acceptodds
Under review as a conference paper at ICLR 2027

Predictive Uncertainty Reads Insulin Resistance: A Latent-Dynamics Readout from Frozen CGM Forecasters and a Controlled Feedback System

Abstract

Probabilistic time-series forecasters output uncertainty, but it is usually treated as an error bar rather than a signal. We ask whether predictive uncertainty carries information about a latent property of the data-generating process, using a frozen continuous glucose monitoring forecaster whose representation tracks gold-standard insulin resistance when read by a supervised probe. We define relative forecast uncertainty (RFU), the predictive scale normalized by recent within-person glucose variability, and find that it discriminates insulin resistance defined by the steady-state plasma glucose (SSPG) reference test in pooled gold-standard cohorts (AUROC 0.78, two-sided permutation p = 0.0009, n = 47); binary discrimination remains above chance after linear adjustment for glucose level, variability, and cohort (cross-fitted AUROC 0.70 [0.53, 0.85], nominally significant). The sign is clinically counterintuitive: more insulin-resistant subjects receive lower relative uncertainty, and the forecaster has greater realized skill relative to persistence in this group. The readout is sensitive to within-window time shuffling, overlaps substantially with long-range persistence, and is reproduced by four frozen forecasters, including a general-purpose time-series model applied zero-shot (subject rankings agree at 0.90 to 0.96), with protocol-dependent magnitude. RFU aligns with the representation-based readout it was derived to explain, and on a controlled feedback system with subject-level mean and variance matched it ranks a held-out latent gain (+0.90). RFU thus provides a scalar readout of SSPG-defined insulin-resistance status without label-based fitting; its physiological interpretation and the added value of the learned scale beyond normalization remain unresolved.

Primary area:
interpretability and explainable AI
Keywords:
predictive uncertainty, uncertainty quantification, calibration, probing frozen models, time-series foundation models, system identification, continuous glucose monitoring, insulin resistance

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_VBrVOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Relative Forecast Uncertainty (the logged subject-average ratio of a frozen glucose forecaster's 1-hour predicted uncertainty scale to the 24-hour input window's standard deviation) as a label-free readout of gold-standard insulin resistance measured by steady-state plasma glucose. Evaluated across 47 pooled subjects (32 at 48 hours), four frozen backbones, a Bergman minimal-model physiological simulator, and a synthetic linear autoregressive system, RFU achieves an unadjusted binary AUROC of 0.78 and a cross-fitted covariate-adjusted AUROC of 0.70 (which fails at a Holm-adjusted p-value of 0.17), with lower relative uncertainty indicating higher insulin resistance.

Critical strengths

The authors provide an algebraic proof (Appendix A) showing that dividing by the window standard deviation simply cancels the model's internal denormalization to read its normalized uncertainty head, and thorough negative controls. Ironically, these controls are so thorough that they refute the paper's headline claim.

Critical weaknesses
  1. No statistically significant gain over trivial window-variance or classical autocorrelation baselines (Driver-Blindness):
  • Denominator dominance: In Figure 3a, discarding the neural network entirely and replacing its predicted uncertainty with the number 1 (computing only the logged average of the inverse window standard deviation, which weights stable overnight/fasting windows) already achieves an AUROC of 0.71 versus 0.78 for full RFU (p = 0.23) while the raw learned uncertainty scale alone is near chance (0.58). Similarly, the 320-dimensional supervised representation probe works just as well on a randomly initialized, untrained network (AUROC 0.715 untrained vs. 0.724 trained).

  • Driver-blind collapse to classical persistence: Because univariate glucose forecasters lack meal and insulin inputs (Appendix H) and default to short-term momentum extrapolation (driver-blindness), all four frozen models simply act as local autocorrelation filters that output narrower normalized error bars on smooth, persistent curves. Consequently, RFU correlates at -0.85 with classical DFA, fails to significantly beat closed-form DFA (AUROC 0.756), 3-step autoregression (0.712), or lag-1 autocorrelation (0.706) in Table 3, ties DFA identically at 48 hours in Table 2, and loses to 1-step autoregression and simple persistence error across every regime of the synthetic benchmark (Figure 5).

    1. Mechanistic falsification in closed-loop physiology and post-hoc statistical fragility:
    • Refutation by the Bergman physiological simulator: The primary models are pretrained on open-loop Type 1 Diabetes data and applied to closed-loop non-T1D cohorts. While the authors argue that RFU reads a linear feedback controller gain (Appendix B), their own Bergman minimal-model simulations (Section 4.5, Figure 7b) falsify this: whenever realistic compensatory beta-cell insulin secretion is active to stabilize mean glucose, RFU completely fails to track insulin sensitivity (correlations of +0.10 and -0.11), succeeding only when compensation is disabled so that mean glucose heavily confounds insulin sensitivity (correlation of -0.88).

    • Post-hoc formula switching and multiplicity failure: Appendix B proves an exact unexplained-variance identity only for the root-mean-square uncertainty ratio, yet under that median root-mean-square protocol (Table 4), the covariate-adjusted correlation with SSPG collapses into noise across all four models (p = 0.11 to 0.65). Even after switching post-hoc to the uncapped arithmetic mean of ratios, the 24-hour adjusted continuous correlation is non-significant (p = 0.092), the adjusted binary AUROC (0.70) fails Holm multiple-testing correction (p = 0.17), and adding the learned scale to baseline covariates in out-of-fold logistic regression yields no significant improvement (Figure 3b).

Questions for authors
  • In your code (run_06_classical.py, Appendix F), DFA, Sample Entropy, Lag-1 autocorrelation, and 12-step autoregressive error are computed on at most 20 windows at a 4-hour stride, whereas RFU uses all qualifying windows at a 2-hour stride. What are the Table 2 and Table 3 numbers when all classical baselines use the exact same windows and aggregation as RFU?

  • Since the constant-scale control (logged mean inverse window standard deviation) achieves AUROC 0.71 on its own (Figure 3a) versus 0.64 for subject-level coefficient of variation (Table 3), does RFU retain any statistically significant association with SSPG when adjusted directly for this constant-scale denominator?

  • Why does Table 4 use Freedman-Lane permutation for the main protocol rows but switch to naive within-cohort outcome permutation for the variant rows?

Reviewer_SAMpOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper asks whether a probabilistic forecaster's predictive uncertainty carries information about a latent property of the process generating the data. It defines relative forecast uncertainty (RFU) as a frozen CGM forecaster's horizon-averaged predictive scale divided by the context window's glucose SD, logged at the subject level. For window-standardizing forecasters this equals the model's normalized-space scale. In pooled gold-standard cohorts (n=47, 16 insulin-resistant), lower RFU marks SSPG-defined insulin resistance

Critical strengths

The paper validates against SSPG rather than a proxy. It uses within-cohort permutation tests, subject-level bootstrap intervals, Freedman-Lane partial correlations, and Holm correction. The release recomputes 32 key reported values. Analyses that need withheld data (calibration, realized error, AI-READI, OGTT, disposition index) can't be checked. This matters because it makes the crude finding (AUROC 0.78, n=47) credible, and it makes the paper's remaining problems concrete and fixable rather than speculative. It doesn't carry acceptance, because the claims beyond that association rest on results that are nominal, inconclusive, or at odds with an unreported large-sample analysis

Four frozen forecasters, including TimesFM-3, applied zero-shot learning, achieving AUROC values of 0.7-0.78. Their subject rankings agree at 0.90-0.96, and binary discrimination holds in all 16 forecaster-by-protocol cells (4 x 4). So it shows that the association doesn't depend on a single model, which is the paper's most solid result. Its scope is limited, though: all four models are evaluated on the same 47 subjects, so this is robustness across models, not replication across people. It doesn't change my recommendation, because the same genericity is what we expect if RFU largely tracks glucose variability.

Critical weaknesses

Figures are definitely not a strength of this paper. Figure 1 is poorly drafted, unengaging, and almost uninformative.

The central adjusted claim is not established, and the only well-powered test of an adjusted association is essentially null and unreported. In the gold-standard cohorts, the evidence is weak:

  • the adjusted result is nominal (cross-fitted AUROC 0.70, Holm-adjusted p 0.17);
  • the adjusted gradient is inconclusive (partial -0.26).

The authors' own AI-READI analysis (n=2,181; logs/v3_07_compression.txt, code/run_07_compression.py lines 76–77) finds that RFU's association with log HOMA-IR falls from −0.27 to −0.03 (roughly [−0.07, +0.01]) after adjusting for mean glucose and CV. It was computed in the same run as the probe alignment reported in the paper, but it is not included in the paper.

HOMA-IR is a fasting, hepatically weighted proxy, and AI-READI includes treated type 2 diabetes, so this result doesn't refute an SSPG-specific signal. But the paper treats log HOMA-IR in AI-READI as an adequate label (it trains the probe on it), so the result should be reported and discussed.

Matching marginals leaves RFU unchanged. The matched and natural regimes give identical RFU-gain correlations, seed by seed (mean +0.899), as per-window normalization implies. So the +0.90 is the raw correlation in a first-order linear system. The low natural-regime adjusted value reflects over-adjustment: series mean and SD explain about 90% of the rank variance in the gain

48-hour context: the authors rightly say this isn't a context effect, but the larger estimates are still used to support it. They reflect which subjects have long records: the same 32 subjects give the same effect at 24h

The two simulator settings designed to separate insulin sensitivity from glucose level can't see it, by construction. Setting (b) produces bit-identical glucose traces for any S_I, and setting (c) pins S_I x secretion except where its bisection hits a bound. In this model class, a glucose-only statistic reads sensitivity x secretion. The question falls to human analysis, which is inconclusive. The paper itself rates specificity "not established"

So these analyses are what move the paper from an association to its title claims. Without them, the evidence is the n=47 association, a nominal adjusted result, and its reproduction across forecasters on the same subjects. The paper's contribution beyond the claim that "glucose variability tracks insulin resistance" depends on the adjusted association. Without it, RFU doesn't significantly beat window SD alone

Questions for authors

Your logs/v3_07_compression.txt reports RFU against log HOMA-IR in AI-READI (n=2,181). The Spearman correlation of −0.274 falls to −0.030 after adjusting for mean glucose and CV. Why is this not reported, and how does it bear on the adjusted SSPG claim, given that the probe is trained on the same label? Please report it. If medication data are available, please repeat it, excluding participants on insulin or other glucose-lowering therapy

The constant-scale control already shows that RFU isn't significantly better than window SD on identical windows. Could you compute DFA, lag-1 autocorrelation and sample entropy on the RFU windows (all windows, 2-h stride) rather than on the first 20 windows at a 4h stride?

Per-window normalization makes RFU invariant to z-scoring a whole series. Your seed-decomposition file shows identical matched and natural RFU–gain correlations, seed by seed. Do you agree that:

  • the matched regime doesn't constrain RFU; and
  • the low natural-regime adjusted value reflects over-adjustment, since series mean and SD explain about 90% of the rank variance in the gain? If so, what does 4.6 show beyond the raw correlation, given that an AR(1) fit does better (+0.95)?

In simulator settings (b) and (c), S_I enters glucose only through its product with secretion. With your code, setting (b) gives bit-identical traces across S_I, and setting (c) pins the product except where the bisection hits its bound. Do you agree that these settings cannot inform on S_I, and that in this model class, a glucose-only statistic reads sensitivity x secretion? What evidence in your human cohorts separates the two, beyond the n=24 disposition-index analysis?

Do the RFU windows include OGTT or standardized-meal periods? The window code excludes no days, and the OGTT log lists 10 tests "with matching CGM." Does the result hold with those periods excluded? Also, why does 4.5 report the n=20 OGTT subset?

Which group had 0.81 coverage and which had 0.86, and what does "centred probability-integral transforms" mean? If insulin-sensitive subjects are over-covered, how much of the RFU gap reflects differential calibration rather than predictability?

Reviewer_183SOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper investigates whether predictive uncertainty from frozen probabilistic continuous glucose monitoring (CGM) forecasters can reveal latent physiological properties, particularly insulin resistance. The authors introduce relative forecast uncertainty (RFU), defined as the predictive scale normalized by recent within-person glucose variability. The motivation is that insulin-resistance information may be encoded in glucose dynamics rather than absolute glucose levels. Experiments conducted on a cohort of 47 subjects show that RFU can discriminate insulin resistance (AUROC 0.78), with lower RFU associated with greater insulin resistance. The authors further evaluate RFU across four frozen forecasters to examine the robustness of the observed association with insulin resistance across different forecasting models.

Critical strengths
  1. The paper explores whether forecasting uncertainty can capture latent physiological properties without supervised label fitting.

  2. The authors evaluate RFU across four frozen forecasters, compare it with classical time-series measures, and include controlled simulations. These experiments provide useful evidence that the observed association is not specific to one forecasting architecture.

Critical weaknesses
  • Limited methodological novelty: the main contribution is an empirical analysis of insulin resistance using a relatively simple normalization of predictive uncertainty, rather than a new forecasting or representation-learning methodology. The paper would benefit from a clearer explanation of its methodological contribution to the broader machine learning community.

  • The constant-scale normalization baseline already achieves AUROC 0.71 compared with 0.78 for RFU, and the improvement is not statistically significant. Combined with the small sample size (n=47), this raises concerns about whether the learned uncertainty provides substantial information beyond conventional variability-based statistics.

Questions for authors
  1. The Introduction states that "Probabilistic forecasters emit an estimate of their own uncertainty, almost always treated as an error bar rather than as a signal in its own right." This statement appears overly broad, given existing research on uncertainty-aware forecasting and uncertainty-based representation learning. Could the authors clarify how their proposed use of predictive uncertainty differs from existing approaches that explicitly model or exploit uncertainty? A more comprehensive discussion of related work would help establish the novelty of the proposed method.

  2. Given that the controlled synthetic system is relatively simple and a classical autoregressive baseline performs better than RFU, could the authors clarify what additional insights RFU provides and whether its advantages would extend to more complex dynamical systems?

  3. The paper would benefit from substantial improvements in clarity, organization, and presentation:

  • Abstract: The abstract moves quickly to experimental results without sufficiently explaining the methodological motivation and contribution.
  • Introduction: Several sentences are unusually short and feel disconnected, making the argument difficult to follow in places.
  • Figures: The paper main figure (Figure 1) contains substantial empty space and could be reorganized for better readability.
  • Related work: The discussion sometimes reads as a list of existing models rather than a critical comparison explaining how the proposed approach differs from prior methodologies.
  • Methodology: Section 3 is relatively brief given the number of datasets, models, and statistical procedures involved. The subsection "Cohorts, labels, and representation probe" is especially dense and would benefit from a more structured explanation of cohort construction, label definitions, and evaluation protocols.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.