How we measure it

How we measure your response

Everything GlucoBeat shows you comes from your own data through calculations you can read here: which formula we use, which time window we look at, how much data is needed and what happens when it is missing. No black box.

Three rules every number follows

It describes what happened in your data; it does not predict what will happen.
It comes with its context: how many times it was seen (n), in what range and with what confidence.
It speaks of associations (“after X came Y”), never of causes: food, exercise, insulin, medication, sleep, time of day and stress all act on your glucose at once.

Period statistics

They are calculated over all readings in the chosen period, weighted by the time each one covers (at most 20 minutes, so that a sensor gap does not count as if there were data).

Time in range (TIR): Percentage of time with glucose inside your target range. We show it with your range (the one you set in Settings) and also with the international standard of 70–180 mg/dL, so you can compare it with what your care team uses.
Time below and above (TBR, TAR): Percentage below the bottom of your range (and, separately, below 54 mg/dL) and above the top (and, separately, above 250 mg/dL).
Mean and variability (CV): Weighted mean of the period and coefficient of variation: standard deviation divided by the mean, as a percentage. The lower, the steadier.
GMI: Glucose management indicator: 3.31 + 0.02392 × mean in mg/dL (Bergenstal et al., 2018). It is an estimate from the mean and does not replace a lab HbA1c.
Episodes: A stretch counts as an episode (low or high) when glucose stays out of range for at least 15 minutes in a row, with no gaps longer than 30 minutes between readings.
Coverage: Fraction of the period with sensor data. When it is low, the period's figures are worth less and we tell you.

Ambulatory glucose profile (AGP)

It is the standardised format used by sensor reports: every day of the period overlaid on a single 24-hour day.

How it is built: Readings are grouped into 15-minute slots by local time and each slot is smoothed with its two neighbours on each side (±30 minutes). For each slot we draw the 5th, 25th, 50th (median), 75th and 95th percentiles.
When it is not drawn: A slot with fewer than 5 readings stays empty: we do not invent the curve.

Response to an event

Every meal, workout, dose, medication or sleep you log is an event. GlucoBeat looks at your curve over the following hours and describes what happened.

Starting level: Median of the readings in the 30 minutes before the event. If there are none, the value interpolated at that moment (only if there are readings less than 30 minutes away on each side).
Window we look at: It depends on the type: meal 3 h, exercise 4 h (or the length of the session if you logged it, then what happened after it ended), insulin 4 h, medication 4 h, treating a low 90 min, sleep 8 h, waking up 3 h.
Rise, peak and time to peak: Difference between the highest value in the window and the starting level, and minutes until that maximum. The same with the minimum (drop) for exercise and insulin.
Return to the starting level: Minutes until the first reading, after the peak, that is within 10 mg/dL of the starting level. Only calculated if the peak rose more than 20 mg/dL above that level.
Out of range: Whether any point in the window went above or below your range.
Overlapping events: If there is another meal, workout, dose or treatment of a low between 60 minutes before and 120 minutes after, the response is flagged as “other events nearby”: what happened cannot be attributed to that one alone.
When it is not shown: If the sensor covers less than half of the window, there is no response to measure and no figure is shown.

Usual response and similar meals

When you have logged the same meal several times, or similar meals, we pool their responses so you can see what repeats.

Which meals are pooled: The whole text of each meal is compared, never single words (“chicken” and “rice” from one “chicken with rice” are not two foods). In Statistics, two meals are the same if they share at least half of their meaningful words (ignoring “with”, “of”, “plate”…), even if a detail changes; the group is labelled with the shortest text. In the meal sheet, meals similar to what you type are found by trigrams and by words in common, with a minimum similarity of 0.3.
What is calculated: How many times (n), the median rise, the usual range (25th and 75th percentiles), the usual time to peak, the return to the starting level and what fraction went out of range. And, when there are enough, the same comparison splitting the times with exercise in the next 2 hours from the times without it.
Minimums: At least 2 measurable responses are needed (with starting level, peak and enough coverage). With fewer, nothing is shown.

Experiments

An experiment is a comparison you choose between two situations, over what you log from the moment you open it. GlucoBeat splits the observations into two groups and aggregates each one with the same calculation as the usual response. No AI.

Meals: Exercise after breakfast, lunch or dinner (a session starting within the next 2 hours), exercise before or after a meal (2 hours before or 2 after; meals with exercise on both sides are left out), breakfast or dinner from a time you choose, dinner shortly before bed (under 2 hours until the sleep entry), light or heavy dinner (from a number of grams of carbs you choose, only dinners with carbs logged), a word in the meal (versus the other meals in the same parts of the day) and two meals head to head (by text similarity, as in the meal sheet). For each group: median rise, usual range, minutes above your limit and return to the starting level.
Exercise sessions: Fasted (no meal in the previous 3 hours) or after eating, short or long (by a duration you choose), before or after a time, light or moderate versus hard (by the logged intensity), or two kinds head to head. Each session is measured from its start, with the starting level from the previous 30 minutes, until 2 hours after it ends (capped at 4 hours; 4 hours when the duration is unknown), to capture the drop during the session and the one afterwards. For each group: median change (the lowest point relative to the starting level), usual range, minutes below your limit and how many minutes in the lowest point arrives.
Whole days: Weekdays and weekends, days with logged exercise and without, days after a night of 7 hours or more versus less (with sleep and waking logged; the night is assigned to the day you get up), early mornings (0:00–6:00) after a dinner less than 2 hours before bed versus those after a dinner with margin, and early mornings after eating something with a word you choose in the hours before bed (you choose those too) versus early mornings without it. A day counts if the sensor covers at least half of it (of the early morning, in that last one); today, incomplete, is left out. For each group: median time in range, its usual range, the mean and how many nights (0:00 to 6:00) had a low. And whole weeks, Monday to Sunday: those with 5 exercise sessions or more versus fewer, with the same figures and the early mornings with a low added up; the current week is left out.
Minimums and confidence: At least 3 measurable observations are needed in each group to say anything; until then the sheet shows how many each group has. Each group's confidence follows the usual-response rule and the experiment's is the lower of the two. In splits by exercise, having exercise nearby is the condition and does not count as contamination.
Conclusion: A template with the two medians, the total number of observations and the confidence, always phrased as an association in your data, never as a cause or a recommendation. It is recalculated every time you open the sheet with what has been logged so far (180 days at most).

Confidence

Each aggregate carries a confidence label that comes from three things: how many observations there are, how many are clean (no other events nearby) and how much sensor coverage there was.

High: 5 or more observations and at least 60% clean.
Medium: 3 or 4 observations.
Low: 2 observations. A hint, not a pattern.
The AI's: In AI results, confidence is declared by the model itself according to the data it received, and it is shown as is, together with whether it relied on your history, on general knowledge or on both.

What the AI does and what it receives

The AI does not see your raw data and does not decide anything for you. It only comes in when you ask, and it receives a pseudonymised summary: no name, email or identifiers, with dates as relative days and times.

Period analysis: It receives the statistics, the 24-hour profile, and each event with its measured response as explained above. It is asked to describe associations in your data (“in your data, X was followed by Y, n times”), to state confidence and data quality, and never to talk about causes, doses or diagnoses.
“How would this affect me?”: First you see your usual response, calculated without AI. Then the AI receives that same usual response, your recent and similar meals with their responses, your glucose over the last 3 hours and what you have done in the last 4, and returns an estimate with a range, labelled as an estimate, with its confidence and its basis. Never a permission verdict or an exact figure.
Guardrails: After the answer is generated, a filter removes any sentence with dose amounts, treatment adjustments or diagnoses, and tells you it was removed.

What we do not do

We do not predict or forecast: the figures describe your past and the estimates come with a range and a label.
There is no clinical validation of these calculations as a medical tool: they are descriptive and indicative, to understand your data and arrive at your appointment well informed.
We do not raise alarms or alerts on glucose values, and we do not work in real time: your sensor and its app do that.
We do not give doses, ratios or treatment adjustments, nor diagnoses. Ever.
We do not claim causes: what we see are associations in your data, which may be due to many things at once.

References

Check it with your own data

Every figure you see in GlucoBeat can be traced back to this page. If something does not add up, write to us.

Start for free Write to GlucoBeat