acceptodds
Under review as a conference paper at ICLR 2027

MealGluco: Physiologically Structured Postprandial Glucose Trajectory Forecasting with Online Personalization

Abstract

Predicting postprandial glucose trajectories is important for proactive metabolic health management but remains underexplored. This task is challenging because a model must capture long-horizon glucose dynamics, integrate the effects of the current meal with the evolving pre-meal glucose state, and account for substantial inter-individual variability. To address these challenges, we propose MealGluco, a physiologically structured model that predicts the complete, high-resolution postprandial glucose trajectory from glucose history and meal composition. MealGluco represents the meal-induced response using a peak-normalized Gamma function, captures background glucose dynamics through a separate state pathway, and recursively updates two participant-specific coordinates to enable lightweight online personalization without retraining the shared network. Across three datasets and five evaluation metrics, MealGluco achieves a mean rank of 1.33 among seven evaluated models. Its online personalization reduces trajectory prediction error relative to the non-personalized model while matching the performance of participant-specific fine-tuning at orders-of-magnitude lower computational and storage costs. MealGluco also consistently outperforms the strongest baseline in cross-dataset evaluation, demonstrating improved generalization across populations and devices.

Primary area:
learning on time series and dynamical systems
Keywords:
postprandial glucose forecasting, continuous glucose monitoring, online personalization, physiological signal modeling, time series

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_q4wTOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes MealGluco for predicting a 2-hour post-meal continuous glucose monitoring (CGM) trajectory from 2 hours of pre-meal CGM, and meal nutrients. The population prediction adds the mealtime glucose anchor to a bounded state-response pathway (combining a frozen CGM-JEPA encoder, four short-horizon summary statistics, and time queries) and a meal-response pathway parameterized as a three-variable unimodal Gamma curve. After each completed meal, participant-specific coordinates scaling meal amplitude and post-peak clearance are updated online.

Critical strengths

Well-motivated formulation and comprehensive evaluation: Predicting complete postprandial trajectories across five complementary shape and timing metrics under participant-disjoint splits is a practical formulation, and the authors include participant-clustered paired bootstrap confidence intervals (Appendix C.6), a comparison against per-participant fine-tuning, and a wrong-subject history control (Table 15).

Lightweight online personalization mechanism by updating only two scalar response coordinates.

Critical weaknesses
  1. Unmatched online adaptation in baseline comparisons and overstated universal ablation claims:
  • Architectural gains are confounded with online adaptation: In both the main tables and cross-dataset transfer (Table 3), MealGluco updates its participant-specific coordinates online using completed target-participant meals, whereas all competing baselines are evaluated with fixed population weights.

  • In Table 20, the 95% confidence intervals show that MealGluco does not significantly beat the strongest baseline on any of the five metrics on BIG IDEAs (and is significantly worse on Recovery-MAE by +1.43 mg/dL) and only significantly improves two of five metrics on T1D-UOM. Moreover, contrary to the claim in Section 4 that removing any proposed module degrades performance across all datasets, Table 28 shows that 0 of 10 component ablations on T1D-UOM and only 3 of 10 on BIG IDEAs have confidence intervals excluding zero for Trajectory-MAE.

  1. Structural priors vs. physiological identification: Training supervises only the sum of the state and meal branches, and T1D-UOM lacks insulin records (Appendix B.1), meaning the conditional forecast cannot disentangle nutritional responses from unobserved prandial insulin dosing. Similarly, the nutrient sensitivities in Figure 5(c) partly reflect imposed structural priors (explicit amplitude-attenuation and peak-delay modifiers, plus a hardcoded gate forcing the meal response to zero whenever carbohydrate is zero, Appendix A.4.2) rather than independent physiological validation. Furthermore, because the meal kernel is nonnegative and the state residual is bounded, the population model is mathematically incapable of predicting glucose drops more than 20 mg/dL below the mealtime anchor (or 40 mg/dL after personalization).

  2. Furthermore, the evaluation deletes all meals with incomplete macronutrients upstream (Appendix B.1) and includes no robustness analysis for noisy or partially missing self-reported food logs. This is a critical gap for free-living deployment, where mislogging errors will also propagate into the recursive personalization state.

Questions for authors
  • Can you reconcile the CGMacros reference numbers between Table 5 (22.68 / 28.16) and Table 12 (22.63 / 26.96), explain why the ablation deltas in Table 28 differ from Table 12, and confirm whether the 20 mg/dL state bound in Appendix B.4.2 was selected strictly on inner validation or on outer-fold test scores?

  • How do the strongest baselines (Stacking Framework and UniCA-Gluformer) perform in within-dataset and cross-dataset evaluations when equipped with a comparable lightweight online calibration fitted on the exact same completed meal histories?

  • Given that an unstructured meal head outperforms the Gamma kernel on Peak-MAE and Recovery-MAE in several datasets and removing the state pathway improves peak timing and recovery on CGMacros, how do you justify the claim that every proposed component provides a consistent benefit across metrics and cohorts?

  • How frequently do real post-meal trajectories in your three datasets drop more than 20 mg/dL (population) or 40 mg/dL (personalized) below the mealtime anchor and how should the meal branch on T1D-UOM be interpreted given the absence of prandial insulin records?

3 comments
AuthorsOpenReview ↗

Response to Reviewer q4wT

Thank you for carefully reviewing our work and for your constructive feedback. We have made corresponding revisions to the paper based on your feedback, and all revisions are highlighted in red in the updated PDF. We respond to your questions point by point below.

Q1.1: Can you reconcile the CGMacros reference numbers between Table 5 (22.68 / 28.16) and Table 12 (22.63 / 26.96), and explain why the ablation deltas reported in Table 28 differ from those implied by Table 12?

A1.1: We apologize for this. During the aggregation of the CGMacros data, we mistakenly included results from different experimental batches in the confidence-interval tables, which led to the discrepancy. All CGMacros results have now been updated using the latest data. The table numbering changed in the revised version, and the complete results are now reported in Table 23 and Tables 39–43 (pages 34 and 40).

Q1.2: Was the 20 mg/dL state bound in Appendix B.4.2 selected strictly on inner-validation results, or did outer-fold test scores contribute to its selection?

A1.2: To clarify, the value Rmax⁡=20R_{\max}=20 mg/dL on CGMacros was selected solely based on the inner-validation Trajectory-MAE results; the outer-fold results were not used in the selection. We clarified this issue in Appendix B.4.2 and added the inner-validation Trajectory-MAE results to Table 6 (page 28). For your convenience, we show the inner-validation Trajectory-MAE results below. The inner-validation Trajectory-MAE varies by only 0.345 mg/dL across the six candidates, indicating that model performance is not sensitive to Rmax⁡R_{\max}.

MetricRmax⁡=10R_{\max}=10Rmax⁡=15R_{\max}=15Rmax⁡=20R_{\max}=20Rmax⁡=25R_{\max}=25Rmax⁡=30R_{\max}=30Rmax⁡=40R_{\max}=40
Inner-val Trajectory-MAE21.84621.69621.50121.52321.51521.634

Q2: How do the strongest baselines (Stacking Framework and UniCA-Gluformer) perform in within-dataset and cross-dataset evaluations when equipped with a comparable lightweight online calibration fitted on the exact same completed meal histories?

A2: As requested, we used the same lightweight online calibration as MealGluco for Stacking Framework and UniCA-Gluformer, and recalculated the within-dataset and cross-dataset metrics using the same historical meal information. For your convenience, we show the relevant results below. The complete results for all five metrics are reported in Table 1, Table 3, Appendix C.1, and Appendix C.5 (pages 7, 9, 29, and 35).

Within-dataset

DatasetStacking FrameworkStacking Framework (Pers.)ΔUniCA-GluformerUniCA-Gluformer (Pers.)Δ
CGMacros23.4822.88-0.6024.2123.77-0.44
BIG IDEAs16.4916.30-0.1918.8118.21-0.60
T1D-UOM32.3030.26-2.0433.3432.13-1.21

Cross-dataset

TransferStacking FrameworkStacking Framework (Pers.)ΔUniCA-GluformerUniCA-Gluformer (Pers.)Δ
CGMacros → BIG IDEAs19.8418.08-1.7618.9917.68-1.31
CGMacros → T1D-UOM31.1730.49-0.6834.8932.17-2.72
BIG IDEAs → CGMacros30.1727.55-2.6225.3626.31+0.95
BIG IDEAs → T1D-UOM38.7532.47-6.2835.2532.61-2.64
T1D-UOM → CGMacros26.5226.01-0.5127.5726.99-0.58
T1D-UOM → BIG IDEAs21.3718.51-2.8621.7619.71-2.05

Q3: Given that an unstructured meal head outperforms the Gamma kernel on Peak-MAE and Recovery-MAE in several datasets and removing the state pathway improves peak timing and recovery on CGMacros, how do you justify the claim that every proposed component provides a consistent benefit across metrics and cohorts?

A3: We agree that some statements in the original manuscript could lead readers to understand that all modules consistently improve every metric across all three datasets. We have revised the main text to read: “The component ablation results confirm that each proposed module contributes to the model’s overall forecasting performance. Specifically, removing any module consistently increases Trajectory-MAE across all three datasets.”, and deleted “Across all three datasets, removing any proposed module leads to a clear performance drop.”, which was too absolute (page 9).

Q4.1: How frequently do real post-meal trajectories in the three datasets drop more than 2020 mg/dL below the mealtime anchor for the population model, or more than 40 mg/dL below it for the personalized model?

A4.1: We counted the proportion of meals in which the real post-meal trajectory falls at least 20 or 40 mg/dL below the mealtime anchor at any point. Larger post-meal declines are more common in T1D-UOM, consistent with the general pattern of insulin action and larger post-meal declines in people with type 1 diabetes.

DatasetMealsDrop >20>20 mg/dLDrop >40>40 mg/dL
CGMacros1,650278 (16.8%)94 (5.7%)
BIG IDEAs53491 (17.0%)30 (5.6%)
T1D-UOM1,487597 (40.1%)338 (22.7%)

We further conducted an Rmax⁡R_{\max} sensitivity analysis on T1D-UOM. The experiment followed exactly the same configuration and protocol as the main results, changing only Rmax⁡R_{\max} from 20 mg/dL to 40 mg/dL before retraining and evaluation. The results have been added to Table 8 (page 28) and are also shown below for your convenience. A negative Relative change indicates lower error for Rmax⁡=40R_{\max}=40. The results show that increasing Rmax⁡R_{\max} does not lead to consistent improvement across metrics. Therefore, a larger Rmax⁡R_{\max} is not an effective way to improve overall predictive performance.

Rmax⁡R_{\max}Trajectory-MAEiAUC-MAEPeak-MAENear-Peak DistanceRecovery-MAE
2030.08 ± 0.551986.8 ± 45.331.87 ± 0.5324.83 ± 1.5424.06 ± 0.55
4029.94 ± 0.491968.0 ± 49.231.71 ± 0.5926.63 ± 1.0324.58 ± 0.38
Relative change−0.46%−0.95%−0.50%+7.26%+2.17%

Q4.2: How should the meal branch on T1D-UOM be interpreted given the absence of prandial insulin records?

A4.2: We would like to clarify that the absence of prandial insulin records does not mean that the meal branch is invalid. People without diabetes also secrete insulin after meals, and this process is not an observed input either. Postprandial glucose prediction can still be performed based on pre-meal CGM, meal nutrient composition, and individual history. Therefore, our meal branch models the nutrient-conditioned component of the observed postprandial response.

Under our protocol, MealGluco and all baselines use the same inputs without insulin. None explicitly models the effect of injected insulin on postprandial glucose, and our experiments show that MealGluco models the effect of diet on postprandial glucose better than the baselines. The larger post-meal declines in T1D-UOM may partly reflect unobserved exogenous insulin and other physiological factors. Although these factors are not provided as explicit input variables, their participant-specific effects are reflected in the historical CGM trajectory and personalization parameters. Adding insulin may improve prediction accuracy for some cases, but it would change the research question and input protocol. This paper focuses on modeling postprandial glucose trajectories based on meal composition and does not model the effects of exogenous insulin. Therefore, the meal branch on T1D-UOM should be interpreted as a conditional prediction component based on meal inputs and individual history, not as an estimate of insulin effects. At the same time, this work focuses on the effects of diet on postprandial glucose.

AuthorsOpenReview ↗

Note on updated table and page numbers

We have completed a second round of revisions to the manuscript. All newly added and revised content is marked in blue. Because an additional Appendix table has been included, the table numbers and page numbers have changed in the latest PDF. We have updated the relevant table and page references in our response comments above. Please use the numbering in the latest manuscript when checking the revisions. We hope that the revised manuscript now addresses your concerns.

Reviewer_m6LJOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MealGluco, a framework for forecasting complete postprandial glucose trajectories from pre-meal CGM history and meal composition. The method decomposes the prediction into a structured meal-induced response and a history-dependent state response, and further introduces two participant-specific latent coordinates for online personalization. These coordinates are recursively updated using previously completed meals while keeping the shared model fixed.

The problem is meaningful, the proposed decomposition is intuitive, and the paper includes extensive experiments on three datasets. However, I have several concerns regarding the evaluation of online personalization, the identifiability of the participant-specific parameters, and whether some of the main claims are fully supported by the reported results.

Critical strengths
  • The paper addresses a practically relevant problem: predicting the complete postprandial trajectory rather than only short-horizon glucose values or summary statistics.
  • The decomposition into meal-induced and pre-meal-state responses provides a useful structural inductive bias.
  • The experimental protocol is generally careful, with participant-disjoint evaluation, multiple datasets and metrics, and a strictly causal online update.
  • The proposed two-dimensional personalization mechanism is computationally lightweight compared with participant-specific fine-tuning.
Critical weaknesses
  1. The comparison with frozen baselines is not fully fair. MealGluco uses completed trajectories from the test participant to update the personalization parameters, while the main baselines remain fixed at test time. Since a noticeable part of the performance gain comes from personalization, it is difficult to separate the benefit of the model itself from the benefit of additional participant-specific information.
  2. Personalization and evaluation are mixed in the current protocol. The model is continuously updated using earlier meals from the same test participant, and later meals are then used for evaluation. This is causal, but I would prefer to see an additional setting where the parameters are estimated from an initial set of meals, then frozen and evaluated on a separate set of later meals.
  3. The two personalization parameters are not sufficiently validated. The paper does not show whether u_rho and u_r are separately identifiable, nor whether their estimates actually stabilize as more meals are observed. Figure 6(c) shows improving prediction error, but this is not the same as showing parameter convergence.
  4. Some claims are stronger than the reported evidence. The structured Gamma response is significantly better than the unstructured head only on CGMacros, but not on the other two datasets. Similarly, personalization is not consistently beneficial across all datasets and metrics, and participant-specific fine-tuning is significantly better on T1D-UOM. These results suggest that some statements about the structured response, personalization, and overall generalization should be more carefully qualified.
Questions for authors
  1. Can the strongest baselines be evaluated with a comparable amount of causal participant-specific adaptation, so that the contribution of the architecture can be separated from the benefit of additional test-time supervision?
  2. Can the authors report an adaptation-then-evaluation experiment, where the participant coordinates are estimated from the first K meals, frozen, and evaluated on subsequent unseen meals?
  3. Do the authors have empirical evidence that u_rho and u_r are separately identifiable and stable? In particular, how do their posterior uncertainties and estimates evolve as more meals are assimilated?
  4. Could the authors moderate the claims regarding the structured Gamma response, online personalization, and cross-dataset generalization to better reflect the dataset-specific statistical results?
  5. Have the authors considered compare your prediction model with FDA-approved models, such as UVA/Padova Type 1 Diabetes Simulator?
7 comments
AuthorsOpenReview ↗

Response to Reviewer m6LJ

Thank you for carefully reviewing our work and for your constructive feedback. We have made corresponding revisions to the paper based on your feedback, and all revisions are highlighted in red in the updated PDF. We respond to your questions point by point below.

Q1: Can the strongest baselines be evaluated with a comparable amount of causal participant-specific adaptation, so that the contribution of the architecture can be separated from the benefit of additional test-time supervision?

A1: To separate the contribution of the architecture from the benefit of additional test-time supervision, following your suggestion, we equipped the strongest baselines with the same personalization module as MealGluco, denoted as UniCA-Gluformer (Pers.) and Stacking Framework (Pers.), respectively. For your convenience, we show the results on CGMacros in the table below, and have added the complete results to Tables 1, 9, and 10 (pages 7 and 30). Personalization improves the predictive performance of both Stacking Framework and UniCA-Gluformer, showing that these models benefit from the additional test-time supervision. Nevertheless, they still do not surpass MealGluco's overall performance. Therefore, the remaining difference under the same personalization conditions is more appropriately attributed to the architecture.

ModelTrajectory-MAEiAUC-MAEPeak-MAENear-Peak DistanceRecovery-MAE
MealGluco21.16 ± 0.171701.7 ± 21.325.00 ± 0.3217.54 ± 0.3319.18 ± 0.32
UniCA-Gluformer (Pers.)23.77 ± 0.431979.8 ± 34.630.80 ± 0.5719.27 ± 0.1020.79 ± 0.10
Stacking Framework (Pers.)22.88 ± 0.021911.2 ± 3.429.27 ± 0.1119.85 ± 0.2219.91 ± 0.07

Q2: Can the authors report an adaptation-then-evaluation experiment, where the participant coordinates are estimated from the first K meals, frozen, and evaluated on subsequent unseen meals?

Following your suggestion, we have added the detailed experimental setup and results for the adaptation-then-evaluation experiment to Appendix C.2.1 (page 32). For your convenience, we show the Trajectory-MAE results in the table below. For each participant, we fixed the first 16 meals as the history. To ensure that all values of KK use exactly the same test set, we fixed 9 meals per participant as the test set. These test meals were sampled at equal intervals in chronological order from eligible meals after the 16th and remained unchanged for K=0,…,16K=0,\ldots,16. The results are consistent with the conclusion in Figure 6(c) that performance improves as the number of completed meals increases.

DatasetK=0K=1K=2K=4K=8K=16
CGMacros23.78 ± 0.3623.94 ± 0.2022.32 ± 0.3821.60 ± 0.3921.47 ± 0.2821.22 ± 0.42
BIG IDEAs16.31 ± 0.2516.32 ± 0.1416.25 ± 0.1216.29 ± 0.2016.02 ± 0.1415.98 ± 0.20
T1D-UOM31.75 ± 0.5531.35 ± 0.6730.69 ± 0.5030.14 ± 0.4530.14 ± 0.6030.43 ± 0.52

Q3.1: Do the authors have empirical evidence that u_rho and u_r are separately identifiable and stable?

A3.1: uρu_\rho and uru_r are not interchangeable parameters in the model: uρu_\rho adjusts the meal-response amplitude through a bounded transformation and primarily controls the peak scale; uru_r affects only the post-peak clearance factor, which is 1 before the peak and gradually decays after the peak. Therefore, the region near the peak primarily provides information about uρu_\rho, while the post-peak recovery region mainly constrains uru_r. The online Gaussian information-form update jointly estimates the two coordinates, with their overlap in the post-peak region explicitly modeled by the off-diagonal terms of the information matrix. Provided that the observations cover both the peak and recovery regions, the two coordinates can be estimated separately under the given parameterization.

Q4.1: Could the authors moderate the claims regarding the structured Gamma response to better reflect the dataset-specific statistical results?

A4.1: Regarding the structured Gamma response, we agree that some statements in the original manuscript could lead readers to understand that all modules consistently improve every metric across all three datasets. We have revised the main text in the PDF to read: “The component ablations confirm that every proposed module contributes to the model's overall forecasting performance, particularly in Trajectory-MAE.” We also deleted “Across all three datasets, removing any proposed module leads to a clear performance drop.”, which was too broad (page 9).

Q4.2: Could the authors moderate the claims regarding the online personalization to better reflect the dataset-specific statistical results?

A4.2: Regarding personalization, we believe that the current statements are already fairly objective. The advantage of online personalization lies not only in effectively improving predictive performance, but more importantly in its substantially lower deployment cost than participant-specific fine-tuning. In addition, the Results section explicitly reports that, relative to participant-specific fine-tuning, online personalization performs better on CGMacros, is nearly on par on BIG IDEAs, and is slightly worse on T1D-UOM, while requiring 583–982 times less computation and 971–14,118 times less storage. This demonstrates the advantages of online personalization. Therefore, we will retain the relevant statements in the main text.

Q4.3: Could the authors moderate the claims regarding the cross-dataset generalization to better reflect the dataset-specific statistical results?

A4.3: We agree that the cross-dataset statement should more accurately reflect the dataset-specific results, and have updated it in the PDF to: “MealGluco shows strong cross-dataset generalization, reducing Trajectory-MAE by 2.2--13.4% relative to the strongest standard baseline in all six transfers. After the two matched-personalization variants are included, MealGluco remains best in five of six transfers. Overall, these results support broad transfer across populations and devices.” (page 9)

Q5: Have the authors considered compare your prediction model with FDA-approved models, such as UVA/Padova Type 1 Diabetes Simulator?

A5: Thank you for the suggestion. First, we would like to clarify that the postprandial glucose trajectory prediction model studied in this paper uses only pre-meal CGM, meal nutrient composition, and participant history, and is evaluated on three real-world datasets. It does not use insulin dose as an input or model glucose--insulin dynamics. UVA/Padova, in contrast, is an insulin-aware physiological simulator that requires individualized physiological parameters and an explicit insulin treatment regimen. It therefore differs from our task definition, available inputs, and evaluation objective. Using it as a comparator would require changing the task inputs and introducing insulin information that is not consistently available across the datasets. Therefore, we did not include UVA/Padova as a direct baseline. We will consider this simulator as an important baseline in future work on insulin-related models.

Q3.2: In particular, how do their posterior uncertainties and estimates evolve as more meals are assimilated?

A3.2: To quantitatively evaluate whether the two participant-specific coordinates become stable as more completed meals accumulate, we define mean absolute coordinate change per additional completed meal. For two adjacent observation points Ki<KjK_i<K_j, this metric is calculated as $ \frac{|u(K_j)-u(K_i)|}{K_j-K_i}$ and averaged across participants with the corresponding meal history, separately for uρu_\rho and uru_r. The complete results have been updated in Table 19 of Appendix C.2.2 (page 33). For the reviewer's convenience, we also include the table of uρ/uru_\rho/u_r results below. In addition, to make the trajectories of uρu_\rho and uru_r easier to inspect, we selected one participant from each of the three datasets and show their complete coordinate trajectories as the number of completed meals increases in Figure 7 (page 33). As can be seen, both coordinates gradually stabilize after the first few meals.

DatasetK=0K=0–11K=2K=2–44K=4K=4–88K=8K=8–1616K=16K=16–3232
CGMacros0.270 / 0.2360.178 / 0.1470.066 / 0.0730.021 / 0.0240.013 / 0.011
BIG IDEAs0.370 / 0.2100.116 / 0.1910.065 / 0.0720.031 / 0.0200.015 / 0.012
T1D-UOM0.177 / 0.2400.080 / 0.1040.092 / 0.0250.033 / 0.0160.010 / 0.005
Reviewer_X6YWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies meal-conditioned forecasting of the complete two-hour postprandial CGM trajectory (24 points at 5-minute resolution) from a two-hour pre-meal CGM window, meal macronutrients (carbohydrate, protein, fat, fiber), clock time, and a fasting-glucose covariate. The proposed model, MealGluco, decomposes the forecast into a bounded residual decoded from a frozen pretrained CGM-JEPA encoder plus recent-history statistics and time queries, and a meal response represented by a peak-normalized Gamma kernel whose amplitude, peak time, and shape are predicted from nutrients and pre-meal state, with protein/fat/fiber modifiers

Critical strengths

Figures are definitely a strength of this paper. Figure 1 is pretty cute and well-designed, and it is publication-quality. Figure 2 is clean and informative (font sizes could be a bit larger). Other charts are good and have nice color patterns.

The two-coordinate online personalization is a distinctive, well-executed idea that stands on its own.

Updating only a participant-level gain and post-peak clearance factor via a closed-form, strictly causal, backprop-free information update, with the shared network frozen and a 48-byte persistent state, is a clean mechanism for adapting to new users. It is also validated more carefully than most personalization claims: significant gains over the population model on all five CGMacros metrics, on Trajectory-MAE, iAUC-MAE, and Recovery-MAE for T1D-UOM, and on iAUC-MAE for BIG IDEAs (Table 21); a wrong-history negative control (Table 15) showing the gains are identity-specific rather than an artifact of extra data; and accuracy on par with per-participant fine-tuning overall. This is a strength because it is the component most likely to be reused. Even if the Gamma inductive bias proves marginal, as the ablations suggest, this module could be attached to any population forecaster.

The evaluation protocol is rigorous and transparent enough to be audited. Participant-disjoint CV with fold-local preprocessing, covariance fitting, and checkpoint selection. Heterogeneous public datasets (healthy, T2D, and T1D populations, multiple CGM devices). Five complementary trajectory-shape metrics; cross-dataset transfer. Three seeds. Participant-clustered bootstrap intervals for every major comparison and an appendix that spells out every constant and procedure

Apart from the R_max sweep on CGMacros' outer folds (a five-value sweep, so the bias is probably modest), I found no leakage of test data into training. This is critical because the reported numbers are credible as measurements, and it means the paper's main problems are correctable by re-framing claims and adding matched baselines rather than by redoing the study.

Critical weaknesses

The paper's central thesis is not fully supported and is contradicted by its own appendix, and the headline comparison is not information-matched.

The main text asserts that the structured Gamma kernel and nutrient modifiers are what make the model work, and that removing any proposed module results in a clear performance drop across all three datasets. The appendix shows otherwise: on Trajectory-MAE no ablation is significant on T1D-UOM and only three of ten are on BIG IDEAs; the unstructured meal head is significantly better than the Gamma kernel on four dataset–metric pairs (CGMacros Recovery, BIG IDEAs Peak and Recovery, T1D-UOM Peak) against one in its favour; only 6 of the 15 main comparisons are significant, none on BIG IDEAs and one significantly against; and per-participant fine-tuning is significantly better on T1D-UOM (all stated otherwise in the abstract and Section 4).

MealGluco adapts to each test participant's history. At the same time, every baseline is a frozen population model. In the matched population-only row, the model still leads on CGMacros (22.63 vs. 23.48) but loses on BIG IDEAs (16.87 vs. 16.49) and roughly ties on T1D-UOM (31.01 vs. 31.19), so the clean sweep in Table 1 depends substantially on a step the baselines were denied.

Two further issues reduce trust in the numbers as printed: R_max was selected on the same CGMacros outer folds used for the main results (a five-value sweep, so probably modest in effect, but the selection procedure for the many other constants is unstated), and the CGMacros ablation means disagree with the bootstrap deltas beyond rounding. In contrast, the other two datasets match exactly. This is problematic because what survives scrutiny is a modest, dataset-dependent gain (strongest on the one dataset used for design decisions) rather than the physiological inductive bias the paper is built around.

The positioning overstates novelty, and the physiologically structured framing does not hold for the T1D evidence. The claimed task formulation omits closely related work on long-horizon postprandial trajectory prediction (e.g., Chen et al., A Switched Neural Process for Long-Term Postprandial Glucose Trajectory Prediction Based on Insulin Sensitivity Estimation. IEEE Transactions on Instrumentation and Measurement, 2025. doi:10.1109/TIM.2025.3554905) and does not connect the personalization module to nonlinear mixed-effects modeling.

On the physiological side, the nonnegative kernel plus +/-20 mg/dL state-bound cannot represent post-meal drops larger than $\sim$20 mg/dL below the anchor ($\sim$40 after personalization), so post-bolus hypoglycemia is structurally out of reach, and zero-carbohydrate meals are hard-wired to produce no meal-kernel response. This is an issue because it reduces the contribution to a sensible engineering combination with dataset-dependent evidence, and because one of the three datasets on which generality is claimed is evaluated using a physiologically incomplete model with an inaccurate data description

Questions for authors

Please report the population-only MealGluco row in Table 1 and at least one baseline given the same history, e.g., a per-participant bias/scale correction on the Stacking Framework or per-participant fine-tuning of UniCA-Gluformer. Does the advantage survive on BIG IDEAs and T1D-UOM?

The paper states that T1D-UOM "contains seven participants" and "does not provide insulin records for the meals used here." The public dataset describes 16-17 participants with bolus-insulin files for most. Please clarify: how were the seven participants selected, and were bolus records unavailable for the selected meals, excluded, or not used? How do T1D-UOM results change for all models if bolus insulin is added as an input?

R_max was selected on the CGMacros outer folds used for the main results. Which other constants were chosen using outer-fold results rather than inner validation? Could you report CGMacros results with all selections restricted to inner validation?

Please explain: Table 12 vs. Tables 28-32 on CGMacros only (e.g., w/o nutrient modifiers: +0.12 from the means vs. +0.36 bootstrap; w/o post-meal time: +0.84 vs. +0.63), while BIG IDEAs and T1D-UOM rows match exactly; Table 5's R_max = 20 row vs. the Table 12 population row; "1.8-7.9%" vs. the 2.2-13.4% implied by Table 3; 101,896 vs. 102,192 trainable parameters; "51 outer folds" in Table 4.

Is the Gamma structure justified beyond CGMacros? The unstructured head is significantly better on CGMacros Recovery, BIG IDEAs Peak and Recovery, and T1D-UOM Peak, and on Trajectory-MAE; no ablation is significant on T1D-UOM. How do you reconcile this with the claim that accuracy "relies on the structured Gamma parameterization," and will the main-text claims be revised to match Tables 20-22 and 28-32?

Which parameters were fine-tuned, with how many steps and what learning rate, and were updates matched to the number of completed meals? Why does the persistent state differ by ~15× across datasets?

What fraction of raw logged meals survives the 180/120-min isolation and completeness filters per dataset, and what fraction of meals is in M_d (used for Near-Peak Distance and Recovery-MAE)?

Does CGM-JEPA's pretraining corpus (Stanford, Colas) include any participants from the three evaluation datasets?

How does the task differ from long-horizon postprandial trajectory prediction, such as Chen et al.'s switched neural process, and why is the personalization module not framed as a linearised nonlinear-mixed-effects estimator?

6 comments
AuthorsOpenReview ↗

Response to Reviewer X6YW

Thank you for carefully reviewing our work and for your constructive feedback. We have made corresponding revisions to the paper based on your feedback, and all revisions are highlighted in red in the updated PDF. We respond to your questions point by point below.

Q1: Please report the population-only MealGluco row in Table 1 and at least one baseline given the same history, e.g., a per-participant bias/scale correction on the Stacking Framework or per-participant fine-tuning of UniCA-Gluformer. Does the advantage survive on BIG IDEAs and T1D-UOM?

A1: Following your suggestion, we added the population-only MealGluco results to Tables 1, 8, and 9, denoted as MealGluco w/o pers., and equipped UniCA-Gluformer and Stacking Framework with the same personalization module as MealGluco, denoted as UniCA-Gluformer (Pers.) and Stacking Framework (Pers.), respectively. For your convenience, the results on the three datasets are shown below. The complete results are available on pages 7 and 29 of the revised PDF. Considering all five metrics, MealGluco still has an overall advantage on all three datasets.

DatasetModelTrajectory-MAEiAUC-MAEPeak-MAENear-Peak DistanceRecovery-MAE
CGMacrosMealGluco21.16 ± 0.171702 ± 2125.00 ± 0.3217.54 ± 0.3319.18 ± 0.32
MealGluco w/o pers.22.63 ± 0.231873 ± 2026.97 ± 0.0820.08 ± 0.2920.48 ± 0.21
UniCA-Gluformer (Pers.)23.77 ± 0.431980 ± 3530.80 ± 0.5719.27 ± 0.1020.79 ± 0.10
Stacking Framework (Pers.)22.88 ± 0.021911 ± 329.27 ± 0.1119.85 ± 0.2219.91 ± 0.07
BIG IDEAsMealGluco16.56 ± 0.181192 ± 1618.76 ± 0.1016.65 ± 0.3918.66 ± 0.40
MealGluco w/o pers.16.87 ± 0.461247 ± 5919.03 ± 0.5915.21 ± 0.1518.02 ± 0.13
UniCA-Gluformer (Pers.)18.21 ± 0.391363 ± 8322.75 ± 1.6616.30 ± 0.6019.87 ± 0.32
Stacking Framework (Pers.)16.30 ± 0.061236 ± 619.60 ± 0.1017.42 ± 0.1317.98 ± 0.14
T1D-UOMMealGluco30.08 ± 0.551987 ± 4531.87 ± 0.5324.83 ± 1.5424.06 ± 0.55
MealGluco w/o pers.31.01 ± 0.792093 ± 8433.15 ± 0.9825.82 ± 1.9624.54 ± 0.47
UniCA-Gluformer (Pers.)32.13 ± 0.252252 ± 838.34 ± 0.1525.75 ± 3.2924.69 ± 0.16
Stacking Framework (Pers.)30.26 ± 0.012097 ± 534.84 ± 0.2124.00 ± 0.4924.07 ± 0.03

Q2.1: How were the seven T1D-UOM participants selected from the public cohort of 16–17 participants?

A2.1: The original T1D-UOM dataset contains more participants, but our main experiments retained only the 7 participants who met both criteria: complete records for the four macronutrients and complete continuous pre-meal/post-meal CGM windows. The selection criteria and final sample size have been added to Appendix B.1; see page 24 of the revised PDF.

Q2.2: For the selected meals, were bolus-insulin records unavailable, excluded during preprocessing, or simply not used as model inputs?

A2.2: Some meals in the original T1D-UOM dataset have bolus-insulin records, but we did not use them. Our study primarily focuses on the effects of different meals on blood glucose.

Q2.3: How would T1D-UOM results change for all models if bolus insulin were added as an input?

A2.3: Our main research question concerns the effects of meal composition and pre-meal CGM on postprandial glucose trajectories. Adding bolus insulin to the inputs of all models would require careful design of the input location and architecture, as well as extensive experiments, which are beyond the scope of our work. However, to address your question, we tried adding bolus insulin as an input to MealGluco and used the dose and injection time to estimate an additional glucose-lowering amount for correction. The results on T1D-UOM are shown below.

MethodTrajectory-MAEiAUC-MAEPeak-MAENear-Peak DistanceRecovery-MAE
MealGluco30.238 ± 0.3411998.4 ± 24.632.178 ± 0.21025.374 ± 0.83724.244 ± 0.345
MealGluco + insulin30.043 ± 0.2781966.5 ± 21.231.820 ± 0.33925.361 ± 0.62624.097 ± 0.424

The results show that adding bolus insulin information leads to a small improvement in the mean of all five metrics on T1D-UOM.

Q3: Rmax⁡R_{\max} was selected on the CGMacros outer folds used for the main results. Which other constants were chosen using outer-fold results rather than inner validation? Could you report CGMacros results with all selections restricted to inner validation?

A3: We thank the reviewer for pointing out that the selection protocol for Rmax⁡R_{\max} was unclear, and we apologize for this. To clarify, the value Rmax⁡=20R_{\max}=20 mg/dL on CGMacros was selected solely based on the inner-validation Trajectory-MAE results; the outer-fold results were not used in the selection. We clarified this issue in Appendix B.4.2 and added the inner-validation Trajectory-MAE results to Table 5 (page 27). For your convenience, we show the inner-validation Trajectory-MAE results below. The inner-validation Trajectory-MAE varies by only 0.345 mg/dL across the six candidates, indicating that model performance is not sensitive to Rmax⁡R_{\max}.

MetricRmax⁡=10R_{\max}=10Rmax⁡=15R_{\max}=15Rmax⁡=20R_{\max}=20Rmax⁡=25R_{\max}=25Rmax⁡=30R_{\max}=30Rmax⁡=40R_{\max}=40
Inner-val Trajectory-MAE21.84621.69621.50121.52321.51521.634

Q4.1: Please explain why Table 12 and Tables 28–32 contain numerical inconsistencies on CGMacros, while the corresponding BIG IDEAs and T1D-UOM values match exactly.

A4.1: We apologize for this. During the aggregation of the CGMacros data, we mistakenly included results from different experimental batches in the confidence-interval tables, which led to the discrepancy. This issue did not occur for BIG IDEAs or T1D-UOM. All CGMacros results have now been updated using the latest data. The table numbering changed in the revised version, and the complete results are now reported in Table 19 and Tables 35–39 (pages 33 and 39).

Q4.2: Please clarify why the main text reports a cross-dataset improvement of "1.8–7.9%", whereas the results implied by Table 3 range from "2.2-13.4%".

A4.2: The original range "1.8–7.9%" also included our method variant, MealGluco w/o pers., as a comparator, but it is an ablation model rather than a baseline. To avoid causing the same confusion for other readers, we have revised the range throughout the paper to "2.2-13.4%" as you suggested (pages 9 and 10).

Q4.3: Please explain why the reported number of trainable parameters differs between 101,896 and 102,192.

A4.3: These two numbers come from different parameter-counting conventions. 101,896 is the number of effective trainable parameters that actually receive nonzero gradients in the final computation path; 102,192 is the total number of parameters marked with requires_grad=True in the model. The additional parameters are structural compatibility parameters retained for compatibility with older model versions and ablations, and are not involved in the final prediction. To avoid confusing readers about the parameter count, we have updated the paper to consistently use the effective trainable parameter count, and no longer report 102,192.

Q4.4: Please clarify what "51 outer folds" in Table 4 refers to and what exactly this number counts.

A4.4: The three datasets use 5, 5, and 7 participant-disjoint outer folds, respectively, for a total of 5+5+7=175+5+7=17 outer folds. Each fold is trained and evaluated with three seeds, resulting in 17×3=5117\times3=51 fold runs.

Q5.1: Is the Gamma structure justified beyond CGMacros? The unstructured head is significantly better on CGMacros Recovery, BIG IDEAs Peak and Recovery, and T1D-UOM Peak, and on Trajectory-MAE; no ablation is significant on T1D-UOM.

A5.1: Most meals in T1D-UOM involve insulin intake, whereas the main focus of this work is modeling the effect of diet on postprandial glucose in a broader population. Therefore, MealGluco faces greater heterogeneity in this cohort. In addition, after preprocessing, only 7 participants remain in T1D-UOM, making it more difficult to obtain stable statistical significance with a small number of participants. This makes T1D-UOM a challenging external validation cohort, but this alone does not imply that the Gamma structure has no value. There is currently no larger public dataset that satisfies our requirements for complete meal records and CGM windows and could replace this evaluation.

Q5.2: How do you reconcile this with the claim that accuracy "relies on the structured Gamma parameterization," and will the main-text claims be revised to match Tables 20-22 and 28-32?

A5.2: First, we would like to clarify that the “accuracy relies on the structured Gamma parameterization” statement you quoted does not appear in the original paper. The original text emphasizes the joint contribution of the structured Gamma parameterization and nutrient-specific modulation: "The unstructured meal-response head and nutrient-modifier ablations show that accurate forecasts rely on both the structured Gamma parameterization and nutrient-specific modulation" (page 9, line 449). We believe this does not single out the Gamma structure, because it describes the joint contribution of the two components to the model's overall trajectory prediction.

Q6.1: Which parameters were fine-tuned, with how many steps and what learning rate?

A6.1: For per-participant fine-tuning, we evaluated three parameter scopes: all_head updates all trainable response/head parameters except the frozen CGM-JEPA encoder; gamma_mlp updates the MLP that generates the Gamma parameters; and gamma_output updates only the final Gamma output layer. The learning-rate candidates were 10−410^{-4}, 5×10−45\times10^{-4}, and 10−310^{-3}, and each update used either 1 or 3 gradient steps. Combining the three parameter scopes with the four learning-rate/step configurations yielded 12 candidates, which were selected entirely using inner validation. We have added the specific details to Appendix C.2 (page 30).

Q6.2: Were updates matched to the number of completed meals?

A6.2: The update schedule strictly matches the number of completed meals. Each time a meal's 120-minute outcome window becomes fully available, exactly one update event is triggered. For each participant, the fine-tuning model is initialized from the population checkpoint only once, and subsequent updates use only historical meals whose outcome windows have ended. The online personalization method and fine-tuning use exactly the same sequence of historical meals.

Q6.3: Why does the persistent state differ by ~15× across datasets?

A6.3: The approximately 15-fold difference in persistent state arises because inner validation selected fine-tuning parameter scopes of different sizes. Fine-tuning must store the selected parameters' weights, Adam first moments, and Adam second moments, totaling 12 bytes per parameter. all_head was selected for 8/15 runs on CGMacros, whereas it was never selected for T1D-UOM, resulting in a difference of 677,651/46,617≈14.5×677{,}651/46{,}617\approx14.5\times. Detailed configurations and selection statistics have been added to Appendix C.2 (page 30).

Q7: What fraction of raw logged meals survives the 180/120-min isolation and completeness filters per dataset, and what fraction of meals is in M_d (used for Near-Peak Distance and Recovery-MAE)?

A7: All data-filtering rules were fixed before model training and evaluation and were applied consistently to all three datasets. For CGMacros, 1,650 of 1,706 raw meal records were retained (96.7%); for BIG IDEAs, 534 of 626 aggregated meals with complete nutrient records were retained (85.3%); and for T1D-UOM, 1,487 of 4,351 raw records were retained (34.2%). MdM_d is the predefined valid subset for these metrics. Its size is 1,404/1,650 (85.1%) for CGMacros, 416/534 (77.9%) for BIG IDEAs, and 1,095/1,487 (73.6%) for T1D-UOM.

Q8: Does CGM-JEPA's pretraining corpus (Stanford, Colas) include any participants from the three evaluation datasets?

A8: Based on the publicly available information, there is currently no evidence that the pretraining participants overlap with the three evaluation datasets or that data leakage occurred. The original CGM-JEPA paper states that its pretraining data came from independent Stanford and Colas cohorts, rather than the CGMacros, BIG IDEAs, or T1D-UOM datasets used in this paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.