acceptodds
Under review as a conference paper at ICLR 2027

GlucoPRISM: Protocol-Supervised Pretraining for Cross-Cohort Glucose Transfer

Abstract

Continuous glucose monitoring (CGM) foundation models encode each day as one vector that mixes who the person is, what happened that day and how the sensor measured it, so measurement differences travel with the representation into new cohorts. We ask whether pretraining can place the measurement component that harms transfer in known coordinates, so that deployment drops them with a fixed projection and no target data. GlucoPRISM supervises a GlucoFM backbone with response contracts, rules for how each part of the representation should change when only the day or only the sensor changes: a 64-dimensional Trait band should change with neither, a 48-dimensional State band only across days, and a 16-dimensional Sensor band should track the device. Synthetic sensor views calibrated on paired Dexcom and Libre recordings, different-day pairs and glucose-summary targets supply this supervision without clinical labels. Across twelve cross-cohort directions, dropping the Sensor band from the same frozen encoder raises mean transfer AUROC from 68.75 to 73.57, against 69.3 for GlucoFM. The gain has two parts. Into CGMacros, the one target cohort recorded with two sensor types, filtering adds 11 to 19 points, and this holds when the generator is recalibrated on half of the participants and scored on the other half. On the other eight directions filtering is neutral and the encoder itself exceeds GlucoFM by 3.5 points. The device remains linearly decodable after filtering, so the filter removes a device-related component that harms transfer rather than all measurement information. The retained bands improve window AUROC on thirteen of fourteen public task–cohort cells (Holm-adjusted ) and correlation on fourteen of fifteen clinical endpoints in an external cohort of 9,303 participants.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Continuous glucose monitoring, foundation models, representation learning, cross-cohort transfer, domain shift, sensor invariance, disentanglement, self-supervised pretraining, digital health

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_3vDQOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CGM features are often varies across different cohorts, devices or sampling rate. When measurements changes dues to different conditions, such invariance will move with the model. Unlike previous method may neglect learning of useful signals, this paper aims to predefine the protocol to add on the existing glucose foundation model to solve the cross-domain generalization problem. Then the paper evaluated this mechanism across serveral cross-dataset transfer settings and demonstrate the performance of propose GlucoPRISM.

Critical strengths

The control experiments and evaluations across datasets demonstrate the effectiveness of the proposed mechanism and other experiments of reporting within-cohort the negative results also provide useful insights for the community.

Critical weaknesses
  1. The motivation and research gap is not clearly defined and supported. In line 43-44, although the paper propose these glucose measurement varies, but how exactly such 'moves with the mode' corresponds to specific learning problem or lead to any problems such as final prediction performance degradation or problem or learnt representation?
  2. The three bands, trait, state and sensor, are they well-defined or established protocols in the field? and does these selection of protocols mainly based on authors observation or chosen by authors themselves, the intuition of using these as supervision is also lack of support and validation.
  3. The protocol reads like a add-up supervision component for all glucoseFM, but the paper only choose one backbone and does not show the generalization of the protocol supervision.
  4. The training framework and process can be framed better, for example, clearly state the backbone and added components are trained jointly during pertaining.
Questions for authors

My questions are listed above.

5 comments

We thank the reviewer for recognizing the cross-dataset controls and the value of reporting negative results. We will clarify the learning problem, the status of the bands, the scope of generalization, and the joint training procedure.

W1. Motivation and research gap.

The learning problem is that sensor–cohort confounding can encourage source-fitted clinical probes to exploit sensor-related features that fail to transfer. The observation mask alone predicts device identity at 100 AUROC (Appendix T). More directly, with the encoder fixed and each readout probed using the same source data and fitting procedure, removing Sensor increases transfer AUROC from 64.71 to 78.83 into CGMacros and from 68.75 to 73.57 overall; filtering is nearly neutral elsewhere (+0.17; Table 8). This connects the representation problem to prediction degradation.

The gap is how to learn a selectively removable sensor-related component during pretraining, avoiding post-hoc eraser estimation. INLP/LEACE fit device-labeled removal operators; domain adversaries suppress device predictability. In separate same-corpus retrains, our fixed filter reaches 73.14 transfer AUROC versus 69.20 for GlucoFM+LEACE and 67.19 for domain-adversarial GlucoFM (Tables 22–25). Erasure on GlucoPRISM itself gives similar performance; the distinction is that our removal operator is predefined. It requires no target-specific fitting, although the sensor-view generator uses CGMacros paired recordings for calibration; held-out-participant controls retain the filtering gain (Appendix G.1).

W2. Band definitions, motivation and validation.

Trait, State and Sensor are our operationally defined representation bands, with author-chosen widths (64/48/16), rather than established clinical protocols. They are motivated by person-specific glycemic patterns (Hall et al., 2018), within-person variability (Danne et al., 2017; Battelino et al., 2019), and sensor differences (Boscari et al., 2018). “Protocol” denotes acquisition settings—sensor, sampling rate and calibration. Table 1 specifies the intended responses: Trait should remain stable across days and sensors; State may vary across days while remaining sensor-stable; Sensor should respond to sensor changes. Simulated sensor views and repeated-day pairs supply this supervision.

The evidence tests both representation behavior and objective contributions. Across 44 paired HPP visits, median cosine similarity is 0.985/0.923/0.435 for Trait/State/Sensor (Table 39). In Appendix O’s separate 24-cell ablation, removing sensor-view alignment lowers retained transfer by 6.70 points (conditional participant-bootstrap 95% interval [4.14, 9.10]). Separation is incomplete: device identity remains decodable after filtering (Appendix K.2). The two-seed Sensor/State allocation sweep tests widths of 8/16/32 for Sensor with Trait fixed; it does not establish an optimal allocation (Table 26). We will state these assumptions and limits explicitly.

W3. Backbone scope and new controls.

We evaluate a GlucoFM-based foundation model; portability of the objectives to other backbones remains untested. GlucoFM was the strongest window-level baseline in our benchmark (65.9 AUROC versus 58.5–64.7; Table 2). To isolate supervision within this architecture, new controls use three seeds and the main 12-direction transfer protocol; the final row below is the main-model reference. All entries are AUROC (%).

ConfigurationTransferInto CGMacrosWindow
Same architecture and recipe, protocol objectives disabled67.3365.2965.12
+ sensor-view alignment, full readout72.5676.0966.55
+ sensor-view alignment, Sensor removed73.5078.7566.17
Full GlucoPRISM, Sensor removed73.5778.8368.01

Adding alignment raises full-readout transfer by 5.23 points [0.09, 11.63]. These alignment rows retain variance regularization; a separate control with alignment as the sole added objective still gains 4.65 [1.88, 8.17]. These new intervals resample directions and seeds. The complete objective also improves window AUROC from 66.17 to 68.01. This supports supervision’s contribution within GlucoFM; we will add the controls to Appendix O and explicitly limit backbone-generalization claims.

W4. Joint pretraining and deployment.

The online GlucoFM backbone, pooling, band projections, Sensor Gaussian channel and auxiliary prediction heads are optimized jointly from random initialization in one 300-epoch stage using Eq. 5; the target encoder follows exponential-moving-average updates (Appendices D/F). At inference, auxiliary prediction heads are discarded, while the learned representation projections are retained and Sensor uses its posterior mean. The encoder is frozen, and the fixed projection PTS=[I112  0]P_{TS}=[I_{112}\;0] retains Trait and State before fitting the downstream probe on source data. We will make these stages explicit in Section 3.3 and Figure 2.

Dear Reviewer 3vDQ,

Thank you for your thoughtful and constructive feedback, particularly your concerns regarding the motivation, the definition of the bands and the training procedure. As the discussion period draws to a close, we would greatly appreciate it if you could revisit our responses and let us know whether any concerns remain unresolved.

In our responses, we have linked sensor-related features directly to prediction degradation in cross-cohort transfer, positioned our fixed filter against fitted erasure and domain-adversarial training, and explained how the Trait, State and Sensor bands are defined, motivated and validated. We also report new matched controls that isolate the contribution of the protocol supervision on the GlucoFM , and we clarified that the backbone and all added components are trained jointly in one pretraining stage. All of these analyses will be included in the revised paper.

We hope our responses and the additional evidence have adequately addressed your concerns. We would be happy to answer any further questions.

Thank you again for your time.

AuthorsOpenReview ↗

Followup on Rebuttal

Dear Reviewer 3vDQ,

Thank you for your thoughtful and constructive feedback, particularly your concerns regarding the motivation, the definition of the bands and the training procedure. As the discussion period draws to a close, we would greatly appreciate it if you could revisit our responses and let us know whether any concerns remain unresolved.

In our responses, we have linked sensor-related features directly to prediction degradation in cross-cohort transfer, positioned our fixed filter against fitted erasure and domain-adversarial training, and explained how the Trait, State and Sensor bands are defined, motivated and validated. We also report new matched controls that isolate the contribution of the protocol supervision on the GlucoFM , and we clarified that the backbone and all added components are trained jointly in one pretraining stage. All of these analyses will be included in the revised paper.

We hope our responses and the additional evidence have adequately addressed your concerns. We would be happy to answer any further questions.

Thank you again for your time.

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

The paper points out the current shortcoming in cross-cohort generalization of continuous glucose monitoring (CGM) foundation models, where participant characteristics, day-specific states, and sensor-related measurement effects are entangled in a single representation. The proposed method introduces protocol-supervised pretraining on a GlucoFM backbone, partitioning the representation into Trait (64 dimensions), State (48 dimensions), and Sensor (16 dimensions) bands. Synthetic sensor views, different-day pairs, glucose-summary prediction, and a Gaussian bottleneck encourage these bands to follow specified response contracts.

The paper conducts comprehensive experiments. The results show that removing the sensor band improves mean transfer AUROC from 68.75 to 73.57, with gains of 11–19 points on transfers into CGMacros. The retained representation also improves window-level AUROC over GlucoFM on 13 of 14 public benchmark cells and improves correlation on 14 of 15 clinical endpoints in the external cohort. However, device information remains decodable after filtering and transfer comparisons are based on only 12 non-independent directions. Overall, the results support protocol-supervised representation partitioning as a promising approach to CGM transfer, while leaving questions about generalizability and the interpretation of the learned bands unresolved.

Critical strengths

The paper is well-motivated and has a clear scope. The proposed method supervises representation bands according to their responses to sensor and day changes, enabling a fixed projection at deployment. Comparing full and filtered readouts on the same frozen encoder provides a relatively clean test of the filtering intervention.

Comprehensive evaluation with informative failure analysis. The paper evaluates multiple CGM cohorts, compares against established foundation-model baselines and erasure methods, and examines clinical information in an external cohort. Importantly, it reports that filtering helps primarily when sensor protocols differ and acknowledges residual device information and limitations in short-horizon glucose prediction.

Critical weaknesses
  1. The largest filtering gains occur in the four transfer directions into CGMacros, the only target cohort with two sensor types. The other eight directions show essentially no filtering benefit. Validation on an independently collected cohort with different devices and protocols would be needed to establish broader generalizability. Moreover, none of the transfer comparisons survives Holm correction, limiting the strength of the statistical evidence.

  2. Device identity remains highly decodable from the filtered representation. Device-group information is also present in every band in the external cohort. Furthermore, Sensor retains useful physiological information for some clinical and short-horizon tasks, while Trait and State remain dependent. These findings raise questions about whether the method isolates a general measurement nuisance or primarily removes a particular transfer-harmful component. The paper would benefit from a clearer characterization of the conditions under which its response contracts yield the claimed benefits.

Questions for authors
  1. Is it possible to evaluate the fixed sensor filter using a generator calibrated on a separate study with different devices, with no participants shared between calibration and evaluation? Would the transfer gains persist under this setting?

  2. The filtered representation achieves approximately 99.97 AUROC for device identification on same-day CGMacros pairs. How do the authors reconcile this result with the intended separation of measurement effects? Can they identify which residual device-related features remain and demonstrate that the removed component, rather than another correlated representation change, explains the transfer gains?

4 comments

We thank the reviewer for the positive review, and for recognising our clear scope, the clean full-versus-filtered test and our informative failure analysis. New results use the released checkpoints, three seeds and the paper's transfer protocol; brackets are 95% intervals over directions and seeds.

W1. Independent cohort and multiple testing

The filtering gain replicates on an independently collected cohort. ShanghaiT2DM (Libre) differs in country, study and population (type 2 diabetes); all but one labelled participant are disjoint from pretraining (excluding that one changes averages by at most 0.17), and it played no role in the generator, calibrated on CGMacros only. Shared IR and HL labels give 10 new directions:

New (3 seeds)FullFilteredGlucoFMFiltered − full
Shanghai → others, IR (3)57.7266.0754.86+8.35 [2.32, 12.04]
Shanghai → others, HL (2)48.1948.2348.56+0.04 [−7.17, 7.57]
Others → Shanghai (5)48.2148.9648.67+0.75 [−2.32, 3.44]

Filtering beats the full readout and GlucoFM in all three IR directions (into CGMacros 72.69 vs 60.74 and 56.29). HL and directions into ShanghaiT2DM are near chance for every model, so they do not test filtering.

Multiple testing. The pre-declared confirmatory family is window AUROC over 14 cells, where only GlucoPRISM survives Holm (pHolm = 0.031); twelve directions sharing three cohorts cannot power a separate Holm family. The intervals agree: filtering +4.82 [1.02, 10.04] and filtered − GlucoFM +4.24 [0.86, 7.39] (Table 8), with a positive estimate whichever target cohort is held out (Table 9); the controlled-study deletion gain is 3.45 [1.89, 4.94] by participant bootstrap (Appendix O.3).

W2 / Q2. Residual device information

The 99.97 is expected, and the removed component explains the gain. Filtering is designed to remove the device-related component that harms transfer, not to make the representation device-blind, as the abstract states.

  • Which features remain (Appendix T). Device identity is in the input: the observation mask alone, or the count of observed samples, identifies the sensor at 100 AUROC; three glucose-level scalars give 99.1; same-day pairs differ by a systematic −31.1 mg/dL. Device-blindness would discard observation count and glucose level; every representation tested, GlucoFM included, decodes the device at 99.9–100.
  • The remainder does not harm transfer. LEACE, fitted with real device labels on all CGMacros windows including same-day ones, removes linear device information yet transfers like the fixed filter (−0.05 [−1.27, 1.18]; INLP +0.33 [−0.83, 1.76]; Table 23).
  • The removed coordinates explain the gain. Only they differ between readouts (same encoder, data and probe). Restoring the 2 least device-predictive Sensor coordinates lowers transfer from 73.57 to 71.63, and 4 to 68.49, near the full readout (68.75; Table 21). Generic suppression does not reproduce the gain (domain-adversarial GlucoFM 67.19 vs 73.14; fitted erasure on GlucoFM 3.94 below the filter).
  • External cohort. HPP device group is confounded with the participant (one manufacturer), leaving little shift to remove (full and filtered both 67.6); its decodability from every band is reported as the contract not established.

W3. Conditions for the benefit

The latter: pretraining places the transfer-harmful, device-related component in 16 known coordinates; it does not remove all measurement information. The contracts are response contracts, not independence claims (Table 1): Trait and State both describe physiology, so dependence is expected (HSIC 0.73), while the required responses hold (within-person cosine across HPP visits: Trait 0.985, State 0.923, Sensor 0.435). When the contracts help:

  1. Target has a new sensor: filtering helps (into CGMacros +14.12 [5.03, 21.34]; ShanghaiT2DM IR +8.35).
  2. No new sensor: neutral (+0.17 [−0.72, 1.01]), as on single-manufacturer HPP.
  3. Within a cohort: free (68.01 vs 67.90).
  4. Tasks that use the Sensor band: it is not a pure device band; keeping it helps short-horizon and meal-response tasks, which also favour GlucoFM.

Q1. Generator calibrated on a separate study

This needs same-day recordings from two sensors on the same participants, which no public cohort other than CGMacros provides. The closest tests show the gain persisting:

  • Participant-disjoint: calibrating on 22 of 44 paired participants and scoring the other 22 leaves the gain unchanged (+17.16 vs +16.80).
  • No paired data: a generator knowing only Libre's public 15-minute sampling rate keeps +13.35.
  • Separate study: ShanghaiT2DM (W1), never used for calibration.

Revisions. Appendix table of all 10 new directions; scope statement; calibration tests made prominent; Appendix K.2: the 99.97 beside Appendix T and LEACE; abstract and Section 1: the W3 claim and conditions.

We hope these answers address the reviewer's concerns, and we would be happy to answer any further questions.

AuthorsOpenReview ↗

Followup on Rebuttal

Dear Reviewer cwdF,

Thank you for your thoughtful and constructive feedback, particularly your questions about generalizability beyond CGMacros and the residual device information after filtering. As the discussion period draws to a close, we would greatly appreciate it if you could revisit our responses and let us know whether any concerns remain unresolved.

In our responses, we have provided ten new transfer directions with the independently collected ShanghaiT2DM cohort, where filtering beats both the full readout and GlucoFM in all three insulin-resistance directions. We also identified which device features remain and why they are present in the input itself, showed with LEACE, INLP and coordinate-restoration tests that the removed coordinates explain the gain, characterized the conditions under which the response contracts help, and reported the participant-disjoint and no-paired-data calibration tests. All of these analyses will be included in the revised paper.

We hope our responses and the additional evidence have adequately addressed your concerns. We would be happy to answer any further questions.

Thank you again for your time.

AuthorsOpenReview ↗

Followup on rebuttal

Dear Reviewer cwdF,

Thank you for your thoughtful and constructive feedback, particularly your questions about generalizability beyond CGMacros and the residual device information after filtering. As the discussion period draws to a close, we would greatly appreciate it if you could revisit our responses and let us know whether any concerns remain unresolved.

In our responses, we have provided ten new transfer directions with the independently collected ShanghaiT2DM cohort, where filtering beats both the full readout and GlucoFM in all three insulin-resistance directions. We also identified which device features remain and why they are present in the input itself, showed with LEACE, INLP and coordinate-restoration tests that the removed coordinates explain the gain, characterized the conditions under which the response contracts help, and reported the participant-disjoint and no-paired-data calibration tests. All of these analyses will be included in the revised paper.

We hope our responses and the additional evidence have adequately addressed your concerns. We would be happy to answer any further questions.

Thank you again for your time.

AuthorsOpenReview ↗

Followup on Rebuttal

Dear Reviewer cwdF,

Thank you for your thoughtful and constructive feedback, particularly your questions about generalizability beyond CGMacros and the residual device information after filtering. As the discussion period draws to a close, we would greatly appreciate it if you could revisit our responses and let us know whether any concerns remain unresolved.

In our responses, we have provided ten new transfer directions with the independently collected ShanghaiT2DM cohort, where filtering beats both the full readout and GlucoFM in all three insulin-resistance directions. We also identified which device features remain and why they are present in the input itself, showed with LEACE, INLP and coordinate-restoration tests that the removed coordinates explain the gain, characterized the conditions under which the response contracts help, and reported the participant-disjoint and no-paired-data calibration tests. All of these analyses will be included in the revised paper.

We hope our responses and the additional evidence have adequately addressed your concerns. We would be happy to answer any further questions.

Thank you again for your time.

Reviewer_2YYZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper looks at CGM pretraining and whether the measurment component that harms cross cohort transfer can be isolated into known coords and dropped. The proposed method splits the model into trait, state and sensor bands and supervises the learning for each head. At deployment the sensor head is dropped.

Critical strengths

The extensive experiments and controls are the key strength. They give pretty good insights into the dynamics and give support to many of the claims made.

The evals themselves are also well thought out and cover most relevant scenarios (but see omissions below).

Critical weaknesses

My key issue with this paper is the assumption/motivation for the work, that pretraining CGM models are harmed in terms of transfer by a latent sensor component. This asserted but never shown and, if anything, there is some evidence that this artefact is introduced by the GlucoPRISM architecture.

GlucoFM transfers better into CGMacros and the 64.71 failure is by GlucoPRISM, which seems odd given the objective.

In the experiment where the raw sensor head is removed before h_A (appendix K), it has no effect. That would suggest the issue is inside the hidden layers of the device classifier, not part of the earlier layers.

Related to this, it is unclear where exactly the encoder gains originate. The paper claims that this is from the alignment but that’s not shown directly. A simple baseline comparison with just GlucoGM and alignment would isolate this much better.

Questions for authors
  1. Is there evidence that the harm is directly connected to the pretraining itself.
  2. What are the per direction stdevs for Hall to CGMacros and Stanford to CGMacros
7 comments

Thank you for identifying the attribution issue. The controls below separate sensor-associated harm in GlucoFM, alignment's contribution and the learned removable band. New experiments use three seeds and 12 transfer directions; brackets are unadjusted 95% intervals from resampling both. GlucoFM, the matched control and GlucoPRISM all use 300 epochs on the same corpus.

W1 / Q1. Harm before GlucoPRISM supervision.

On GlucoFM retrained with our corpus (Appendix K.1), fitted LEACE and INLP improve transfer into CGMacros by 5.25 and 3.86 AUROC points, versus 0.29 and 0.22 elsewhere (Table 23). Removing device-associated information thus helps an encoder without our architecture or objectives. These erasers use CGMacros inputs and device labels; they support device-associated harm without identifying a unique causal sensor factor.

Our new matched-architecture control keeps our pooling/projections, sampler, corpus and schedule but disables the protocol objectives. It scores 67.33 overall and 65.29 into CGMacros; deleting its final 16 coordinates gives −0.78 [−3.27, 2.03]. So the protocol objectives are not necessary for this failure. With protocol supervision but no KL penalty, sensor deletion yields +10.04 [4.96, 15.16] into CGMacros and +0.19 [−0.31, 0.68] elsewhere, supporting learned coordinate-level removability without that penalty.

Adding 31.12 mg/dL to CGMacros Libre windows at test time changes transfer by +0.16 [−2.08, 2.42] (full), +1.17 [−1.89, 3.96] (filtered) and −0.54 [−3.97, 2.10] (control). Correcting the mean offset does not recover the filtering gain, nor identify the remaining mechanism. These controls will replace the single-seed scaling observation as primary evidence for the motivation.

W1b. Full readout vs GlucoFM.

The full readout scores 64.71, below GlucoFM's 73.45; the deployed filtered readout reaches 78.83 (Table 8). We will separate this comparison from the larger within-encoder deletion effect, which shows the readout intervention is useful, not that the unfiltered encoder is uniformly better.

W2. The raw-block experiment.

hAh_A is distinct from the device classifier. The deployed embedding is z=[hT(T);hS(S);hA(A)]z=[h_T(T);h_S(S);h_A(A)]; a separate training-only head classifies devices from zAz_A . Sensor-view and day-pair supervision act on projected bands; other auxiliary losses also constrain raw blocks.

Appendix K compares deletion in the raw [T;S;A][T;S;A] and projected zz readouts, with separately fitted source probes. The negligible raw-block effect, −0.07 [−0.63, 0.69], localizes useful deletion to the final projected representation, not inside the discarded device classifier or uniquely within hAh_A. Figure 2 will distinguish retained projections from training-only heads.

W3. Contribution of alignment.

The build-up experiment fixes architecture and recipe and adds sensor-view alignment of zTz_T and zSz_S, without day, device-classification, cross-band correlation or KL losses. Alignment rows retain variance regularization; a separate contrast removes it. Values are AUROC (%); GlucoFM and full GlucoPRISM are reference rows.

ConfigurationAll 12Into CGMacrosWindow
GlucoFM69.3373.4565.85
Matched control, full67.3365.2965.12
+ alignment, full72.5676.0966.55
+ alignment, Sensor removed73.5078.7566.17
Full GlucoPRISM, Sensor removed73.5778.8368.01

Alignment improves full-readout transfer by 5.23 [0.09, 11.63] before any deletion; without the variance floor, +4.65 [1.88, 8.17]. All models use 300 epochs, so extra training does not explain this. This isolates alignment within our architecture.

Deletion adds 0.94 [−0.69, 4.07] to the alignment model, versus 4.82 in full GlucoPRISM; the complete recipe also raises filtered window AUROC from 66.17 to 68.01. Appendix O's 24-cell study finds alignment contributions of 5.82 [3.60, 7.82] (full) and 6.70 [4.14, 9.10] (filtered), with conditional participant-bootstrap intervals; adding and removing alignment both support its contribution. These comparisons do not assign the complete recipe's larger deletion benefit uniquely to device classification.

Q2. Per-direction SDs into CGMacros.

Mean ± SD over three seeds (Table 18), AUROC / AUPRC (%); IR = insulin resistance.

Source, taskGlucoFMGlucoPRISM
Hall, diabetes77.2±8.8 / 87.9±5.081.7±1.7 / 89.0±2.0
Hall, IR80.3±0.9 / 91.0±0.676.6±4.2 / 87.9±3.5
Stanford, diabetes71.1±3.1 / 84.3±2.285.6±0.8 / 91.6±0.3
Stanford, IR65.3±6.9 / 84.0±2.971.4±2.1 / 86.7±1.0

GlucoPRISM improves both metrics in three of four directions; Hall→CGMacros IR remains a negative result. These are seed SDs, not uncertainty over new cohorts; the paper's transfer comparisons do not survive Holm correction. We will move these rows beside Table 4 and add the new controls, loss configurations and uncertainty procedures to Appendix O.

AuthorsOpenReview ↗

Follow up on Rebuttal

Dear Reviewer 2YYZ,

Thank you for your thoughtful and constructive feedback, particularly your questions about whether transfer harm arises from a sensor component and where the encoder gains originate. As the discussion period draws to a close, we would greatly appreciate it if you could revisit our responses and let us know whether any concerns remain unresolved.

In our responses, we have provided the analyses you suggested: evidence of sensor-associated transfer harm in GlucoFM itself, a matched control on the GlucoFM backbone that adds sensor-view alignment alone, with and without variance regularization, and a test-time offset control. We also clarified that the raw-versus-projected deletion experiment concerns the retained projection, not the training-only device classifier, and reported the requested per-direction standard deviations. All of these analyses will be included in the revised paper.

We hope our responses and the additional evidence have adequately addressed your concerns. We would be happy to answer any further questions.

Thank you again for your time.

AuthorsOpenReview ↗

Follow up on Rebuttal

Dear Reviewer 2YYZ,

Thank you for your thoughtful and constructive feedback, particularly your questions about whether transfer harm arises from a sensor component and where the encoder gains originate. As the discussion period draws to a close, we would greatly appreciate it if you could revisit our responses and let us know whether any concerns remain unresolved.

In our responses, we have provided the analyses you suggested: evidence of sensor-associated transfer harm in GlucoFM itself, a matched control on the GlucoFM backbone that adds sensor-view alignment alone, with and without variance regularization, and a test-time offset control. We also clarified that the raw-versus-projected deletion experiment concerns the retained projection, not the training-only device classifier, and reported the requested per-direction standard deviations. All of these analyses will be included in the revised paper.

We hope our responses and the additional evidence have adequately addressed your concerns. We would be happy to answer any further questions.

Thank you again for your time.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of revisions

We thank the reviewers and the AC for their constructive feedback. We have uploaded a revised PDF in which every change is highlighted in yellow. New tables are numbered R1 and R2, so all table and appendix numbers cited in our responses are unchanged.

  • Motivation (2YYZ, 3vDQ): Section 1 now states the learning problem and the evidence of sensor-associated harm in GlucoFM itself: fitted erasure raises transfer into CGMacros by 5.25 (LEACE) and 3.86 (INLP) AUROC points (Appendix K.1). Section 1 also states when filtering helps. Section 4.5 marks the single-seed scaling study as secondary to these controlled comparisons.
  • Matched controls (2YYZ, 3vDQ): New Appendix O.5 and Table R2 report the matched-architecture control, sensor-view alignment with and without the variance floor, the no-KL arm and the test-time offset control. They also give the loss configurations, the shared 300-epoch budget and the interval procedures. Section 4.3 and the abstract summarize the alignment result (+5.23 AUROC, 4.65 without the variance floor).
  • Readouts and training (2YYZ, 3vDQ): Section 3.3 and the Figure 2 caption state that the backbone and all added components are trained jointly in one stage. They also list which projections are kept at deployment and which heads are used only in training. Section 4.1 separates the full-readout comparison with GlucoFM from the within-encoder deletion effect.
  • Seed variability (2YYZ): The per-direction seed SDs (AUROC and AUPRC) for the four directions into CGMacros are in Table 18, which the Table 4 caption references.
  • Independent cohort (cwdF): New Appendix G.2 and Table R1 report the ten ShanghaiT2DM directions. Filtering improves all three insulin-resistance directions (+8.35 [2.32, 12.04]). The Section 5 scope statement is updated, and the calibration tests are in Section 4.1 and Appendix G.1.
  • Residual device information (cwdF): Appendix K.2 explains why the 99.97 AUROC is expected, alongside Appendix T, the LEACE and INLP results and coordinate restoration (Table 21).
  • Bands and scope (3vDQ): Section 3.1 states that the bands and their widths are our operational definitions, defines "protocol" and adds supporting references. Section 5 states the limits on band widths and that the evaluation uses one backbone.

We would be happy to answer any further questions.

AuthorsOpenReview ↗

Summary of the discussion

Dear Area Chair and Senior Area Chair,

We thank you and all reviewers for your time and for the constructive discussion, which has substantially strengthened our manuscript. The revised paper is uploaded, with all changes highlighted in yellow. New tables are numbered R1 and R2, so all table and appendix numbers cited in our responses are unchanged. We briefly summarize the outcome:

Reviewer 2YYZ (2). The reviewer engaged with our new results and described the contribution of sensor-view alignment as "an interesting finding". The remaining concern was that the paper's framing did not match this result. Following this suggestion, the revision now names sensor-view alignment as the main source of the transfer gain in the abstract, Section 1, the contributions, Section 5 and the Conclusion (+5.23 AUROC over a matched control, Appendix O.5). Sensor filtering is kept as the second part of the gain, measured within the same frozen encoder (+4.82 overall, +14.12 into CGMacros, Table 8). The earlier concerns were addressed with:

  • evidence of sensor-associated harm in GlucoFM itself (LEACE +5.25, INLP +3.86 into CGMacros, Appendix K.1);
  • matched-architecture controls with and without the variance floor (Table R2);
  • a test-time offset control;
  • a clarification of the raw-block experiment;
  • the requested per-direction standard deviations (Table 18).

Reviewer cwdF (3, weak accept). The reviewer described the paper as "well-motivated" with a "comprehensive evaluation with informative failure analysis". The concern on generalizability beyond CGMacros is addressed by an independently collected cohort: filtering improves all three ShanghaiT2DM insulin-resistance directions (+8.35, Appendix G.2, Table R1). The participant-disjoint calibration test leaves the gain unchanged (+17.16 vs +16.80). The residual device information is explained in Appendix K.2 and Appendix T, and the LEACE, INLP and coordinate-restoration tests (Table 21) show that the removed coordinates explain the gain.

Reviewer 3vDQ (1). The reviewer recognized that the control experiments and cross-dataset evaluations "demonstrate the effectiveness of the proposed mechanism". The concerns were mainly about presentation, and each is now addressed in the revised text:

  • Motivation. Section 1 states the learning problem and the evidence of sensor-associated harm.
  • Band definitions. Section 3.1 gives the operational definitions of Trait, State and Sensor, with supporting references.
  • Backbone scope. Section 5 states it explicitly.
  • Joint training. Section 3.3 and Figure 2 state that the backbone and all added components are trained jointly in one stage.

We hope these revisions will be taken into account.

Thank you again for your guidance.

Kind regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.