Research · The human body · Nov 16, 2026

Glucose an hour ahead, on days nobody had read

Met19 people, the last 30% of each record held out

The test at a glance

What it tests
Accuracy against reality: sensor glucose 60 minutes ahead, with the insulin and carbohydrate in between given.
Data
HUPA-UCM Diabetes Dataset: 25 people with type 1 diabetes, glucose every five minutes with insulin and carbohydrate; development on the first 70% of each record. 19 people record enough doses and meals to score.
Held out
The last 30% of each of the 19 records: 75,144 forecasts.
Compared against
The fitted body model alone; a learned model per person; a learned model across people.
Bar
Root-mean-square error at 60 minutes lower than all three.
Result
26.4 mg/dL against 29.5, 26.6 and 26.9: met. 30 minutes: 14.5 against 18.4, 14.7 and 14.7. Above the 12.5 mg/dL sensor accuracy.

For 19 people with type 1 diabetes, Living Worlds forecast sensor glucose an hour ahead on the last part of each record, which it had never read, given the insulin and carbohydrate in between: a root-mean-square error of 26.4 mg/dL over 75,144 forecasts, ahead of the fitted body model alone (29.5), a learned model per person (26.6) and one learned across people (26.9), and ahead of the per-person learned model in 16 of the 19 people. Each person has a small model of their own, six fitted numbers, and one correction is learned across everyone in that model's units. The margin over the learned models is small, and the error is above the 12.5 mg/dL accuracy of the sensor. Registered before the held-out parts were read, and registered knowing it had not beaten the learned model in development; scored once.

Protocol, result and scorer

Next

More people and longer records, people who make their own insulin, and exercise and sleep beside insulin and food.

A glass tree whose crown holds many small worlds like fruit

Which world will you live in?

What grows here is meant to reach the real world. Start with a world of your own.