What the Curve Contains: Level, Variation, and Memory
In CGM-derived glucose data from adults without diagnosed diabetes, the challenge is not a shortage of measurements. It is knowing which measurements describe genuinely different features of the curve and which merely repeat the same information in new forms. The key question is how to reduce a dense time series to a small set of interpretable dimensions without losing the pattern that makes continuous monitoring useful. The answer developed in recent research is a three-axis model: average level, extent of variation, and short-lag temporal structure.
Those axes are useful because they describe different properties of glucose behavior. They are not three labels for “good” or “bad” glucose, and they are not a complete readout of a person’s physiology. They are a way to organize the signal so that a reader can understand what the data show before asking what those features may be associated with.
From a list of readings to a shape
Imagine writing down a glucose value every few minutes. The list is long, but length alone does not tell us what matters. One way to summarize it is to calculate the average. Another is to ask how far the readings spread around that average. A third is to ask whether nearby readings tend to resemble one another and how that resemblance changes as the time gap grows.
These questions describe different aspects of the same record. Two people might have a similar average glucose but different degrees of fluctuation. Two others might have comparable averages and spreads but different temporal patterns: one curve may move in smooth, persistent stretches, while another may reverse direction more quickly. A summary that keeps only the average cannot distinguish them. A summary that keeps average and spread still may miss the arrangement of the changes over time.
This is what a CGM trajectory adds to a snapshot. It does not make every wiggle biologically important. It gives researchers enough repeated observations to ask whether the wiggles have a stable structure across a recording and whether different people vary along more than one dimension.
The research terminology is precise but manageable:
- Mean is the average glucose level across the relevant readings. It describes where the curve is centered.
- Std, or standard deviation, describes the extent to which readings vary around that center. It is a measure of spread, not a count of how many times a person crossed an arbitrary threshold.
- AC_Var is an autocorrelation-based feature. Autocorrelation asks how much a time series resembles itself after a delay; AC_Var captures how that relationship varies across short lags. It therefore describes temporal structure, not just level or spread.
The word “memory” in the title is a shorthand for that last idea, not a claim that glucose has memory in a conscious or literal sense. If nearby readings are related in a characteristic way, the sequence contains information about how changes unfold. AC_Var is an operational statistic calculated from that sequence. It should not be silently renamed “stability,” “insulin resistance,” or any other clinical diagnosis.
Why three axes can replace a crowded dashboard
CGM software can generate many indices. Some measure similar things with slightly different formulas; others combine existing measures. Presenting all of them can create the impression of precision while making it harder to see the underlying pattern. The study of 8,025 people in the 10 K Project used dimensional-reduction methods to ask how many independent directions were actually needed to represent the variation among the measures.
One of those methods, exploratory factor analysis, groups observed variables according to shared underlying patterns. In this analysis it identified three factors. A mean-related group, a variance-related group, and an autocorrelation-related group together explained 82% of the total variance in the CGM-derived measures. Principal component analysis, a related method that finds major directions of variation, produced a similar result: its first three components explained 85% of the total variation.
The percentages do not mean that 82% or 85% of a person’s physiology has been measured. They mean that, within these data and under the study’s preprocessing and selection decisions, most of the differences among the recorded CGM metrics could be organized along three statistical directions. This distinction matters. A compact description can preserve much of the information in a dataset without becoming a complete theory of the body.
The authors then chose representative measures for those directions: Mean for the level-related axis, Std for the variation-related axis, and AC_Var for the autocorrelation-related axis. The choice was not a declaration that every other index is useless. It was a way to limit redundancy and keep the analysis interpretable. Mean, Std, and AC_Var were also selected because the study treated them as relatively distinct rather than as three highly overlapping versions of the same quantity.
The complementarity is easiest to see with simple curve shapes. Two people can have the same Mean while one has a narrow, quiet spread and the other has wider excursions; Std separates those records. Two people can have similar Mean and Std while one moves through a smooth rise and fall and the other changes direction in a more irregular temporal pattern; AC_Var is intended to preserve information about that ordering. None of these differences is automatically a disease signal. They are distinctions in the data that let a later analysis ask a more specific question.
That is also why the three measures should be read together rather than ranked from “best” to “worst.” Mean is not the normal metric and AC_Var the advanced metric. They are different coordinates. A person’s average level may matter even when their short-lag pattern is unremarkable; a temporal feature may add information even when the average and spread look similar to someone else’s. The point of the representation is to prevent one coordinate from standing in for the whole curve.
The agreement between factor analysis and principal component analysis strengthens the description of the dataset because two related ways of finding structure pointed toward the same three broad directions. It does not turn the result into a biological law. Statistical methods can reveal recurring patterns in recorded variables, but the meaning of those patterns still depends on the people studied, the measurements available, and the outcomes used to test them. The authors therefore pre-specified one representative measure for each direction rather than testing an unlimited collection of near-duplicates. That choice improves interpretability and reduces the temptation to treat the most favorable metric as the answer.
A helpful way to picture the result is this:
The diagram is not a recipe for turning a home sensor into a clinical score. It shows the direction of the reasoning: many measurements can be compressed into three interpretable axes, and those axes can be tested by seeing whether they preserve the shape of a glucose response.
Compression is tested by reconstruction
A statistical grouping is useful only if it preserves something important. To test that, the researchers used an autoencoder model on postprandial glucose profiles from 863 participants in the PREDICT1 Study. An autoencoder compresses an input into a smaller internal representation and then attempts to reconstruct the original pattern from that compressed representation. Here, the relevant question was whether a small number of dimensions could reproduce the shape of a post-meal glucose trajectory.
A two-dimensional representation achieved a correlation of 0.88 between measured and reconstructed profiles. Expanding the representation to three dimensions raised the correlation to 0.92, with a 95% confidence interval of 0.92 to 0.93. Adding more dimensions produced only minimal improvement, with the reported change in correlation below 0.01. That pattern supports three dimensions as a useful compromise in the studied setting: two did not preserve as much, while additional dimensions added little to reconstruction performance.
The researchers next tested whether the three simple, pre-specified measures could stand in for the latent dimensions. Mean, Std, and AC_Var independently predicted the three latent dimensions with correlations of 0.93, 0.95, and 0.83. The values are not equally strong, and they should not be read as a single “accuracy” score. They show that the selected measures tracked the compressed representation well enough to serve as interpretable proxies for its main axes.
This is a mechanism of representation, not a mechanism of disease prediction. Reconstructing a waveform means that the compressed features retain information about the observed waveform. Predicting future diabetes, forecasting a vascular event, identifying a cause, or selecting treatment would require a different outcome, a different validation design, and evidence that the prediction improves a real decision. A model can reproduce the input pattern without answering any of those questions.
Why the autocorrelation axis is not just another variability score
Mean and Std are intuitive because they describe location and spread. AC_Var requires more care. Standard deviation can tell us how widely readings are distributed, but it does not tell us how those readings are arranged in time. The same set of values can be ordered in different ways and produce different trajectories. Autocorrelation adds information about that ordering by comparing the series with delayed versions of itself.
The paper chose AC_Var rather than AC_Mean as its representative autocorrelation feature. AC_Var is based on how the autocorrelation function changes across short lags, and the authors describe it as more robust to noise in prior work. In practical terms, it asks not only whether adjacent readings are related, but how the strength of that relationship varies over the short delays used in the calculation.
That distinction gives AC_Var a different role from Std. A curve can have moderate spread but a distinctive temporal organization, or broad fluctuations without the same short-lag pattern. The three axes can therefore separate questions that a single “variability” label would blur together.
The physiological interpretation must remain one step removed from the calculation. The authors discuss AC_Var as potentially reflecting combinations of insulin sensitivity, beta-cell responsiveness, and insulin clearance, based on prior clamp- and oral-glucose-tolerance studies. Those are hypotheses about what the pattern may represent biologically. AC_Var itself is still a statistic extracted from a glucose time series. A high value in an individual does not prove impaired insulin clearance, just as a high Mean does not by itself prove insulin resistance.
This separation between measurement and interpretation is not pedantic. It keeps the data honest. First state what the feature calculates. Then state what it was associated with or what researchers propose it may reflect. Do not skip directly from a named metric to a clinical conclusion.
What the three-axis model leaves out
Every compression has a boundary. The study’s three axes summarize most observed variation among the selected CGM measures in the analyzed cohorts. They do not guarantee the same percentages in every population, device, age group, disease state, or recording duration. A representation that works well for one dataset can lose information when the conditions that generated the data change.
The model also does not remove the need for adequate data. The autocorrelation-based measure depends on densely sampled, contiguous time-series data, and the study was built around multi-day CGM windows; the 10 K Project analysis used recordings of at least seven days. The authors note that longer recordings may improve estimation stability. A short, interrupted record may still be visually interesting, but it is not automatically equivalent to the research data used to define the feature.
There is a second boundary in the difference between controlled and everyday settings. A standardized meal creates a clearer starting point and a more predictable challenge. Ordinary life includes meals close together, imperfectly recorded timing, and other factors that are not represented in the glucose series. A three-dimensional model reconstructed the controlled postprandial profiles well, but its correlation was lower when applied to real-life meal profiles. That does not invalidate the model. It tells us that the meaning of a compact representation depends partly on how orderly the underlying context is.
Finally, the three axes are not a replacement for all conventional clinical information. The study reports that they capture important variation beyond fasting and two-hour glucose values, but it did not directly compare them with a comprehensive multivariable risk model combining conventional markers in the 10 K Project. More dimensions can improve description without proving that a decision made with those dimensions is better.
The durable mental model is therefore modest but powerful. Mean tells us where the curve tends to sit. Std tells us how widely it moves around that level. AC_Var tells us about a feature of its short-lag temporal organization. Together, they can preserve much of the structure researchers see in CGM data from people without diagnosed diabetes. They make the curve easier to study, not the person easier to diagnose.
The natural question now is what these axes relate to beyond reconstruction: whether their differences appear alongside meaningful metabolic, vascular, or liver-related markers, and how strong that evidence really is.