperciva]

I've found that the interstitial delay is pretty minimal; maybe 5-10 minutes behind fingerpricks (which themselves lag somewhat behind arterial blood, especially if your hands are cold). The dominating factor is the 15-20 minutes it takes before the median of a window reflects the new data.

[shanselman]

Sure, but we agree every minute counts. Also, CGMs in arms differ from belly, etc. 10 min when insulin starts in 30 is a non-trivial amount of time. Point is, we def need faster insulin and faster, more accurate CGMs.