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

Adinath Madhavrao Dukre; Abdullah Mohammad; Yifan Lu; Eran Segal; Ahmed Metwally; Imran Razzak

cs.LGOct 8, 2026

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

Authors: Adinath Madhavrao Dukre, Abdullah Mohammad, Yifan Lu, Eran Segal, Ahmed Metwally, Imran Razzak

Organizations: Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE · Vitalverse, Abu Dhabi, UAE · Google, USA

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 p=0.031) and correlation on fourteen of fifteen clinical endpoints in an external cohort of 9,303 participants.

Figures & tables

Figure 1: GlucoPRISM treats a CGM day like light passing through a prism. (a) Schematic. A day of glucose readings from a CGM sensor worn on the back of the upper arm mixes who the person is, what happened that day and how it was measured. Protocol supervision disperses the representation into a Trait block (64 dimensions) that should not change across days or sensors, a State block (48) that may change across days but not across sensors, and a Sensor block (16) that tracks the device. At deployment the Sensor block is filtered out and the remaining 112 dimensions are used with the encoder frozen. (b) Measured. Mean cross-cohort transfer AUROC over twelve directions (Tables 2 and 3 ). Bar colours show which blocks each readout keeps; the horizontal axis starts at 60. The average combines a large gain on the four directions into CGMacros with a neutral effect elsewhere (Section 4.1 ).

Figure 2: GlucoPRISM pretraining and fixed readout. (a) Synthetic sensor views and repeated-day views provide distinct supervision. (b) The shared GlucoFM backbone produces a 128-dimensional representation partitioned into Trait (64), State (48) and Sensor (16); auxiliary objectives and a Sensor bottleneck shape these blocks during training. (c) At deployment the fixed filter $P_{TS}=[I_{112}\;0]$ removes the Sensor band and passes the 112 Trait and State dimensions to the downstream probe. The drawing summarizes the flow; statistical pooling (256 to 128), block projections and raw-block CMP heads are specified in the method. The Sensor mean and projection remain active during feature extraction. Traces are schematic.

	Trait (64)	State (48)	Sensor (16)
Across days	unchanged	may change	not constrained
Across sensors	unchanged	unchanged	changes
Supervision	day-pair alignment, sensor-view alignment, level targets	day contrast, sensor-view alignment, variability targets	device prediction, KL bottleneck

Table 1: Protocol supervision as response contracts. Each block is defined by how it should respond when only the recording day or only the sensor changes. The supervision terms are given in Equations 2 to 4 .

		Frozen third-party checkpoints							Retrained, public corpus			Ours
		Chronos-2 small	Chronos -2	MOMENT large	MOMENT small	CGM- former	Mantis	Mantis V2	CGM- JEPA	X-CGM- JEPA	Gluco FM	Gluco PRISM
Window	PR	48.7	49.9	50.3	52.1	53.7	54.0	55.3	52.3	52.1	57.9	59.1
	ROC	58.5	59.6	59.9	62.5	62.9	63.6	64.7	60.9	60.8	65.9	68.0
	F1	52.8	53.5	53.5	55.2	55.8	56.2	56.4	53.8	53.5	57.6	57.8
Subject	PR	64.3	64.1	67.7	66.2	64.9	68.1	67.8	66.9	67.2	66.3	67.2
	ROC	66.1	65.7	69.3	67.3	66.3	70.0	69.2	67.3	67.8	69.0	69.4

Table 2: Frozen-encoder results averaged over fourteen within-cohort cells and twelve transfer directions. Window and Subject rows are within-cohort evaluations. GlucoPRISM uses the filtered readout

z_{TS}

. Trained models use three shared seeds. PR, ROC and F1 are reported in percentage points;

p_{\mathrm{Holm}}

refers to the window-AUROC comparison with GlucoFM in the declared family. Subject-level and transfer tests are given separately in the appendix.

Readout	Transfer ROC-AUC	$\Delta$	ECE	Brier
Full readout (all three bands)	68.75 $\pm$ 2.67	—	0.163 $\pm$ 0.022	0.241 $\pm$ 0.014
INLP ( Ravfogel et al., 2020 ) , 8 iterations	72.94 $\pm$ 0.40	$+4.19$ $\pm$ 2.44	0.136 $\pm$ 0.024	0.218 $\pm$ 0.009
Filtered readout (ours)	73.57 $\pm$ 0.54	$+4.82$ $\pm$ 2.18	0.114 $\pm$ 0.009	0.207 $\pm$ 0.006

Table 3: Readout interventions on GlucoPRISM: transfer AUROC, ECE and Brier score over three seeds. Differences are relative to the same encoder’s full readout. INLP fits eight projections using real device labels and accesses target inputs when CGMacros is the target; the fixed Sensor filter does neither. The paired filter-minus-INLP AUROC difference is

+0.63

points (95% interval

[-0.09,+1.35]

p=0.079

Direction	Task	CGM-JEPA	X-CGM-JEPA	GluFormer (tiny)	GlucoFM	GlucoPRISM
Stanford $\to$ Hall	Diabetes	70.4	71.5	72.0	68.4	76.6 (+4.6)
	IR	63.9	64.0	60.7	61.7	62.1 ( $-$ 1.9)
Hall $\to$ Stanford	Diabetes	63.9	64.0	64.7	66.3	76.0 (+9.7)
	IR	56.6	56.7	56.2	58.2	58.0 ( $-$ 0.2)
CGMacros $\to$ Hall	Diabetes	69.6	70.7	72.1	78.2	78.0 ( $-$ 0.2)
	IR	61.1	61.6	63.2	68.5	70.0 (+1.5)

Table 4: Cross-dataset transfer following Li et al. (2026) : a frozen-encoder probe is fitted on the source cohort and evaluated on the target cohort for diabetes and insulin resistance (IR). GlucoPRISM uses the filtered Trait–State readout

z_{TS}

. Values are means over three pretraining seeds in percentage points. Bold marks the best and underline the second-best model in each row; the bracket gives GlucoPRISM’s difference from the best baseline. ROC-AUC is shown; the full table with PR-AUC and seed standard deviations is Table 18 in Appendix H . † marks directions into CGMacros, the target cohort recorded with two sensor types.

Figure 3: Frozen-encoder prediction of all 15 baseline clinical endpoints on HPP. Paired bars show GlucoFM (blue, diagonal hatching) and GlucoPRISM (orange, cross-hatching) for each endpoint; the two panels show correlation and explained variance. Scores use embedding-only Ridge probes on the common test pool, with 301–1,214 labelled participants per endpoint. Labels are concurrent clinical measurements, not prospective disease outcomes. Bars are reported estimates without uncertainty intervals; the panels use different horizontal scales.

	No learned representation				CGM foundation models					GlucoPRISM readouts
	Mean glucose	Raw window	Indices full wear	Indices day-matched	CGM- JEPA	X-CGM- JEPA	Glu- Former	GlucoFM	GlucoFM trunc-112	Filtered $[z_{T}\\|z_{S}]$	$z_{A}$ kept	$z_{T}$ only	$z_{S}$ only	$z_{A}$ only
DR	65.4	63.6	70.3	70.5	67.3	67.5	68.6	69.5	69.5	71.8	71.7	71.5	71.4	68.8
HL	53.1	53.0	59.7	58.2	55.8	56.7	58.0	59.8	59.9	59.9	59.7	59.7	59.5	58.1
OB	57.0	65.1	69.7	68.8	66.5	66.7	68.1	68.7	68.7	71.2	71.3	69.4	69.3	65.8
Avg	58.5	60.6	66.6	65.8	63.2	63.6	64.9	66.0	66.0	67.6	67.6	66.9	66.7	64.2

Table 5: External frozen-encoder evaluation on HPP: subject-level AUROC from ten repetitions of five-fold subject-grouped probing. Clinical labels train the probes within HPP; encoders remain frozen and were not pretrained on HPP. DR: laboratory HbA1c

\geq 5.7\%

; HL: hyperlipidemia; OB: obesity. Kept/deleted rows use the same encoder. Classical indices are shown with full wear and with input days matched to the model. These comparisons are descriptive and separate from the confirmatory family in Table 2 .

Figure 4: Do the bands keep their contracts? (a) Kernel dependence (HSIC) between bands: Sensor is nearly independent of Trait and State, while Trait and State remain dependent. (b) Within-person cosine between two HPP clinic visits (44 participants; bars are medians, diamonds are means). (c) Device-group AUROC on HPP by readout. Device group varies between participants wearing one manufacturer’s sensor, and every band decodes it, so this contract is not established in that setting. (d) AUROC gain from keeping each term in a controlled one-term-off study with its own protocol (24 transfer cells, three seeds); lines are 95% participant-bootstrap intervals.

Appendix figures & tables51 assets

Supplementary material from the paper’s appendix.

Appendix

Evaluation	Encoder pretraining	Probe fitting and scoring	Question tested
Public windows / subjects	Public corpus	Subject-disjoint folds within each cohort	Representation quality
Cross-cohort transfer	Public corpus	Fit on source; score on target	Filtering the Sensor band
External HPP	Public corpus	HPP training/test participants	Clinical information
HPP corpus-size study	Nested HPP subsets	Public benchmark splits and transfer directions	Transfer at larger pretraining sizes

Appendix

Table 6: Evaluation settings. Encoders are frozen for all downstream probes. Public refers to the five-cohort pretraining corpus; the public benchmark contains four evaluation cohorts. The two HPP uses have opposite train–evaluation directions.

		CGMacros				ShanghaiT2DM			Stanford			Hall
Method	Metric	DR3	HL	IR	Ob	HL	Hy	IR	$\beta$ C	DR	IR	DR	GT	HL	IR	Avg
General time-series foundation models (frozen third-party checkpoints)
Chronos-2-small	PR	55	28	84	61	32	16	61	60	66	58	51	49	20	41	48.7
	ROC	70.0	46.8	68.9	58.7	49.5	53.3	52.2	63.2	64.8	67.3	64.5	59.0	50.4	50.1	58.5
	F1	48	47	61	54	48	50	50	58	60	57	57	53	47	49	52.8
Chronos-2	PR	56	29	90	64	31	18	61	59	65	55	49	51	23	47	49.9

Appendix

Table 7: Linear-probe performance on every public task–cohort cell (window level, frozen encoders, subject-disjoint folds, three shared seeds for trained models). Bold marks the best and underline the second-best value in each column for each metric. ROC-AUC cells are given to one decimal; PR-AUC and Macro-F1 cells are rounded to integers, and Avg is the unrounded 14-cell mean from Table 2 . Tasks: DR diabetes risk, DR3 3-class diabetes risk, IR insulin resistance,

\beta

\beta

-cell dysfunction, HL hyperlipidemia, Ob obesity, Hy hypoglycemia, GT glucotype.

Figure 5: Filtering the Sensor band. (a) Transfer AUROC of the full readout (open circles) and the filtered readout

z_{TS}

(filled circles) for each direction, from the rounded means of Table 19 ; the column on the right gives the difference, and shading marks CGMacros targets. DR: diabetes risk; IR: insulin resistance. (b) Absolute AUROC in the separate two-seed KL sweep (mean

\pm

seed SD). Panel (b) measures window AUROC, distinct from the transfer metric in panel (a). Larger gaps need not imply higher filtered-readout accuracy.

	8 directions not into CGMacros	4 directions into CGMacros	All 12 directions
Full readout	70.76	64.71	68.75
Filtered $z_{TS}$	70.94	78.83	73.57
GlucoFM	67.27	73.45	69.33
Filtered $-$ full	$+0.17$ $[-0.72,+1.01]$	$+14.12$ $[+5.03,+21.34]$	$+4.82$ $[+1.02,+10.04]$
Full $-$ GlucoFM	$+3.49$ $[+0.93,+6.16]$	$-8.74$ $[-21.51,+4.24]$	$-0.59$ $[-7.12,+4.49]$
Filtered $-$ GlucoFM	$+3.66$ $[+0.73,+6.59]$	$+5.38$ $[-3.09,+12.42]$	$+4.24$ $[+0.86,+7.39]$

Appendix

Table 8: Cross-cohort transfer ROC-AUC by target group, computed from per-seed, per-direction values over the three shared seeds. Differences are averaged over directions; brackets are 95% intervals from resampling directions and seeds together (

n=8

n=4

and

n=12

directions). Directions share cohorts, so Table 9 also holds out one target cohort at a time.

Held-out target cohort	Filtered $-$ full	Filtered $-$ GlucoFM
CGMacros	$+0.17$ $[-0.74,+1.01]$	$+3.66$ $[+0.80,+6.55]$
Stanford	$+7.09$ $[+1.44,+14.16]$	$+3.94$ $[-0.56,+8.31]$
Hall	$+7.20$ $[+1.74,+14.42]$	$+5.11$ $[+0.56,+9.35]$

Appendix

Table 9: Leave-one-target-cohort-out transfer: differences averaged over the directions whose target is not the held-out cohort, with 95% intervals from resampling directions and seeds together. Filtered minus GlucoFM stays positive whichever target cohort is held out; its interval spans zero only when Stanford is held out.

Context	Endpoint	GluFormer (tiny)	CGM-JEPA	X-CGM-JEPA	GlucoFM	GlucoPRISM
Frozen representation	Trajectory (mg/dL)	25.86	23.73	23.43	25.84	25.87
	Positive iAUC (mg/dL $\cdot$ h)	36.05	32.59	32.04	35.80	36.20
	Peak rise (mg/dL)	32.59	29.74	29.31	32.62	33.32
	Peak time (min)	30.89	30.51	30.47	31.27	30.36
+ 1 h pre-meal CGM	Trajectory (mg/dL)	24.96	23.27	23.08	23.46	23.31
	Positive iAUC (mg/dL $\cdot$ h)	34.69	32.16	31.86	32.20	32.08

Appendix

Table 10: Postprandial glycemic response (PPGR) across cumulative context, averaged over Dexcom (5-min) and Libre (15-min) evaluations, following the protocol of Li et al. (2026) . Entries are mean absolute error (lower is better) over 10 probe seeds; bold marks the best and underline the second-best model in each row, ranked on unrounded values. All baselines are our reproductions on the public-only pretraining corpus, so they differ from published values. Device-specific results with seed standard deviations are in Appendix S .

Figure 6: Few-shot adaptation with frozen encoders, reported as task-averaged PR-AUC over the public within-cohort cells, following the protocol of Li et al. (2026) . Left: number of labelled support subjects per class. Right: fraction of each subject’s observations available for probe fitting. Points are mean estimates without uncertainty intervals; vertical axes differ between panels.

Figure 7: Transfer advantage across HPP pretraining corpus sizes. Each row is one metric and each column one nested corpus size (500 to 8,000 participants); cells give GlucoPRISM minus GlucoFM in percentage points, with orange where GlucoPRISM is higher and grey-blue where GlucoFM is higher. Window and subject scores average 14 task–cohort cells; transfer scores average 12 directions. One pretraining seed is available, so no uncertainty or significance is inferred. Absolute scores are shown in Figure 15 .

Model	Params	Reproduced			Published		mean	cells
	(M)	PR	ROC	F1	PR	ROC	$\|\Delta$ ROC $\|$	$\leq 2$
GlucoFM ( Li et al., 2026 )	0.72	57.9 $\pm$ 0.6	65.9 $\pm$ 1.1	57.6 $\pm$ 0.6	58.8	66.7	3.53	5 / 14
GluFormer-tiny ( Lutsker et al., 2026 )	0.65	52.3 $\pm$ 0.3	61.3 $\pm$ 0.5	55.1 $\pm$ 0.2	53.0	61.9	3.43	6 / 14
CGM-JEPA ( Muhammad et al., 2026 )	0.52	52.3 $\pm$ 0.1	60.9 $\pm$ 0.2	53.8 $\pm$ 0.3	54.7	62.4	3.47	5 / 14
X-CGM-JEPA ( Muhammad et al., 2026 )	0.52	52.1 $\pm$ 0.1	60.8 $\pm$ 0.1	53.5 $\pm$ 0.1	54.6	62.1	3.23	7 / 14
raw 288-point window	—	51.2	59.9	53.6	—	—	—	—

Appendix

Table 11: Retrained CGM baselines and published reference results. Task-averaged over the same 14 task–cohort cells, the same frozen folds and the same untuned probe as everywhere else in this paper. Reproduced is what this work produces; published is GlucoFM’s Table 3, whose pretraining corpus combines public data with non-public Wear-CGM recordings. It reports 477 subjects and 109,066 recording hours, compared with 514 subjects and 10,952 daily windows covering 237,879 distinct hours here. The published scores are therefore a reference under a different corpus and preprocessing pipeline, rather than a like-for-like column. The last two columns give per-cell agreement: mean absolute ROC-AUC difference over the 14 cells, and how many of them land within 2 points. GlucoFM’s lead over the next-best model reproduces at

+4.5

AUROC against

+4.4

published , which is the ordering this benchmark rests on. The three models beneath it span

0.5

AUROC here against

0.4

published, so neither table resolves their internal order.

Cohort	Device	Rate (min)	Subjects	Windows	Obs. frac.
REPLACE-BG ( Aleppo et al., 2017 )	Dexcom G4	5	226	9,040	0.978
Stanford ( Hall et al., 2018 )	Dexcom	5	27	622	0.983
ShanghaiT2DM ( Zhao et al., 2023 )	FreeStyle Libre	15	40	722	0.333
Colás ( Brunner et al., 2012 )	Medtronic iPro	5	205	391	0.997
BIG IDEAs ( Bent et al., 2021 )	Dexcom	5	16	177	0.935
Total (public only)	—	—	514	10,952	—

Appendix

Table 12: Pretraining corpus. All five cohorts are public and redistributable. ShanghaiT2DM’s low observation fraction is not missingness: a 15-minute sensor can fill at most 96 of 288 five-minute grid slots, so coverage thresholds are applied relative to device rate. REPLACE-BG is capped at 40 windows per subject; uncapped it would contribute roughly 28,000 windows and turn the corpus into a type 1 diabetes distribution.

Setting	Value	Provenance
Architecture
Window grid	288 steps at 5 min (24 h)	inherited
Patches	24 of 12 steps (1 h)	inherited
Causal filter $\sigma$	init 6.0, learned in $[2.0,12.0]$	inherited
Filter truncation / lookback	3 $\sigma$ / 9 steps	inherited
Stream dim	64 per stream	inherited

Appendix

Table 13: GlucoPRISM configuration, generator settings and probe parameters. HPP scaling uses 96 positions and patch width four instead of the public-corpus grid.

Figure 8: Real Dexcom–Libre disagreement on same-day paired windows. The offset is systematic and one-directional rather than symmetric noise, which is what makes it a learnable device signature rather than a nuisance that averages out.

Fitted parameter	Point estimate	Leave-one-subject-out range	95% bootstrap
Calibration offset (mg/dL)	-31.12	[-32.39, -30.20]	[-34.60, -27.71]
Deming slope	0.88	[0.86, 0.89]	[0.83, 0.92]
Mean abs. relative difference (%)	24.25	[23.86, 24.56]	[22.54, 25.99]
Correlation with source window	0.74	[0.73, 0.75]	[0.69, 0.78]

Appendix

Table 14: Subject-level sensitivity of the paired-sensor generator. Parameters are fitted on 374 windows from 44 subjects and refitted using leave-one-subject-out analysis and subject bootstrap resampling. The calibration offset remains negative in all 2,000 bootstrap refits; the reported ranges quantify variation in its magnitude and the other fitted parameters.

Arm	Lag	Smoothing $\sigma$	Fitted on
Control (released recipe)	$-1\pm 1.0$	1.5	all 44 CGMacros paired participants
(a) Uncalibrated	0	0	nothing (only the Libre 15-minute decimation)
(b) Half-calibrated	$-1\pm 1.48$	1.29	22 of the 44 paired participants

Appendix

Table 15: Synthetic-sensor generator arms used to test calibration leakage. Lag is in five-minute steps; all arms apply no level shift (bias 0).

Projected readout (deployed)
Arm	Filtered ROC, all 12	Gain, into CGMacros (all 44)	Gain, into CGMacros (held-out 22)	Gain, 8 not into CGMacros	Arm $-$ control (held-out 22)
Control	73.14 $\pm$ 0.62	14.39 $\pm$ 2.44	16.80 $\pm$ 4.05	0.03 $\pm$ 0.77	—
(a) Uncalibrated	73.05 $\pm$ 0.53	12.42 $\pm$ 5.42	13.35 $\pm$ 6.83	0.45 $\pm$ 0.89	$-3.46$ [ $-0.05$ , $-5.54$ , $-4.78$ ]
(b) Half-calibrated	73.17 $\pm$ 0.58	15.14 $\pm$ 5.30	17.16 $\pm$ 7.13	0.03 $\pm$ 0.91	$+0.35$ [ $+4.73$ , $-0.00$ , $-3.67$ ]
Raw pre-head blocks
Control	67.69 $\pm$ 3.12	$-0.29$ $\pm$ 0.88	$-0.70$ $\pm$ 0.97	0.25 $\pm$ 0.21	—

Appendix

Table 16: Filtering gain (filtered minus full transfer ROC-AUC) under each generator, mean

\pm

SD over three seeds. Held-out 22 restricts CGMacros-target scoring to the participants not used by arm (b). The last column gives each arm minus the control on the held-out 22, averaged over seeds, with per-seed values in brackets.

	window ROC-AUC			window PR-AUC			subject ROC-AUC
Model	$\Delta$	won	$p_{\mathrm{Holm}}$	$\Delta$	won	$p_{\mathrm{Holm}}$	$\Delta$	won	$p_{\mathrm{Holm}}$
GlucoPRISM	+2.16	13/14	0.031 ∗	+1.16	9/14	0.208	+0.50	7/14	1.000
Chr-2s	-7.38	4/14	0.107	-9.21	3/14	0.019 ∗	-2.85	4/14	1.000
Chr-2	-6.24	3/14	0.221	-8.08	3/14	0.047 ∗	-3.21	5/14	1.000
MOM-l	-5.92	3/14	0.282	-7.66	3/14	0.019 ∗	+0.31	8/14	1.000
CGM-JEPA	-4.94	3/14	0.293	-5.64	3/14	0.122	—	—	—

Appendix

Table 17: Paired significance against GlucoFM, on all three metrics at once. Wilcoxon signed-rank over the 14 task–cohort cells, Holm-corrected over the confirmatory family

k=10

of declared before any result was read.

\Delta

is the mean paired difference and “won” the number of cells in which the model beats GlucoFM. On the confirmatory axis, window ROC-AUC, exactly one comparison survives correction, and it is ours. The three further starred entries all sit in the PR column and are baselines falling below GlucoFM, not improvements over it. Reading across a row rather than across three separate tables is the point: the ordering is stable on window ROC and PR, and dissolves at subject level.

Direction	Task	Metric	CGM-JEPA	X-CGM-JEPA	GluFormer (tiny)	GlucoFM	GlucoPRISM
Stanford $\to$ Hall	Diabetes	PR	53.8 $\pm$ 0.7	55.2 $\pm$ 0.5	63.2 $\pm$ 5.1	59.5 $\pm$ 1.6	68.0 $\pm$ 2.0
		ROC	70.4 $\pm$ 0.9	71.5 $\pm$ 0.4	72.0 $\pm$ 3.2	68.4 $\pm$ 0.3	76.6 $\pm$ 0.8
	IR	PR	48.7 $\pm$ 2.0	48.5 $\pm$ 0.7	48.7 $\pm$ 2.5	53.1 $\pm$ 1.9	50.7 $\pm$ 2.6
		ROC	63.9 $\pm$ 1.1	64.0 $\pm$ 0.6	60.7 $\pm$ 2.9	61.7 $\pm$ 1.4	62.1 $\pm$ 0.5
Hall $\to$ Stanford	Diabetes	PR	66.3 $\pm$ 0.1	66.3 $\pm$ 0.4	68.4 $\pm$ 1.1	69.9 $\pm$ 1.2	78.5 $\pm$ 1.1
		ROC	63.9 $\pm$ 0.2	64.0 $\pm$ 0.3	64.7 $\pm$ 1.1	66.3 $\pm$ 1.1	76.0 $\pm$ 1.1

Appendix

Table 18: Cross-dataset transfer with seed variability: mean

\pm

SD over three pretraining seeds, in percentage points, for all directions in Table 4 . Bold : best; underline : second best.

Backbone and frozen third-party baselines
	CGM $\to$ Hall		CGM $\to$ Stan		Hall $\to$ CGM		Hall $\to$ Stan		Stan $\to$ CGM		Stan $\to$ Hall
Model	DR	IR	DR	IR	DR	IR	DR	IR	DR	IR	DR	IR	Avg
GlucoPRISM	78.0	70.0	80.9	65.8	81.7	76.6	76.0	58.0	85.6	71.4	76.6	62.1	73.6
$z_{A}$ kept	78.0	70.7	80.8	63.9	67.6	57.8	75.7	59.2	74.3	59.2	75.9	61.9	68.7
GlucoFM	78.2	68.5	74.3	62.6	77.2	80.3	66.3	58.2	71.1	65.3	68.4	61.7	69.3
MantisV2	68.3	64.0	69.5	71.0	77.6	64.8	65.8	65.0	68.3	72.5	66.1	68.8	68.5

Appendix

Table 19: Cross-cohort transfer: all twelve (source, target, task) directions. The probe is fitted on the entire source cohort and evaluated on the entire target cohort; target clinical labels are used only for scoring, not fitting or tuning and there is no cross-validation, so this measures direct transfer rather than within-cohort generalisation. ROC-AUC, higher is better; bold is the best in a column, underline the runner-up. The second row is the same model read without discarding the Sensor block, so the first row minus the second isolates the intervention this paper proposes. Cohorts abbreviated CGM (CGMacros), Stan (Stanford), Hall; tasks DR (diabetes risk) and IR (insulin resistance).

Endpoint	GlucoFM	GlucoPRISM	$\Delta$ vs FM	GlucoPRISM [ $z_{A}$ kept]	deletion gain
Dexcom $\rightarrow$ Libre
Diabetes risk (3-cls)	81.85	80.21	$-1.63$	61.09	$+19.12$
Hyperlipidemia	63.91	54.07	$-9.84$	53.00	$+1.07$
Insulin resistance	83.27	85.25	$+1.97$	54.09	$+31.16$
Obesity	66.78	64.66	$-2.12$	60.41	$+4.25$
Libre $\rightarrow$ Dexcom

Appendix

Table 20: Paired cross-device evaluation over all 8 task–direction combinations. GlucoPRISM has a mean difference of -1.34 AUROC points versus GlucoFM and wins 3 comparisons. The same encoder’s full readout is weaker for Dexcom

\rightarrow

Libre (mean 57.1) than in the reverse direction (74.1). This asymmetry motivates further analysis of sampling, calibration and participant effects; it does not isolate a unique mechanism.

Figure 9: Individual block readouts and width-matched controls. Cross-cohort transfer ROC-AUC over three seeds. Trait and State each outperform their controls. Trait alone and the released

[z_{T}\|z_{S}]

readout differ by less than the seed standard deviation. Applying the same truncations to the GlucoFM embedding yields no corresponding gain. These values come from separate block-control runs on different checkpoints, not from the released model’s runs in Table 3 , so the two are not expected to match.

	Cross-cohort transfer			Within-cohort
Readout	ROC-AUC	ECE	Brier	window	subject
keep all 16 (full readout)	68.75 $\pm$ 2.67	0.163 $\pm$ 0.022	0.241 $\pm$ 0.014	67.90 $\pm$ 0.66	69.55 $\pm$ 2.48
keep 12 of 16	68.76 $\pm$ 2.45	0.163 $\pm$ 0.026	0.242 $\pm$ 0.017	—	—
keep 8 of 16	68.44 $\pm$ 2.09	0.167 $\pm$ 0.027	0.244 $\pm$ 0.019	—	—
keep 4 of 16	68.49 $\pm$ 2.12	0.169 $\pm$ 0.025	0.245 $\pm$ 0.018	67.75 $\pm$ 0.55	69.15 $\pm$ 3.14
keep 2 of 16	71.63 $\pm$ 2.38	0.137 $\pm$ 0.013	0.218 $\pm$ 0.006	67.97 $\pm$ 0.62	68.71 $\pm$ 3.25

Appendix

Table 21: Partial Sensor deletion. Coordinates are ranked by device predictability using real Dexcom/Libre labels, and the least informative

m

are restored. Removing all 16 coordinates gives the highest transfer and window AUROC among these readouts; the full readout has the highest subject AUROC. Restoring four coordinates reduces transfer AUROC from 73.57 to 68.49. ECE and Brier are lower-is-better.

Projected readout
Method	ROC-AUC	PR-AUC	ECE	Brier
GlucoPRISM, filtered (Sensor removed)	73.14 $\pm$ 0.62	72.33 $\pm$ 1.32	0.128 $\pm$ 0.013	0.214 $\pm$ 0.008
GlucoPRISM, full readout	68.35 $\pm$ 0.55	69.40 $\pm$ 1.62	0.167 $\pm$ 0.030	0.243 $\pm$ 0.022
GlucoPRISM, full + INLP (8 it.)	73.46 $\pm$ 0.88	72.46 $\pm$ 1.22	0.123 $\pm$ 0.007	0.213 $\pm$ 0.004
GlucoPRISM, full + LEACE	73.09 $\pm$ 1.01	72.35 $\pm$ 1.24	0.136 $\pm$ 0.010	0.219 $\pm$ 0.010
GlucoFM	67.26 $\pm$ 2.53	68.27 $\pm$ 1.32	0.237 $\pm$ 0.026	0.283 $\pm$ 0.023

Appendix

Table 22: Fitted erasure versus the fixed Sensor filter: transfer scores averaged over twelve directions, mean

\pm

SD over three pretraining seeds. Separate same-corpus retrains from Table 3 . The upper panel uses the projected readout deployed in the paper; the lower panel uses the raw pre-head blocks. GlucoFM rows are identical in both panels because GlucoFM has no block heads.

Method	Directions	$\Delta$ ROC-AUC	$\Delta$ PR-AUC	$\Delta$ ECE	$\Delta$ Brier	$\Delta$ ROC-AUC (raw)
GlucoPRISM, full readout	All 12	$-4.78$ $[-9.21,-1.04]$	$-2.93$ $[-6.24,-0.28]$	$+0.039$ $[-0.004,+0.101]$	$+0.030$ $[-0.000,+0.079]$	$-0.07$ $[-0.63,+0.69]$
	8 not into CGMacros	$-0.03$ $[-0.98,+1.14]$	$+0.24$ $[-0.78,+1.55]$	$+0.008$ $[-0.010,+0.035]$	$+0.002$ $[-0.006,+0.013]$	$-0.25$ $[-0.70,+0.25]$
	4 into CGMacros	$-14.29$ $[-19.43,-9.77]$	$-9.27$ $[-14.55,-4.60]$	$+0.101$ $[-0.002,+0.225]$	$+0.084$ $[+0.009,+0.182]$	$+0.29$ $[-1.19,+1.95]$
GlucoPRISM, full + INLP	All 12	$+0.33$ $[-0.83,+1.76]$	$+0.13$ $[-1.03,+1.44]$	$-0.006$ $[-0.039,+0.030]$	$-0.001$ $[-0.017,+0.016]$	$+0.61$ $[-0.98,+2.79]$
	8 not into CGMacros	$-0.32$ $[-1.06,+0.48]$	$-0.19$ $[-1.57,+1.17]$	$+0.004$ $[-0.013,+0.029]$	$+0.001$ $[-0.005,+0.011]$	$-0.38$ $[-0.94,+0.17]$
	4 into CGMacros	$+1.62$ $[-1.34,+4.53]$	$+0.76$ $[-1.60,+3.16]$	$-0.025$ $[-0.103,+0.062]$	$-0.006$ $[-0.046,+0.041]$	$+2.59$ $[-2.10,+7.64]$

Appendix

Table 23: Each method minus the fixed filter (GlucoPRISM, filtered) on the projected readout, by direction group. Paired by seed and direction; brackets are 95% intervals from a joint seed-and-direction bootstrap. Negative

\Delta

ROC and

\Delta

PR, and positive

\Delta

ECE and

\Delta

Brier, favour the fixed filter. The last column repeats

\Delta

ROC for the raw pre-head readout.

$\lambda$	Held-out MCR+TD loss	Held-out device AUROC	Within loss budget
0.1	0.0300	99.91	yes (selected)
0.3	0.2372	98.99	no
1	0.2061	99.53	no
3	0.3003	97.67	no

Appendix

Table 24: Selection of the adversarial weight on the pretraining held-out split, with no downstream data. The rule keeps weights whose held-out MCR+TD loss is within 10% of that at

\lambda=0.1

and picks the lowest held-out device AUROC among them.

Model	Window PR	Window ROC	Window F1	Subject PR	Subject ROC	Subject F1	Transfer ROC
GlucoPRISM, filtered	59.06 $\pm$ 0.49	67.90 $\pm$ 0.36	57.87 $\pm$ 0.11	67.07 $\pm$ 0.29	69.26 $\pm$ 1.15	58.78 $\pm$ 0.28	73.14 $\pm$ 0.62
GlucoPRISM, full	58.87 $\pm$ 0.45	67.88 $\pm$ 0.41	57.81 $\pm$ 0.36	67.07 $\pm$ 2.20	70.06 $\pm$ 2.61	60.09 $\pm$ 1.69	68.33 $\pm$ 0.53
GlucoFM	56.04 $\pm$ 0.30	62.99 $\pm$ 0.22	56.47 $\pm$ 0.26	66.17 $\pm$ 0.87	68.98 $\pm$ 0.75	60.18 $\pm$ 0.36	67.25 $\pm$ 2.54
DANN-GlucoFM, $\lambda=0.1$ (selected)	54.30 $\pm$ 0.56	61.85 $\pm$ 0.74	55.68 $\pm$ 0.17	67.26 $\pm$ 3.02	69.48 $\pm$ 4.42	60.50 $\pm$ 2.56	67.19 $\pm$ 1.77
DANN-GlucoFM, $\lambda=0.3$	54.97 $\pm$ 1.25	63.71 $\pm$ 0.69	56.58 $\pm$ 0.81	64.73 $\pm$ 2.95	67.00 $\pm$ 3.82	58.19 $\pm$ 2.77	61.73 $\pm$ 3.21
DANN-GlucoFM, $\lambda=1$	53.18 $\pm$ 1.60	61.43 $\pm$ 1.65	55.03 $\pm$ 1.04	63.84 $\pm$ 3.23	65.68 $\pm$ 3.68	56.56 $\pm$ 3.26	64.02 $\pm$ 3.93

Appendix

Table 25: Domain-adversarial GlucoFM (DANN) against GlucoFM and GlucoPRISM, mean

\pm

SD over three seeds, from the separate same-corpus retrains. Upper block: within-cohort 14-cell averages and 12-direction transfer ROC-AUC. Lower block: transfer PR-AUC and calibration, and linear device information on same-day CGMacros pairs. Bold : best; underline : second best (device AUROC is not ranked).

Sensor-block width $d_{A}$ (at $\beta=0.1$ )
Setting	Full readout	$z_{A}$ dropped	Gain (window)	Gain (subject)
8	67.67 $\pm$ 0.43	68.01 $\pm$ 0.06	$+0.35$ $\pm$ 0.37	$-1.11$ $\pm$ 1.89
16 (released)	67.52 $\pm$ 0.17	68.03 $\pm$ 0.21	$\mathbf{+0.51}$ $\pm$ 0.04	$\mathbf{+0.56}$ $\pm$ 1.02
32	67.62 $\pm$ 0.59	67.91 $\pm$ 0.29	$+0.30$ $\pm$ 0.30	$-1.49$ $\pm$ 0.00
KL weight $\beta$ (at $d_{A}=16$ )
0.03	67.56 $\pm$ 1.17	68.16 $\pm$ 0.74	$+0.60$ $\pm$ 0.43	$-0.03$ $\pm$ 3.21

Appendix

Table 26: Sensor width and KL-weight sweep with two matched seeds per arm. Width changes hold Trait at 64 dimensions and allocate the remaining width between State and Sensor. Larger deletion gains occur at the two higher KL weights, but the sequence is nonmonotonic and uncertain.

Encoder	Paired subj.	Frozen encoder	$z_{T}$	$z_{T}\\|z_{S}$	$\Delta$
$1\times$ , no factorization loss	15	65.5	64.7	64.6	$-0.80$
	20	65.5	66.0	65.7	$+0.50$
	30	65.5	66.8	65.7	$+1.30$
	45	65.5	65.6	65.2	$+0.10$
$1\times$ , post-hoc heads	15	63.5	64.0	63.4	$+0.50$
	20	63.5	64.1	64.4	$+0.60$

Appendix

Table 27: Fitting the block decomposition post hoc on a frozen encoder, using real paired windows rather than synthetic ones, while sweeping how many paired subjects are available for the fit. The pre-specified comparison is the Trait block against the frozen encoder’s own unsplit output: it is

-0.08

ROC-AUC over 12 configurations (

t=-0.32

), positive in 7 of 12. The tested post-hoc procedure yields no resolved average improvement.

Figure 10: Training losses with and without strong protocol objectives. The masked-latent objective increases during joint training while auxiliary losses decrease. This indicates competing optimization outcomes, without directly measuring gradient conflict.

Objective	Endpoint	Without	Complete	Gain	Interval
Sensor-view alignment	Full transfer	62.52	68.34	+5.82	[3.60, 7.82]
Sensor-view alignment	Retained transfer	65.09	71.79	+6.70	[4.14, 9.10]
Sensor-objective bundle	Retained transfer	67.00	71.79	+4.79	[2.26, 7.24]
Variance regularization	Full window	66.89	68.23	+1.35	[0.44, 2.52]
Variance regularization	Retained transfer	68.35	71.79	+3.44	[1.13, 5.86]
Sensor KL regularization	Full window	67.61	68.23	+0.62	[0.05, 1.14]

Appendix

Table 28: Objective contributions whose intervals exclude zero in the controlled study. All entries are AUROC percentage points. The complete configuration is compared with the indicated objective removed, with the other coefficients fixed. Intervals are conditional 95% Bayesian participant-bootstrap intervals.

Metric	Retained minus full	Favorable direction
AUROC	+0.0345	Higher
Average precision	+0.0221	Higher
Macro-F1	+0.0806	Higher
Brier score	$-0.0717$	Lower
Expected calibration error, ten bins	$-0.0968$	Lower

Appendix

Table 29: Favorable mean changes after removing the 16 Sensor coordinates in the controlled study’s transfer panel. Changes use each metric’s original scale. The accompanying intervals in the text apply to AUROC; the other entries are descriptive mean changes.

Readout	Without variance	Complete	Paired consistency
Full projected	4.708	6.519	12/12
Retained projected	3.310	5.210	12/12

Appendix

Table 30: Mean singular-value entropy effective rank with and without variance regularization. Paired consistency counts the cohort–seed records with higher effective rank when the objective is active.

Representation	$d$	ROC-AUC			PR-AUC			Transfer ROC-AUC
		DR	HL	OB	DR	HL	OB	DR	HL	OB
cgm_mean	1	65.4 $\pm$ 1.9	53.1 $\pm$ 1.6	57.0 $\pm$ 1.7	41.6 $\pm$ 2.6	22.0 $\pm$ 1.8	21.6 $\pm$ 1.6	65.1	55.1	57.1
iglu		70.3 $\pm$ 1.8	59.7 $\pm$ 1.5	69.7 $\pm$ 1.5	49.5 $\pm$ 3.1	26.4 $\pm$ 1.8	33.1 $\pm$ 2.6	67.6	60.8	67.8
iglu_daymatched	39	70.5 $\pm$ 1.8	58.2 $\pm$ 1.5	68.8 $\pm$ 1.4	–	–	–	–	–	–
raw	288	63.6 $\pm$ 1.9	53.0 $\pm$ 1.8	65.1 $\pm$ 1.6	40.5 $\pm$ 2.4	21.5 $\pm$ 1.7	27.0 $\pm$ 2.0	64.3	53.3	64.8
cgm_jepa	96	67.3 $\pm$ 1.8	55.8 $\pm$ 1.8	66.5 $\pm$ 1.6	44.7 $\pm$ 2.7	23.9 $\pm$ 1.8	28.0 $\pm$ 2.0	66.5	55.1	66.2

Appendix

Table 31: HPP per cell: dispersion, PR-AUC and the cross-device split. Every row of Table 5 , with the two metrics that table omits. ROC-AUC and PR-AUC are mean

\pm

standard deviation over

10\times 5

subject-grouped folds. Transfer fits on the cohort’s first sensor variant and scores on the second, dropping the

52

subjects who hold both (

9251

remain); it is one fit on a fixed split, so it carries no fold dispersion. d is the probe input width, which matters for reading the block rows: a narrow input is better regularised at any fixed

n

, so

z_{A}

at 16 dimensions is not comparable to

z_{T}

at 64 without the width-matched controls of Appendix J .

Comparison	Metric	DR	HL	OB
GlucoPRISM $-$ GlucoFM	AUROC	$+2.24$ [+1.36, +3.13] ∗	$-0.00$ [-1.20, +1.24]	$+2.51$ [+1.37, +3.66] ∗
GlucoPRISM $-$ GlucoFM	PR-AUC	$+4.36$ [+3.10, +5.53] ∗	$+0.69$ [-0.43, +1.78]	$+3.20$ [+1.65, +4.76] ∗
GlucoPRISM: $z_{A}$ deleted $-$ kept	AUROC	$+0.05$ [-0.11, +0.20]	$+0.18$ [-0.18, +0.55]	$-0.06$ [-0.41, +0.25]
GlucoPRISM: $z_{A}$ deleted $-$ kept	PR-AUC	$+0.12$ [-0.12, +0.34]	$+0.12$ [-0.21, +0.45]	$+0.24$ [-0.18, +0.67]
indices (day-matched) $-$ GlucoFM	AUROC	$+1.06$ [-0.04, +2.19]	$-1.57$ [-3.07, -0.18] ∗	$+0.07$ [-1.40, +1.60]
indices (day-matched) $-$ GlucoFM	PR-AUC	$+2.85$ [+1.32, +4.42] ∗	$-1.17$ [-2.41, +0.02]	$+1.28$ [-0.82, +3.21]

Appendix

Table 32: Paired participant-clustered bootstrap on HPP.

2000

replicates; participants are resampled with replacement and every window and prediction from one participant travels together, because windows within a participant are not independent. Both arms of a comparison are scored on the same resample, so these are intervals on the difference rather than on either score. Point estimates pool out-of-fold predictions and score once, where Table 5 scores each fold and averages; the two differ by a few tenths because AUROC is not linear in the pooled sample. ∗ marks

p<0.05

uncorrected. HPP stays descriptive and outside the family of 10 corrected in Section 4.4 .

Endpoint	Dev.	Test	$r$	MAE	$R^{2}$	$r$	MAE	$R^{2}$
			GlucoFM			GlucoPRISM
Apnea–hypopnea index	4505	683	0.2273	7.7357	0.0467	0.3026	7.5818	0.0882
Blood glucose	3397	542	0.6348	6.8216	0.3956	0.6698	6.7141	0.4483
Body mass index	7956	1214	0.3803	3.0570	0.1347	0.4364	3.0105	0.1902
Diastolic pressure	7971	1214	0.2214	7.8867	0.0442	0.2442	7.8359	0.0562
Body-fat field	4892	736	0.2601	0.0649	0.0438	0.4266	0.0605	0.1813

Appendix

Table 33: All baseline clinical endpoints, embedding-only probes. Higher

r

and

R^{2}

, and lower MAE, are better. MAE retains the recorded target units.

Endpoint	Dev.	Test	$r$	MAE	$R^{2}$	$r$	MAE	$R^{2}$
			GlucoFM			GlucoPRISM
Apnea–hypopnea index	4505	683	0.3395	7.3081	0.0903	0.4290	7.0974	0.1809
Blood glucose	3397	542	0.6553	6.8444	0.4229	0.6781	6.7304	0.4596
Body mass index	7956	1214	0.4009	3.0107	0.1518	0.4523	2.9717	0.2042
Diastolic pressure	7971	1214	0.2986	7.6949	0.0866	0.3120	7.6483	0.0964
Body-fat field	4892	736	0.6980	0.0464	0.4870	0.7405	0.0443	0.5439

Appendix

Table 34: All baseline clinical endpoints, embeddings plus common covariates. Higher

r

and

R^{2}

, and lower MAE, are better. MAE retains the recorded target units.

Endpoint	Dev.	Test	$r$	MAE	$R^{2}$	$r$	MAE	$R^{2}$
			GlucoFM			GlucoPRISM
Apnea–hypopnea index	1181	184	0.1252	6.7930	0.0007	0.2644	6.5406	0.0695
Blood glucose	338	53	0.6740	7.7609	0.3696	0.7324	6.9653	0.4845
Body mass index	1896	279	0.3638	2.8492	0.1308	0.3557	2.8879	0.1247
Diastolic pressure	1905	280	0.2016	7.1894	0.0317	0.1943	7.2523	0.0312
Body-fat field	1785	271	0.1933	0.0684	0.0028	0.4067	0.0637	0.1609

Appendix

Table 35: Two-year future clinical levels on the common test, embedding-only probes. Higher

r

and

R^{2}

, and lower MAE, are better. MAE retains the recorded target units.

Endpoint	Dev.	Test	$r$	MAE	$R^{2}$	$r$	MAE	$R^{2}$
			GlucoFM			GlucoPRISM
Meal incremental area	7079	1069	0.4936	1140.4195	0.2410	0.3527	1223.9176	0.1231
Meal peak rise	7079	1069	0.4025	13.9851	0.1602	0.3163	14.6269	0.0978
Glucose at +120 min	7998	1218	0.7438	9.2290	0.5530	0.7085	9.8908	0.5017
Glucose at +60 min	7998	1218	0.7597	9.0759	0.5769	0.7078	10.0525	0.5009
Next-day glucose CV	7819	1196	0.5504	0.0294	0.3027	0.5266	0.0302	0.2772

Appendix

Table 36: Meal and short-horizon tasks, embedding-only probes. Next-day fraction uses the threshold 180 mg/dL. Higher

r

and

R^{2}

, and lower MAE, are better. MAE retains the recorded target units.

Endpoint	Trait	State	Sensor	Trait+State	All
Apnea–hypopnea index	0.2843	0.3046	0.2830	0.3026	0.3076
Blood glucose	0.6226	0.6692	0.5811	0.6698	0.6509
BMI	0.4036	0.4211	0.3202	0.4364	0.4349
Diastolic pressure	0.2416	0.2526	0.2067	0.2442	0.2532
Body-fat field	0.3669	0.3781	0.2434	0.4266	0.4262
HbA1c	0.7116	0.7263	0.5691	0.7307	0.7336

Appendix

Table 37: Baseline clinical readout ablation: Pearson correlation (higher is better).

Endpoint	Trait	State	Sensor	Trait+State	All
Meal incremental area	0.2771	0.3022	0.0848	0.3527	0.3669
Meal peak rise	0.2657	0.2785	0.1038	0.3163	0.3257
Glucose +120 min	0.6769	0.6993	0.4237	0.7085	0.7089
Glucose +60 min	0.6714	0.6945	0.3864	0.7078	0.7107
Next-day glucose CV	0.4996	0.4992	0.2833	0.5266	0.5291
Next-day fraction $>180$	0.8360	0.8848	0.2568	0.8905	0.8916

Appendix

Table 38: Meal and short-horizon readout ablation: Pearson correlation (higher is better).

Readout	Device AUROC	Device AP	Bal. acc.	Mean cosine	Median cosine
Trait	0.8467	0.7793	0.7564	0.4378	0.9848
State	0.8515	0.7836	0.7741	0.4024	0.9229
Sensor	0.8300	0.7354	0.7277	0.2667	0.4347
T+S	0.8599	0.7982	0.7703	0.4275	0.8028
All	0.8607	0.7992	0.7670	0.4145	0.7113

Appendix

Table 39: Device information and paired-visit stability by readout. The first three columns concern device discrimination (1,219 people); the last two concern cosine similarity (44 people). These are different tasks and are not combined into a single score.

Baselines that need no learned representation
Readout	dim	pooled $r$	between $r$	within $r$	recovers pre-meal
Meal time of day (2 features)	2	0.083	0.028	0.097	–
Macronutrients, winsorised	4	0.195	0.150	0.218	–
Pre-meal glucose (1 feature)	1	0.265	-0.006	0.437	–
Macronutrients + clock + pre-meal	7	0.343	0.067	0.483	–
CGM foundation models, frozen-encoder probing

Appendix

Table 40: Postprandial response for

136true374

meals from

8282

participants. Frozen encoders read the preceding 24 hours; incremental area uses the following two hours. No response sample enters the input, but the last pre-meal input reading is subtracted as the label baseline. Ridge probes use subject-grouped folds. Between-participant correlation compares participant means; within-participant correlation centers within each participant. The final column separately probes retention of pre-meal glucose.

Context	Endpoint	GluFormer (tiny)	CGM-JEPA	X-CGM-JEPA	GlucoFM	GlucoPRISM
Frozen representation	Trajectory (mg/dL)	26.55 $\pm$ 0.26	24.90 $\pm$ 0.28	24.65 $\pm$ 0.18	26.64 $\pm$ 0.33	26.43 $\pm$ 0.14
	Positive iAUC (mg/dL $\cdot$ h)	37.85 $\pm$ 0.75	35.16 $\pm$ 0.50	34.67 $\pm$ 0.33	37.92 $\pm$ 0.71	37.79 $\pm$ 0.32
	Peak rise (mg/dL)	35.07 $\pm$ 1.09	32.66 $\pm$ 0.63	32.25 $\pm$ 0.47	34.91 $\pm$ 0.72	35.53 $\pm$ 0.31
	Peak time (min)	30.52 $\pm$ 1.62	30.29 $\pm$ 0.33	30.44 $\pm$ 0.49	31.50 $\pm$ 1.77	30.04 $\pm$ 0.75
+ 1 h pre-meal CGM	Trajectory (mg/dL)	25.75 $\pm$ 0.24	24.37 $\pm$ 0.15	24.27 $\pm$ 0.17	24.25 $\pm$ 0.13	24.08 $\pm$ 0.21
	Positive iAUC (mg/dL $\cdot$ h)	36.57 $\pm$ 0.57	34.61 $\pm$ 0.30	34.41 $\pm$ 0.30	34.21 $\pm$ 0.29	33.92 $\pm$ 0.44

Appendix

Table 41: PPGR on the Dexcom evaluation (5-min native grid). Mean absolute error (lower is better), mean

\pm

SD over 10 probe seeds. Bold : best; underline : second best.

Context	Endpoint	GluFormer (tiny)	CGM-JEPA	X-CGM-JEPA	GlucoFM	GlucoPRISM
Frozen representation	Trajectory (mg/dL)	25.17 $\pm$ 0.21	22.56 $\pm$ 0.47	22.21 $\pm$ 0.49	25.03 $\pm$ 0.13	25.31 $\pm$ 0.10
	Positive iAUC (mg/dL $\cdot$ h)	34.24 $\pm$ 0.53	30.03 $\pm$ 0.82	29.41 $\pm$ 0.97	33.68 $\pm$ 0.17	34.61 $\pm$ 0.30
	Peak rise (mg/dL)	30.12 $\pm$ 0.88	26.82 $\pm$ 0.68	26.38 $\pm$ 0.70	30.33 $\pm$ 0.26	31.10 $\pm$ 0.38
	Peak time (min)	31.26 $\pm$ 0.54	30.73 $\pm$ 0.64	30.51 $\pm$ 0.63	31.04 $\pm$ 0.86	30.68 $\pm$ 0.55
+ 1 h pre-meal CGM	Trajectory (mg/dL)	24.18 $\pm$ 0.40	22.17 $\pm$ 0.47	21.90 $\pm$ 0.44	22.66 $\pm$ 0.20	22.54 $\pm$ 0.20
	Positive iAUC (mg/dL $\cdot$ h)	32.81 $\pm$ 0.75	29.71 $\pm$ 0.84	29.31 $\pm$ 0.79	30.19 $\pm$ 0.42	30.24 $\pm$ 0.36

Appendix

Table 42: PPGR on the Libre evaluation (15-min native grid). Mean absolute error (lower is better), mean

\pm

SD over 10 probe seeds. Bold : best; underline : second best.

Figure 11: Two-dimensional projections of frozen representations for GlucoFM (top) and GlucoPRISM (bottom), coloured by label: (a) diabetes status on CGMacros (Libre), (b) insulin resistance on CGMacros (Dexcom), (c) glucotype on Hall, (d) $\beta$ -cell dysfunction on Stanford, and (e) the same Stanford labels with a label-supervised projection. Both rows use the same windows. Panel (e) separates classes by construction and is shown for reference only.

Figure 12: Effect of multi-day observation for GlucoPRISM: paired PR-AUC change from one-day to $K$ -day subject representations ( $K=2,\dots,7$ ), in percentage points, for each task and cohort. Following Li et al. (2026) , Hall uses $K\leq 4$ and the other cohorts use $K\leq 7$ ; dashes mark values of $K$ that are not evaluated. Cells are point estimates without uncertainty intervals.

Figure 13: Multi-day observation for GlucoPRISM on ShanghaiT2DM insulin resistance under concat(mean, max) aggregation, following Li et al. (2026) : paired PR-AUC change from one-day to

K

-day subject representations, in percentage points. Bars are point estimates without uncertainty intervals.

Days $K$	GlucoFM	GlucoPRISM	GlucoPRISM [ $z_{A}$ kept]
1	61.42	64.94	63.94
2	64.80	67.15	66.37
3	65.74	68.41	67.13
5	68.26	68.88	67.71
7	68.81	69.01	68.06

Appendix

Table 43: Subject-level ROC-AUC using mean embeddings from

K

randomly drawn 24-hour windows per subject; the windows need not be consecutive. The effect of additional days differs by model and readout; these comparisons do not establish that a representation change substitutes for a fixed number of clinical monitoring days.

Figure 14: Design ablations for GlucoPRISM, reported as scores averaged over the within-cohort benchmark cells. (a) Training-time augmentation. (b) Input interpolation and masking. (c) Temporal-dynamics weight

\lambda_{TD}

; the shaded band marks the default. The default configuration is shown in orange. Bars and solid lines are means; diamonds and dashed lines are a single seed-1 run. Values come from the ablation protocol and are not directly comparable with Table 2 .

Cohort	7-day GMI rule	GlucoPRISM	$\Delta$
Stanford	47.6	57.1 $\pm$ 6.2	$+$ 9.5
CGMacros-Dexcom	39.6	59.1 $\pm$ 2.3	$+$ 19.5
CGMacros-Libre	57.1	51.3 $\pm$ 2.6	$-$ 5.8
CGMacros-Fused	44.9	54.1 $\pm$ 3.3	$+$ 9.2
Average (4 cohorts)	47.3	55.4	$+$ 8.1

Appendix

Table 44: Macro-F1 (%) of a seven-day GMI threshold rule and GlucoPRISM. GlucoPRISM values are mean

\pm

SD; the rule is deterministic.

\Delta

is GlucoPRISM minus the rule, in percentage points. Bold marks the higher score in each row.

Figure 15: HPP pretraining corpus-size comparison on public cohorts. AUROC (top) and average precision (AP; bottom) are percentages; higher is better. Each panel compares identical nested participant subsets and downstream splits. Window and subject scores average 14 task–cohort cells; transfer scores average 12 directions. One pretraining seed is available. Lines connect observations without smoothing; no uncertainty or significance is inferred. Axis ranges differ across panels.

Figure 16: Corpus composition and window geometry for the GlucoFM backbone. (a) Window AUROC as the retained fraction of REPLACE-BG varies; the shaded band is

\pm 1

seed s.d. This panel comes from a separate sweep and is not seed-matched to panel (b). (b) Window AUROC by window and patch geometry. The dark bar is the geometry used for every model in this paper (40% day overlap, all windows kept); published is the original GlucoFM geometry. Bars marked

\dagger

are single runs; the others are three-seed means.

Continuous Glucose Monitoring

See all 16 papers

May 29, 2026

GlucoFM: A Dual-Stream Foundation Model for Continuous Glucose Monitoring

Continuous glucose monitoring (CGM) provides a dense view of daily metabolic physiology, yet existing generic time-series and CGM-specific foundation models often encode glucose traces as entangled single-stream sequences, leaving the distinct temporal structure of glycemic dynamics only implicitly modeled. We present GlucoFM, a lightweight CGM foundation model that aligns irregular recordings to a 24-hour chronological grid, preserves observation masks, and decomposes glucose dynamics into slow physiological state and transient event streams, capturing low-frequency glycemic baselines and short-term deviations that may reflect acute physiological responses or sensor artifacts. GlucoFM is pretrained on 109,066 hours of unlabeled CGM recordings from 477 subjects with two complementary objectives: masked contextual latent prediction over fused daily representations and temporal dynamics prediction over state and event streams. Across four diverse cohorts and seven clinical prediction tasks, GlucoFM achieves the strongest subject-disjoint linear-probing performance among evaluated baselines, improving average PR-AUC by 4.1 points over the best CGM-specific foundation model. Its gains are most pronounced on core metabolic outcomes, leading PR-AUC on all diabetes-risk and

β

-cell dysfunction tasks and on 3 of 4 insulin-resistance tasks. GlucoFM also achieves the best overall cross-dataset transfer performance and strong few-shot adaptation among evaluated methods, and consistent gains when aggregating multiple days for subject-level prediction, highlighting physiology-aware decomposition as an effective inductive bias for transferable CGM representation learning.

Zechen Li, Keerthana Natarajan, Weizhi Zhang +11

Google Research · Work done during an internship at Google · University of New South Wales, Sydney

Irregular Time-Series Modeling Latent Dynamics Modeling Clinical Outcome Prediction Wearable Health Monitoring Time-Series Foundation Models Continuous Glucose Monitoring

Ask PDF

May 1, 2026

CGM-JEPA: Learning Consistent Continuous Glucose Monitor Representations via Predictive Self-Supervised Pretraining

Continuous Glucose Monitoring (CGM) can detect early metabolic subphenotypes (insulin resistance, IR;

β

-cell dysfunction), but population-scale deployment faces two coupled problems. First, the same physiological state appears through multiple views (CGM time series, venous OGTT, Glucodensity summaries), so single-view representations fail to transfer when deployment shifts the modality or setting. Second, baselines perform inconsistently across these shifts. Both problems point to one remedy: representations that abstract away from any single view to capture higher-level temporal and distributional structure. We propose CGM-JEPA, a self-supervised pretraining framework which predicts masked latent representations rather than raw values, yielding abstraction that transfers across modalities. X-CGM-JEPA adds a masked Glucodensity cross-view objective for complementary distributional information. We pretrain on

\sim

389k unlabeled CGM readings from 228 subjects and evaluate on two clinical cohorts (

N=27

and

N=17

public-release subsets) across three regimes (cohort generalization, venous-to-CGM transfer, home CGM) under 20-iteration

\times

2-fold cross-validation. X-CGM-JEPA ranks first or second on AUROC for both endpoints across all three regimes while no baseline does, exceeding the strongest baseline by up to

+6.5

pp in cohort generalization and

+3.6

pp in venous-to-CGM transfer (paired Wilcoxon,

p<0.001

). Under modality shift, it matches mean AUROC while redistributing toward weaker subgroups (ethnicity AUROC gap shrinks 25-54%); on sparse in-domain venous data, the distributional view lifts label-aware clustering (ARI

+39\%

, NMI

+40\%

). Code and weights: https://github.com/cruiseresearchgroup/CGM-JEPA

Hada Melino Muhammad, Zechen Li, Flora Salim +1

University of New South Wales · Google Research

Cross-Modal Representation Learning Self-Supervised Time Series Representation Learning Joint-Embedding Predictive Architecture Predictive Representation Learning Time Series Representation Learning Continuous Glucose Monitoring

Ask PDF

Sep 23, 2026

Transferable Evidence Reconstruction for Longitudinal Glucose Representations

Long physiological recordings contain many routine measurements, while predictive information often lies in rare events, sustained burden, and recurring patterns. These properties can be computed as label-free evidence, but directly using them as features leaves limited labeled data to separate reproducible associations from sample-specific ones. Learning to reconstruct evidence can exploit unlabeled recordings, yet joint reconstruction does not explicitly require the decoding rule to transfer across individuals. We introduce transferable evidence reconstruction (TER): a Ridge regressor fits evidence from representations in one group and predicts it in an identity-disjoint group without refitting. The transfer error trains the encoder through the differentiable fit. For continuous glucose monitoring (CGM), clock-aware encoding preserves the multi-day content and timing needed for evidence recovery. Matched interventions connect the gains to reduced fitting-group sensitivity, with structured targets improving on raw recovery. Across ten leading CGM and time-series baselines, TER sets a new best metric on 12/14 phenotype tasks and exceeds the strongest prior overall PR-AUC/ROC-AUC/Macro-F1 by 4.95/4.43/0.66 percentage points; the PR-AUC and ROC-AUC gains are

2.6\times

and

2.2\times

the respective gaps between the two strongest baselines. Meal-response and future-CGM studies further demonstrate predictive utility. TER thus uses meaningful signal properties to supervise not only what a representation preserves, but how reliably it can be read across individuals.

Tian Zhou, Bingqing Peng, Linxiao Yang +6

Ant Group · Independent Researcher

Representation Learning Self-Supervised Time Series Representation Learning Wearable Health Monitoring Continuous Glucose Monitoring

Ask PDF

Abstract

Continuous Glucose Monitoring

Explore similar work

GlucoFM: A Dual-Stream Foundation Model for Continuous Glucose Monitoring

CGM-JEPA: Learning Consistent Continuous Glucose Monitor Representations via Predictive Self-Supervised Pretraining

Transferable Evidence Reconstruction for Longitudinal Glucose Representations