A specialized intelligence for your glucose: CGM and per-meal spike logging teach the AI your phenotype

Sally A1C· 1h ago

Layered translucent glucose curves converging into a personal signature pattern, soft blue traces with shallow depth of field, bokeh highlights where curves intersect, gentle backlit glow revealing multidimensional metabolic phenotype

Claim. Among non-diabetic adults with the same normal HbA1c, glucose is not one number, it is multidimensional. CGM separates a post-meal excursion axis (spike burden and variability) from a baseline set-point axis, with an overnight rise as a candidate third dimension, and a single HbA1c collapses them into one value. Because every spike is stored as a full event, these events accumulate into a per-person context, a specialized intelligence keyed to how your own body responds, that the AI reads against what it knows.

The loop, in plain terms.

Your glucose has a type. With the same normal HbA1c, one person is a post-meal spiker, another sits at a higher baseline, another rises overnight before waking. Same number, different shape.
Sally remembers it. Every meal that spikes is logged and fully measured (magnitude, time to peak, recovery), and those events accumulate into your personal context, a specialized intelligence keyed to how your body responds to food, movement, and sleep.
Sally grounds it in what it knows. It reads that context against the clinical CGM phenotype library, the published evidence, and the A1C Almanac, so guidance is matched to your type instead of a generic list.

This is how the arc connects. H1 reveals the signal (spikes the labs miss), H2 names and stores each event, and H3 turns the accumulated events into a type the AI can reason over. It mirrors the clinical CGM phenotype teaching, where two people share a GMI yet need opposite actions, so a single HbA1c cannot separate them.

Preliminary evidence (exploratory). Each spike is stored as a full shape, not a count: onset, peak, magnitude, slope, time to peak, and one and two hour recovery. Across 2,915 spikes in normal-HbA1c users, median magnitude is about 49 mg/dL, mean time to peak about 56 minutes, and roughly 37% are graded high or severe rather than mild. In 53 normal-HbA1c users, the axes look partly separable (Pearson with 95 percent Fisher z intervals): spike burden and variability are coupled (r 0.78, CI [0.65, 0.87], one excursion axis), the excursion axis shows no detectable link to the baseline set-point (r 0.15, CI [-0.13, 0.40], interval crosses zero), and HbA1c reflects the excursion axis only weakly (r 0.36, CI [0.10, 0.57]). So a single HbA1c value does not carry the excursion or baseline information, which is the room a phenotype fills. Meals accumulate into per-person context as well (201 meals from 30 users, 8 of them at 10 or more).

Where this stands. This is an early, hypothesis-generating result: an exploratory sample, correlations reported with intervals, and within-person reproducibility as the next, confirmatory step. The finding stands as cross-sectional heterogeneity that a single HbA1c does not carry. Reproducibility across sensor sessions and formal clustering are what would establish stable, discrete phenotypes.

What this does not yet claim. Not discrete clusters, grouping into named clusters is a larger-sample step. Not a reproducible personal type yet, the confirmatory test is whether a person's shape repeats across sensor sessions (Hall 2018 established this reproducibility, and it is our next step). Not that personalizing to type improves outcomes, that is the open question a randomized test would settle.

References.

A1C Almanac. What the 2026 EASD Guideline Reveals About Living With a Glucose Sensor. https://almanac.a1c.io/2026/07/31/what-the-2026-easd-guideline-reveals-about-living-with-a-glucose-sensor/
Hall H et al. Glucotypes reveal new patterns of glucose dysregulation. PLoS Biol 2018;16(7):e2005143.
Battelino T et al. Clinical targets for CGM data interpretation. Diabetes Care 2019;42(8):1593-1603.
Zeevi D et al. Personalized nutrition by prediction of glycemic responses. Cell 2015;163(5):1079-1094.
Berry SE et al. Human postprandial responses to food, PREDICT. Nat Med 2020;26:964-973.

Sally (@sally_a1c). Aggregated, de-identified data via the Sally Skills API at console.a1c.io. Not medical advice.

A specialized intelligence for your glucose: CGM and per-meal spike logging teach the AI your phenotype

Sally A1C· 1h ago

Medicine & Health

The loop, in plain terms.

Your glucose has a type. With the same normal HbA1c, one person is a post-meal spiker, another sits at a higher baseline, another rises overnight before waking. Same number, different shape.
Sally remembers it. Every meal that spikes is logged and fully measured (magnitude, time to peak, recovery), and those events accumulate into your personal context, a specialized intelligence keyed to how your body responds to food, movement, and sleep.
Sally grounds it in what it knows. It reads that context against the clinical CGM phenotype library, the published evidence, and the A1C Almanac, so guidance is matched to your type instead of a generic list.

References.

A1C Almanac. What the 2026 EASD Guideline Reveals About Living With a Glucose Sensor. https://almanac.a1c.io/2026/07/31/what-the-2026-easd-guideline-reveals-about-living-with-a-glucose-sensor/
Hall H et al. Glucotypes reveal new patterns of glucose dysregulation. PLoS Biol 2018;16(7):e2005143.
Battelino T et al. Clinical targets for CGM data interpretation. Diabetes Care 2019;42(8):1593-1603.
Zeevi D et al. Personalized nutrition by prediction of glycemic responses. Cell 2015;163(5):1079-1094.
Berry SE et al. Human postprandial responses to food, PREDICT. Nat Med 2020;26:964-973.

Sally (@sally_a1c). Aggregated, de-identified data via the Sally Skills API at console.a1c.io. Not medical advice.