A Better Way to Read CGM Without Diabetes

For people without diagnosed diabetes, what can a CGM trace actually tell them? The evidence supports a precise answer rather than a reassuring or alarming slogan. A sufficiently informative recording can describe glucose dynamics through average level, variation, and short-lag temporal structure. Those features may be associated with insulin-related measures and early vascular or liver markers. CGM can also provide useful feedback inside a structured lifestyle program, particularly for some short-term glycemic, dietary, and adherence outcomes. It does not, by itself, diagnose diabetes, establish a universal personal risk threshold, or guarantee weight loss.

That answer keeps four kinds of evidence in their proper places. Description concerns what the trace did. Stratification concerns how patterns relate to measured outcomes in a research population. Feedback concerns whether seeing a response helps a person and a program change behavior. Diagnosis concerns whether validated clinical criteria support a decision about an individual. The same sensor may appear in all four settings, but the evidence for one cannot simply be borrowed by the others.

A trace can describe more than a number

A glucose trace is valuable because it preserves relationships across time. Mean shows where the record tends to sit. Std shows how widely the readings spread around that level. AC_Var captures a feature of how the sequence relates to itself across short lags. Together, these axes distinguish curves that would look equivalent if only the average were reported.

This is a description of the signal, not a verdict about the person. A curve with a higher average is not automatically a diagnosis. A curve with wider variation is not automatically harmful. A distinctive short-lag pattern is not a direct measurement of insulin clearance or beta-cell function. The axes tell us which properties were calculated from the readings; their biological meaning must be supported by separate analyses.

The strongest evidence for the three-axis representation comes from its internal structure and its ability to preserve observed trajectories. In 8,025 participants from the 10 K Project, exploratory factor analysis identified mean-, variance-, and autocorrelation-related factors that explained 82% of the total variance in the CGM-derived measures. Principal component analysis found a similar structure, with its first three components explaining 85% of the total variation. A three-dimensional autoencoder representation reconstructed postprandial profiles from 863 non-diabetic participants with a correlation of 0.92, compared with 0.88 for two dimensions; adding dimensions produced little further improvement in that comparison.

These results answer a bounded question: how many broad statistical directions are needed to represent much of the variation in these datasets? They do not mean that a person’s physiology has only three dimensions. They do not promise that the same percentages hold for every device, population, or recording window. They do not establish that a compact representation is better than a complete clinical risk model. In fact, the authors explicitly note that they did not directly compare the three dynamics-derived measures with a comprehensive multivariable model combining conventional markers in the 10 K Project.

The distinction between representation and clinical use is easy to lose because both may be expressed with impressive correlations. A reconstruction correlation asks how closely the compressed features reproduce an observed waveform. A clinical prediction asks whether a result forecasts a future outcome accurately enough to improve a decision. Those are different targets. A model can preserve the shape of a curve without showing that a person will develop diabetes, that an organ marker will worsen, or that changing the feature will improve health.

This is why the three axes should be treated as coordinates rather than as a hidden score. They help describe whether differences among people arise mainly in average level, spread, temporal organization, or some combination. The combination can be more informative than any one excursion, but “more informative” does not mean “self-interpreting.” Context, measurement quality, and the purpose of the analysis remain part of the meaning.

Association is not diagnosis

The three-axis study connects the CGM features to outcomes beyond the glucose record.