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

1 INTRODUCTION

• CGM days mix participant, day, and sensor variation, while measurement differences vary across cohorts and are readily learned from observation masks.A linear probe identifies the recording device from the observation mask alone at 100 AUROC in the public corpus.Source: §1 · p. 1
• Response contracts supervise Trait, State, and Sensor bands to change with their intended sources, enabling deployment to drop the 16 Sensor coordinates with a fixed projection.The remaining 112 Trait and State coordinates are kept unchanged; supervision uses synthetic sensor views, different-day pairs, and glucose-summary targets.Source: §1 · p. 2
• 68.75 to 73.57 mean cross-cohort transfer AUROC across twelve directions after removing the Sensor band from the same frozen encoder, without target-cohort fitting.The average combines gains on four directions into CGMacros with little filtering effect on the other eight.Source: §1 · p. 2, Figure 1 · p. 2

2 RELATED WORK

• CGM foundation models learn autoregressive, masked-token, or latent-prediction representations, yet return a single vector in which variation sources remain mixed.GlucoPRISM retains the GlucoFM backbone and changes how its representation is supervised and read out.Source: §2 · pp. 2–3
• Domain-adversarial and invariant-risk objectives suppress nuisance information during training, while post-hoc erasure methods such as INLP and LEACE fit a removal operator afterward.GlucoPRISM instead places measurement information in coordinates fixed during pretraining for filtering at deployment.Source: §2 · p. 3
• Unlike partitions based on domain labels or augmentation-preserved content, GlucoPRISM defines each block by response to simulated sensor and day changes, then removes measurement coordinates with a fixed projection.The paper notes that learned blocks are not identifiable as physiological factors without further assumptions [Locatello et al., 2019; 2020].Source: §2 · p. 3

3.1 FORMULATION: THREE BANDS FROM ONE ENCODER

• A pooled GlucoFM representation is split into 64-dimensional Trait, 48-dimensional State, and 16-dimensional Sensor blocks, with full and filtered readouts differing only by the removed block.Trait should remain unchanged across days and sensors; State may change across days but not sensors; Sensor predicts acquisition.Source: §3.1 · p. 3

3.2 PROTOCOL SUPERVISION: DISPERSING A CGM DAY

• Synthetic sensor views align Trait and State while Sensor predicts device, and different-day pairs keep Trait similar while encouraging State to differ.Glucose-summary heads train raw Trait on level and time-in-range summaries and raw State on variability summaries, without clinical labels; the synthetic sensor generator is fitted on paired Dexcom–Libre windows.Source: §3.2 · pp. 3–4, Table 1 · p. 4

3.3 KEEPING THE SENSOR BAND NARROW, AND FILTERING IT AT DEPLOYMENT

• A KL penalty bounds the Sensor Gaussian, while independence and variance-floor penalties accompany the complete objective; deployment applies a fixed projection to remove Sensor without target data.For CGMacros, the generator still uses calibration data, whereas INLP fits projections using target inputs; the Sensor mean and projection remain active during feature extraction.Source: §3.3 · pp. 4–5, Table 3 · p. 5

4.1 FILTERING THE SENSOR BAND IMPROVES CROSS-COHORT TRANSFER

• Filtering the same frozen encoder raises mean transfer AUROC from 68.75 to 73.57 across twelve directions, while lowering expected calibration error from 0.163 to 0.114 and Brier score from 0.241 to 0.207.The full readout averages 68.75 against 69.3 for GlucoFM.Source: §4.1 · p. 6
• Into CGMacros, filtering adds 11.3 to 18.8 AUROC points; recalibration on half the paired participants preserves the gain when scoring the other half.On the other eight directions, filtering changes transfer by +0.17 points, while the full readout already exceeds GlucoFM by +3.49.Source: §4.1 · p. 6

4.2 THE RETAINED BANDS CARRY CLINICAL INFORMATION IN AN EXTERNAL COHORT

• On HPP, the filtered readout reaches 67.6 mean AUROC across three clinical classifications, and the frozen-encoder probes improve correlation and R2 on fourteen of fifteen baseline endpoints.HPP includes 9,303 participants not seen in pretraining; body-fat correlation rises from 0.260 to 0.427 and liver elasticity from 0.038 to 0.265.Source: Figure 3 · p. 7, §4.2 · p. 8
• GlucoPRISM is weaker on seven meal-response and short-horizon glucose tasks, with postprandial correlation of 0.212 against 0.355 for GlucoFM.Under a separate error-based postprandial protocol, it has lower error than GlucoFM in 15 of 20 settings.Source: §4.2 · p. 8

4.3 THE BANDS RESPOND AS THEIR CONTRACTS PREDICT

• Sensor is nearly independent of Trait and State, and Trait is more stable across two HPP visits than State or Sensor; however, every band decodes device group in that HPP setting.The controlled one-term-off study attributes 6.70 transfer points to sensor-view alignment and 3.44 to variance regularization.Source: §4.3 · pp. 8–9, Figure 4 · p. 8

4.4 FILTERING IS FREE WITHIN A COHORT

• Filtering costs nothing on within-cohort window AUROC: the filtered and full readouts reach 68.01 and 67.90, while the filtered readout exceeds GlucoFM by 2.16 points (Holm-adjusted p = 0.031) across thirteen of fourteen cells.GlucoPRISM also has higher few-shot PR-AUC than GlucoFM at every support size and observation fraction.Source: §4.4 · p. 9

4.5 THE TRANSFER GAP PERSISTS AS THE PRETRAINING CORPUS GROWS

• Across nine nested HPP corpus sizes, GlucoPRISM has higher window and transfer scores; at 8,000 participants its transfer AUROC advantage is 3.24 points.This is a single-seed study, so it does not estimate uncertainty or significance.Source: §4.5 · p. 9

5 DISCUSSION AND CONCLUSION

• Across twelve non-independent transfer directions among three cohorts, no transfer comparison survives Holm correction, and why the harmful device component becomes separable remains open.Filtering helps where source and target measurements differ; elsewhere it is roughly neutral and the pretrained encoder supplies the advantage.Source: §5 · p. 9
• Trait and State remain dependent, Sensor carries physiological signal, and device and cohort are confounded, so Sensor is better interpreted as a measurement-protocol band than a pure device band.State's contribution to transfer is not established, and the retained bands do not remove all measurement information.Source: §5 · p. 9
• Transfer is evaluated among three public cohorts, and testing generator calibration on a second paired cohort with other devices remains the most direct test.Clinical probes assess representation quality, not clinical utility.Source: §5 · p. 10
• The conclusion is that protocol-supervised Trait, State, and Sensor bands permit target-free Sensor filtering, improving transfer into cohorts with a second sensor type while retaining within-cohort accuracy and clinical information.Source: §5 · p. 10

B RELEASE, MODEL IDENTITY AND PROVENANCE

• The released artifacts include code, configuration files, weights, and a shared subject-to-fold assignment; HPP data remain under controlled access, with aggregate results exported.The separately released seed-5 model was selected on the public-benchmark window ROC-AUC and is not used in comparisons.Source: REPRODUCIBILITY STATEMENT · p. 11, B RELEASE, MODEL IDENTITY AND PROVENANCE · pp. 14, 16

• Pretraining and downstream participant sets were intended to be disjoint, but one ShanghaiT2DM participant contributes separate sessions to both, affecting 14-cell averages by at most 0.17 points.Excluding the participant changes any model's ShanghaiT2DM result by at most 0.8 points.Source: C CORPUS, EVALUATION PROTOCOL AND… · pp. 16, 18
• The corpus-size comparison uses one pretraining seed, so its score differences are reported without uncertainty or significance inference.The figure averages window and subject scores over fourteen task–cohort cells and transfer scores over twelve directions.Source: Figure 7 · p. 17
• The reproduced GlucoFM baseline uses a public-only corpus unlike the published corpus, which includes non-public Wear-CGM recordings, so their scores are contextual rather than matched-corpus comparisons.The table reports a reproduced GlucoFM lead of +4.5 AUROC and a published lead of +4.4.Source: Table 11 · p. 17
• REPLACE-BG is capped at 40 windows per subject because uncapped data would contribute roughly 28,000 windows and shift the corpus toward type 1 diabetes.ShanghaiT2DM's 0.333 observation fraction reflects its 15-minute sensor on a five-minute grid, not missingness.Source: Table 12 · p. 17, C CORPUS, EVALUATION PROTOCOL AND… · p. 18
• Probes use repeated five-fold subject-grouped cross-validation with a shared, prewritten fold assignment; Holm correction applies to the declared confirmatory family, while ablations remain exploratory.Shared folds support paired comparisons but do not remove seed or fold uncertainty.Source: C CORPUS, EVALUATION PROTOCOL AND… · p. 18

E SEED VARIABILITY

• Headline comparisons use GlucoPRISM seeds 0, 1, and 2 shared with every baseline; the six-seed row is shown separately because it cannot be paired against three-seed comparators.The six-seed window mean is 68.44 versus 68.01 for the matched-seed row, while its standard deviation is 0.53 versus 0.15.Source: D IMPLEMENTATION AND HYPERPARAMETERS · p. 18, E SEED VARIABILITY · p. 19

D IMPLEMENTATION AND HYPERPARAMETERS

• Internal subject-level probes concatenate each subject's embedding mean, standard deviation, and 10th and 90th percentiles; HPP instead uses the mean alone.The retained readout yields 448-dimensional internal subject inputs, compared with 512 for the 128-dimensional embedding.Source: D IMPLEMENTATION AND HYPERPARAMETERS · p. 18
• The implementation uses a Sensor KL weight of 0.1, cosine-based block distances, and a variance-floor objective; the synthetic generator is fitted on real paired-sensor constants.The measured calibration level shift is not applied to synthetic views.Source: D IMPLEMENTATION AND HYPERPARAMETERS · pp. 18–19
• All downstream encoders remain frozen, with subject-grouped folds and probes fitted on training subjects only.The multi-day analysis averages randomly sampled 24-hour windows rather than adjacent days.Source: D IMPLEMENTATION AND HYPERPARAMETERS · pp. 18–19

E SEED VARIABILITY

• Task-averaged seed variability is the run-to-run spread relevant to headline means; per-cell standard deviations are roughly tenfold larger and would overstate that uncertainty.For GlucoPRISM, the reported comparison is 1.56 per cell against 0.15 task-averaged.Source: E SEED VARIABILITY · p. 19

F OBJECTIVE IMPLEMENTATION AND NOTATION

• Mean and sample standard deviation across all 24 contextual tokens are projected and normalized before the 64/48/16 Trait–State–Sensor split; the Sensor channel uses samples in training and its mean in evaluation.Observation-density weighting and separate block normalization are disabled; exported full and retained features share identical Trait–State coordinates.Source: F OBJECTIVE IMPLEMENTATION AND NOTATION · p. 20
• The repeated-day objective samples another day uniformly for the same subject, then a uniform window, and penalizes State cosine similarity above 0.3 with effective hinge coefficient 0.1.It imposes no minimum time gap or nonoverlap constraint and is skipped when no eligible pairs occur in the batch.Source: F OBJECTIVE IMPLEMENTATION AND NOTATION · p. 20
• The variance floor applies to unit-normalized Trait and State embeddings, with thresholds 0.125 and approximately 0.14434; Sensor is excluded under the active bottleneck.Without the floor, within- and between-subject cosine similarity both reach 0.9999.Source: F OBJECTIVE IMPLEMENTATION AND NOTATION · p. 20
• The cross-block correlation penalty standardizes raw Trait, State, and Sensor across the batch, while Sensor KL is computed analytically from its mean and clipped log-variance at coefficient 0.1.The KL information-bound interpretation applies to the stochastic channel, not automatically to its deterministic evaluation mean.Source: F OBJECTIVE IMPLEMENTATION AND NOTATION · p. 20
• CMP uses two mean-reduced SmoothL1 losses: five level/occupancy targets supervise raw Trait and five dispersion targets supervise raw State; the projected Sensor head predicts eight logits.Audited native labels include Dexcom (9,839 windows), Libre (722), and iPro (391); synthetic partners receive Libre labels, and circadian CMP targets are disabled.Source: F OBJECTIVE IMPLEMENTATION AND NOTATION · p. 20
• Inputs preserve glucose level in mg/dL: a +60 mg/dL perturbation changes the representation by 0.136 with level-preserving preprocessing and 8.7×10−6 when level information is removed.This is an offset-sensitivity control, separate from clinical prediction performance.Source: F OBJECTIVE IMPLEMENTATION AND NOTATION · p. 20

G THE PAIRED-SENSOR GENERATOR, AND ITS ROBUSTNESS

• 43 of 44 subjects show the paired-sensor offset in the same direction; correcting window-start alignment and using Deming rather than ordinary least squares addresses timing and errors-in-variables in generator fitting.The uncorrected alignment produced correlation 0.31 with lag pegged at the search boundary; OLS estimated slope 0.674 versus 0.878 for Deming regression.Source: G THE PAIRED-SENSOR GENERATOR, AND ITS… · pp. 20–21, Figure 8 · p. 21, Table 14 · p. 21

G.1 CALIBRATION WITHOUT THE EVALUATED PARTICIPANTS

• On held-out participants, half-calibration yielded a +17.16 filtering gain versus +16.80 for the all-participant control, while the uncalibrated generator retained +13.35.With three seeds, a calibration contribution of a few points is possible but unresolved; the eight directions not into CGMacros remained near-neutral across arms.Source: §G.1 · p. 21, Table 16 · p. 22, Table 15 · p. 22

H FULL MODEL TABLE AND PAIRED SIGNIFICANCE

• +1.16 PR-AUC on 9 of 14 cells did not survive Holm correction (pHolm = 0.2080); among the three metrics tested together, only one window ROC-AUC comparison survived correction.The PR-AUC column also contains three Holm-significant general-purpose-model deficits against a 0.72M-parameter CGM-specific encoder under class imbalance.Source: H FULL MODEL TABLE AND PAIRED… · p. 22, Table 17 · p. 22

I EXTENDED PUBLIC, DEVICE AND HPP COMPARISONS

• In paired-device evaluation, GlucoPRISM averaged 1.34 AUROC points below GlucoFM and won 3 of 8 comparisons, while filtering improved paired-device transfer within the same encoder.HPP uses locally trained probes on frozen public-corpus encoders; the extended readouts report Sensor-retention benefits for clinical tasks and higher GlucoFM scores on dynamic tasks.Source: I EXTENDED PUBLIC, DEVICE AND HPP… · p. 23

J BLOCK CONTROLS AT MATCHED WIDTH

• Trait and Sensor each outperform matched-width controls: zT by +3.27 and +8.65 AUROC, and zS by +1.67 and +5.74.The controls compare each block with random and PCA projections of matched width and with the same truncation of the baseline embedding.Source: J BLOCK CONTROLS AT MATCHED WIDTH · p. 23
• Trait alone reaches 74.19 AUROC against 72.54 for the released [zT ∥zS] readout; its 1.65-point advantage was consistent across seeds but not statistically resolved.Within the prism view, Trait is defined to remain unchanged across both days and sensors, so it is expected to transfer best on its own.Source: J BLOCK CONTROLS AT MATCHED WIDTH · pp. 23–24
• Transfer results use twelve non-independent directions, and no comparison on this axis survives Holm correction in either direction.The reported baseline is 7.4 ROC-AUC behind GlucoFM.Source: J BLOCK CONTROLS AT MATCHED WIDTH · p. 24
• Deleting Sensor improves the full readout by +13.9 AUROC points for Dexcom →Libre and +2.5 in the reverse direction; these directional averages do not establish general device invariance.The table retains all eight task–direction combinations.Source: J BLOCK CONTROLS AT MATCHED WIDTH · p. 24, Table 20 · p. 24
• Sensor alone reaches 68.88 transfer AUROC, but its predictive information does not establish a pure physiological or device factor.The comparison motivates evaluating deletion directly rather than inferring its value from the block's name.Source: J BLOCK CONTROLS AT MATCHED WIDTH · p. 24

K ERASURE, CALIBRATION AND PARTIAL DELETION

• The fixed filter adds 4.82 AUROC points versus 4.19 for INLP; their paired difference is +0.63 over 12 directions, with a 95% interval of [−0.09, +1.35].The comparison does not establish superiority or equivalence, and the methods use different information: INLP accesses target inputs for CGMacros, while the fixed filter uses a training-specified band.Source: K ERASURE, CALIBRATION AND PARTIAL… · pp. 24–25
• Dropping zA improves expected calibration error from 0.163 to 0.114 and Brier score from 0.241 to 0.207; restoring four coordinates lowers transfer AUROC from 73.57 to 68.49.Removing all 16 coordinates gives the highest transfer and window AUROC among the tested readouts, while the full readout has the highest subject AUROC.Source: K ERASURE, CALIBRATION AND PARTIAL… · p. 25, Table 21 · p. 25

K.1 FITTED ERASURE ON GLUCOFM

• On projected readouts, fitted INLP and LEACE were indistinguishable from the fixed filter, while adversarial GlucoFM reached 67.19 transfer AUROC versus 73.14 for filtered GlucoPRISM; device identity remained 99.97-AUROC-decodable after filtering.The adversarial method also lowered window scores and did not match the fixed filter's calibration; higher weights raised held-out reconstruction loss roughly seven- to tenfold.Source: §K.1 · p. 26, §K.2 · p. 27

L SENSOR-BLOCK CAPACITY AND KL WEIGHT

• Deletion gains were nonmonotonic across β; gains at Sensor widths 8, 16, and 32 were 0.35, 0.51, and 0.30 points, and the two-seed sweep establishes neither an optimal width nor a systematic width effect.At β = 1.0, full and retained-readout AUROC were 66.42 and 67.99, versus 67.52 and 68.03 at released β = 0.1; the larger deletion gain did not mean higher absolute performance.Source: L SENSOR-BLOCK CAPACITY AND KL WEIGHT · pp. 27–28, Table 26 · p. 28

M POST-HOC FACTORIZATION OF A FROZEN ENCODER

• Trait-versus-unsplit post-hoc factorization averaged −0.08 AUROC over 12 configurations, with positive differences in 7; the tested fit used real paired recordings and device labels.The prespecified comparison found no consistent improvement with more paired subjects; pretraining instead combines native device labels with synthetic sensor partners.Source: M POST-HOC FACTORIZATION OF A FROZEN… · p. 28, Table 27 · p. 29

N SENSITIVITY TO STRONGER AUXILIARY-OBJECTIVE WEIGHTS

• Full-weight joint objectives cost 5.28 AUROC at 1× and 7.23 at 4.97×; masked-latent loss rose during training, but gradient conflict was not measured.The paired-sensor augmentation alone was neutral, and the results compare training configurations rather than isolated losses.Source: N SENSITIVITY TO STRONGER… · p. 29, Figure 10 · p. 29

O.2 CONDITIONAL CONTRIBUTIONS OF THE OBJECTIVES

• Sensor-view alignment improves full and retained transfer by 5.82 and 6.70 AUROC points; on the prespecified 24-cell panel, filtering adds 3.45 ± 1.26 points, concentrated in directions targeting CGMacros.The controlled study used three paired pretraining seeds; its transfer intervals support positive alignment effects under both reported resampling procedures.Source: §O.2 · p. 30, §O.3 · p. 30

O.4 VARIANCE REGULARIZATION AND REPRESENTATION GEOMETRY

• Variance regularization increases effective rank in both projected readouts in every matched comparison and raises total variance across all 12 cohort–seed comparisons.The reported effective rank is computed from normalized singular values; these measurements accompany improved window discrimination and retained transfer.Source: §O.4 · p. 31, Table 30 · p. 31

P THE HPP EXTERNAL COHORT: CONSTRUCTION AND AUDIT; R.5 SCOPE OF CONTRIBUTION ATTRIBUTION

• HPP access is controlled by Pheno.AI, and the released evaluation driver supports reproducibility only for groups with cohort access; the readout analyses do not attribute gains to particular losses.HPP labels include no insulin or C-peptide, so insulin resistance is unavailable; blood-panel labels can also be temporally distant from CGM wear.Source: P THE HPP EXTERNAL COHORT:… · pp. 31–32, §R.5 · p. 39

Q.1 DESIGN AND PREDICTOR SETS

• The evaluation uses frozen released encoders, with GlucoPRISM’s prespecified Trait–State readout; adjusted probes add available covariates but omit medication adjustment and lack covariate-only or persistence baselines.Clinical labels are visit measurements selected within 90 days, so concurrent endpoints are not prospective forecasts.Source: §Q.1 · pp. 33–34, §Q.2 · p. 34

Q.2 CONCURRENT CLINICAL ENDPOINTS

• GlucoPRISM exceeds GlucoFM on correlation and R2 for 14/15 baseline endpoints and on MAE for 15/15, while absolute explained variance remains modest for several endpoints.Sleep correlation is 0.1407 versus 0.1424 for GlucoFM; these are point estimates, not significance tests.Source: §Q.2 · p. 34, Table 33 · p. 35

Q.3 TWO-YEAR FUTURE LEVELS AND CHANGES

• GlucoPRISM leads GlucoFM on 12/15 future-level correlations, 12/15 MAEs, and 13/15 R2 values; change prediction is less consistent, with leads on 9 endpoints for each metric but positive R2 on only 4.Future HbA1c and LDL have negative R2 for both models; the change counts are descriptive, not independent replications.Source: §Q.3 · pp. 35–36, Table 35 · p. 36

Q.4 MEAL RESPONSE AND SHORT-HORIZON GLUCOSE PREDICTION

• GlucoFM leads GlucoPRISM on correlation, R2, and MAE for all seven meal and short-horizon tasks in both predictor sets.The authors note that these tasks depend on within-day structure, which is finer than the scale of Trait and State supervision and partly shared with Sensor.Source: §Q.4 · p. 36, Table 36 · p. 37

R.1 WHAT IS COMPARED; R.2 CLINICAL INFORMATION IS OVERLAPPING AND PARTLY COMPLEMENTARY

• Trait–State has higher correlation than either block alone on 9/15 baseline endpoints, but block utility overlaps and adding a block is not uniformly helpful.LDL correlation is highest for Trait alone (0.2050 versus 0.1703 jointly); the readouts differ in width and lack an equal-width random-coordinate control.Source: §R.2 · p. 37, §R.1 · p. 37, Table 37 · p. 38

R.3 THE SENSOR BAND CARRIES PART OF THE SHORT-HORIZON DYNAMICS

• Adding Sensor improves meal-area correlation from 0.3527 to 0.3669 and peak-rise correlation from 0.3163 to 0.3257, yet all-block GlucoPRISM remains below GlucoFM on correlation for all seven tasks.The readout comparisons keep the encoder fixed and do not isolate the effect of any training loss.Source: Table 38 · p. 38, §R.3 · p. 38, §R.5 · p. 39

R.4 DEVICE IDENTITY AND VISIT-TO-VISIT STABILITY

• HPP device group is decodable from every readout (AUROC 0.83 to 0.86), but its two sensor variants are held mainly by different participants, confounding hardware with population differences.For 44 paired participants, Trait has the highest mean cosine (0.4378) and median cosine (0.9848); stability is not itself a prediction metric.Source: §R.4 · pp. 38–39

S POSTPRANDIAL RESPONSE ON HPP

• Within-participant correlation centers truth and prediction on each participant’s own mean, whereas between-participant correlation compares participant means; pooled correlation is reported for comparability, not as the headline.The label variance is 22.9% between participants and 77.1% within; the stated between-participant share is an upper bound on trait-like variation.Source: S POSTPRANDIAL RESPONSE ON HPP · p. 40
• Stacked on the full classical baseline, the largest within-participant gain from any representation is +0.0186 (95% interval [+0.0138, +0.0235]), estimated by a participant-clustered paired bootstrap over 2000 replicates.Between-participant and within-participant results should be interpreted separately; neither band alone isolates meal-level variation, which is finer than the day-level scale at which State is supervised.Source: S POSTPRANDIAL RESPONSE ON HPP · p. 40
• Pre-meal glucose alone reaches 0.437 correlation, while the raw-window baseline reaches 0.440, making the retained pre-meal value a strong baseline for this endpoint.The label is incremental area over the two hours after the meal; no response sample enters the preceding 24-hour input.Source: S POSTPRANDIAL RESPONSE ON HPP · pp. 39–40
• A 15-minute sensor provides only eight samples in a two-hour response, making incremental area coarse and blunting the peak through sensor smoothing.Meal logs are self-reported, and pre-meal glucose is part of the input and label baseline, so some predictable variance reflects where glucose started.Source: S POSTPRANDIAL RESPONSE ON HPP · p. 40
• GlucoPRISM has lower error than GlucoFM in 13 of 20 Dexcom settings and 14 of 20 Libre settings, while X-CGM-JEPA has the lowest error in 16 of 20 Libre settings.These device-specific results use mean absolute error, a protocol different from the correlation analysis.Source: S POSTPRANDIAL RESPONSE ON HPP · p. 40

T BLOCK DEPENDENCE AND REPRESENTATION STABILITY

• Trait and State remain strongly dependent (|r| = 0.86, HSIC = 0.73), while Sensor dependence is lower (HSIC 0.04 against Trait and 0.07 against State); all three blocks outperform width-matched controls.Lower measured dependence does not establish a pure device factor.Source: T BLOCK DEPENDENCE AND REPRESENTATION… · pp. 40–41
• A linear probe predicts device from the observation mask alone at 100 AUROC, while three glucose-level scalars without mask information reach 99.1 AUROC.Paired recordings also show a systematic offset of −31.1 ± 15.8 mg/dL, with 43 of 44 subjects in the same direction; population device predictability may additionally reflect cohort composition.Source: T BLOCK DEPENDENCE AND REPRESENTATION… · pp. 41–42

U EMBEDDING VISUALISATION

• Figure 11’s label-coloured projections are qualitative illustrations; its label-supervised panel separates classes by construction and is not evidence about either encoder.The other panels are unsupervised projections of the same windows.Source: Figure 11 · p. 42, U EMBEDDING VISUALISATION · p. 42

V MULTI-DAY AGGREGATION

• Multi-day aggregation increases diabetes PR-AUC on CGMacros and Hall and glucotype PR-AUC on Hall, but decreases PR-AUC for every Stanford task and for insulin resistance on CGMacros and Shanghai.These paired changes are relative to each model’s own one-day score; GlucoPRISM remains higher than GlucoFM at every number of days in the subject-level analysis.Source: Figure 12 · p. 43, V MULTI-DAY AGGREGATION · p. 43
• For ShanghaiT2DM insulin resistance, concat(mean, max) raises the GlucoPRISM gain from 1.5 points at K = 2 to 8.4 points at K = 7, with a small loss at K = 4.This differs from mean aggregation for the same cell, so the effect also depends on aggregation choice.Source: Figure 13 · p. 43, V MULTI-DAY AGGREGATION · p. 43

W ADDITIONAL ABLATION STUDIES

• 68.41: GlucoPRISM at three randomly drawn monitoring days exceeds its seven-day score of 68.06 and GlucoFM at five days’ 68.26.These windows need not be consecutive, so the comparisons do not establish substitution for a fixed number of clinical monitoring days. A device-adversarial Sensor head at weight 3 raised its clinical probe score to 72.80 AUROC versus 71.38 for control, motivating the released models’ KL capacity penalty.Source: W ADDITIONAL ABLATION STUDIES · pp. 43–44, Table 43 · p. 44

Y COMPARISON WITH A GMI RULE

• 19.5 points: GlucoPRISM’s Macro-F1 advantage over the seven-day GMI rule was largest on CGMacros-Dexcom; the rule led by 5.8 points on CGMacros-Libre.GlucoPRISM also scored higher on Stanford and CGMacros-Fused, while its standard deviations were large relative to some differences, particularly on Stanford.Source: Table 44 · p. 44, Y COMPARISON WITH A GMI RULE · p. 44

X DESIGN ABLATIONS

• 68.5 ROC-AUC: dense interpolation with sparsification scored above the default 66.9; removing augmentation or keeping only value perturbation scored 68.0 and 68.3.Settings stayed close to the default, which was not the best-scoring setting in any panel; the largest ROC-AUC difference was 1.6 points, so the ablations did not identify a clearly better setting.Source: X DESIGN ABLATIONS · p. 44, Figure 14 · p. 45

Z CORPUS COMPOSITION AND WINDOW GEOMETRY

• 1.79 points: removing 177 BIG IDEAs windows lowers GlucoFM window ROC-AUC to 64.06 against 65.85; removing 9,040 REPLACE-BG windows lowers it 0.23 to 65.62.The same ordering holds for transfer ROC-AUC in both models, but each removal is a single run compared with three-seed full-corpus means, so differences are indicative rather than resolved. The released models use the full corpus, including smaller cohorts.Source: Figure 16 · p. 46, Z CORPUS COMPOSITION AND WINDOW GEOMETRY · p. 46
• 65.85: overlapping consecutive days by 40% and keeping all windows produced this GlucoFM window AUROC and geometry used for every model in the paper.Relative to published geometry at 65.48, two-hour patches cost 2.72 AUROC and 30-minute patches 0.39; count-matched overlap variants were single runs.Source: Figure 16 · p. 46, Z CORPUS COMPOSITION AND WINDOW GEOMETRY · p. 46

EXTENDED RELATED WORK

• Synthetic sensor partners and different-day pairs imperfectly proxy measurement and physiology changes, leaving multiple explanations for persistent Trait–State dependence.The passage identifies supervision, objectives, and the shared encoder as possible explanations.Source: EXTENDED RELATED WORK · p. 47
• Fixed-band filtering removes the measurement component at inference, between training-time invariance and post-hoc erasure requiring deployment-space probing with device labels.The discovery cost is paid once at design time, while removal itself is free.Source: EXTENDED RELATED WORK · p. 47
• The objective-weight sweep tests fixed weights, not adaptive methods such as GradNorm or PCGrad, and does not identify directional gradient conflict.The stronger degradation at higher capacity motivates measuring gradient alignment during training.Source: EXTENDED RELATED WORK · p. 47