Adaptation Beats Scale: What Actually Drives Postprandial CGM Forecasts and Cross-Cohort Transfer

Question: How do foundation models and deep sequence models for continuous glucose monitoring forecast postprandial glucose excursions, and how well do they transfer across cohorts and eating contexts?

Answer

Taken together, these papers support a deflationary position on scale and an affirmative one on adaptation. Zero-shot time-series foundation models do not reliably beat strong task-specific baselines on CGM, and their advantage materialises only after CGM-specific fine-tuning or a tabular/physiological prior is added [4] [12] [5]. Cross-cohort transfer does occur — CGM-pretrained encoders generalise to held-out patient groups and to external cohorts spanning multiple countries and devices [2] [6] — but the measured benefit is a few AUROC points on average, and it is concentrated where device or modality composition shifts rather than where physiology shifts [14] [10]. Postprandial error specifically is governed less by architecture than by whether meal context is supplied and how: residual or gated dietary fusion cuts postprandial RMSE by roughly 15% while cutting overall RMSE by only ~3% [4], and a gated meal-informed residual outperforms LSTM, transformer, and GluFormer architectures on peak, time-to-peak and incremental AUC [11]. Across these papers, the field's binding constraints are sensor heterogeneity, eating-context measurement, and evaluation fragmentation — not model capacity.

Zero-shot time-series foundation models do not beat tuned baselines on CGM; the gain arrives only with CGM-specific fine-tuning

In a unified protocol across eight public CGM datasets spanning type 1, type 2, and non-diabetes populations, zero-shot foundation models did not consistently outperform task-specific baselines such as Elastic Net and PatchTST [4]. Lightweight fine-tuning changed the picture: fine-tuned Chronos-Bolt reduced RMSE by 6.5–18.4% in the T1D cohort and 8.6–18.2% in the non-diabetes/T2D cohort, with comparable gains in-distribution and out-of-distribution [4]. The same pattern holds outside forecasting: for week-ahead CGM metrics, AutoML and tabular foundation models (tabPFN, AutoGluon) yielded modest accuracy gains over CatBoost/XGBoost at substantially higher computational cost — 699 s versus 0.04 s per 1,000 cases for tabPFN versus XGBoost [12]. A 10K-parameter architecture-agnostic training framework (region-aware loss, genetic weight optimisation) matched TimesNet (18.7M), BG-BERT (2.1M) and GluFormer (11.2M) on OhioT1DM and AZT1D [5].

CGM-native pretraining transfers across cohorts, but by a few AUROC points, not a regime change

CGM-LSM, a decoder transformer pretrained on 1.6 million CGM records, reports a 48.51% RMSE reduction at the one-hour horizon and consistent zero-shot performance across held-out patient groups [2]. GluFormer, trained on over 10 million measurements from 10,812 adults, generalises to 19 external cohorts (n=6,044) spanning five countries, eight CGM devices and diverse pathophysiological states [6]. Where transfer is quantified as a delta rather than a demonstration, the magnitude is modest: dropping a sensor-associated representation band raised mean cross-cohort transfer AUROC from 68.75 to 73.57 across twelve transfer directions, against 69.3 for the GlucoFM backbone [14]. A predictive-self-supervised alternative (X-CGM-JEPA) exceeded the strongest baseline by up to +6.5 percentage points in cohort generalisation and +3.6 points in venous-to-CGM transfer [10]. Multimodal transformers also forget: AttenGluco improved RMSE by ~10% and MAE by ~15% over a multimodal LSTM across healthy, prediabetes and T2D cohorts on AI-READI, while exhibiting forgetting behaviour as new cohorts were introduced [3].

Sensor and device heterogeneity is the removable part of transfer failure; physiology and eating context are not

The explicit ask in GlucoPRISM is whether the measurement component that harms transfer can be placed in known coordinates; gains were concentrated in CGMacros, the one target cohort recorded with two sensor types, where filtering added 11 to 19 points, while on the other eight directions filtering was neutral and the encoder itself exceeded GlucoFM by 3.5 points [14]. The device remained linearly decodable after filtering, indicating a device-related component was removed rather than all measurement information [14]. Modality shift is a distinct axis: under venous-to-CGM transfer, X-CGM-JEPA matched mean AUROC while shrinking the ethnicity AUROC gap by 25–54% [10]. The size of the measurement problem is visible in device-agnostic clinical data — CGM versus point-of-care agreement showed a MARD of 17.4% in older inpatients with cognitive impairment [26] — which sets a noise floor against which forecast RMSE differences of a few mg/dL must be read.

Dietary context is the largest single lever on postprandial error, and gated residual fusion extracts it without requiring complete logs

Residual-based fusion of CGM with food images and macronutrient records reduced overall RMSE by approximately 3% but postprandial RMSE by approximately 15% relative to a CGM-only baseline [4]. MealRes-Gate produces a complete CGM-only forecast and then adds a horizon-scaled, gated meal-informed residual controlled jointly by self-reported and CGM-derived evidence of meal presence, achieving the lowest MAE across horizons, meal-state strata and glucose ranges against LSTM, transformer and GluFormer, and improving peak glucose, time-to-peak and positive incremental AUC [11]. Two findings qualify the value of dietary inputs. First, dietary effects are temporally non-stationary within the response window — fibre blunted responses after 90 minutes while fats reduced early rises within 50 minutes [20]. Second, within-individual meal-to-meal variability dominates free-living postprandial responses, with between-individual variation concentrated in two dimensions, and three days of monitoring approximating fourteen [19]. On the input side, image-derived macronutrients approach the performance of manually reported macronutrients when combined with clinical and microbiome data [22]; a QLoRA-tuned 4B language model estimated meal carbohydrate with MAE 12.2 g versus 22.6 g for a food composition table, and the resulting carbohydrate signal fed a gradient-boosting model reaching r=0.63 in CGMacros and r=0.49 in a Shanghai T2D cohort when transferred between datasets [21]. Carbohydrate and calorie content alone are weak: a model using nutritional content, habits, biometrics and laboratory data reached r=0.69 and AUC 0.80 against r=0.39 for carbohydrate-only and r=0.57 for calorie-only models in 488 Indian adults with T2D [18].

Short-horizon error has hit an information ceiling where architecture stops mattering; long horizons are a bias problem

Under subject-grouped, leakage-controlled evaluation, 30-minute-ahead forecasting from CGM history saturated at RMSE 13.90 ± 0.58 mg/dL, with ordinary linear regression matching gradient-boosted trees, fully convolutional and temporal convolutional networks — convergence the authors read as an information ceiling rather than a modelling limitation [17]. Independently, patch-wise transformers reached 15.6 mg/dL at 30 minutes on OhioT1DM as an external test set [1], and a physiologically constrained TCN reported median MAE of 16.24 mg/dL at 30 minutes in 25 AZT1D participants [16]. A time-of-day mean forecast — no learned model at all — reduced MSE by an average of 12.2% versus a naive baseline at 5-minute bins [25], which is the benchmark any postprandial claim must clear. Beyond one hour, the problem becomes bias compounding: a sequence-to-sequence LSTM with patient-matched physiological states achieved 240-minute MAE of 39.28 mg/dL and mean error of −10.62 mg/dL, reducing bias by 13.89 mg/dL over a recursive LSTM and MAE by 28.62 mg/dL over an ODE digital twin [27]. Accounting for meal absorption and exercise improved 60-minute RMSE from 26.68 to 23.89 mg/dL in silico and from 37.02 to 35.96 mg/dL clinically [13], and a patient-specific pharmacokinetic encoder improved accuracy by up to 16.4% on simulated and 4.9% on real-world data during severe excursions [7].

Evaluation fragmentation, not modelling capacity, is the binding constraint on transfer claims

The absence of standardised preprocessing workflows and evaluation protocols is explicitly identified as the obstacle to reproducibility and fair comparison, motivating a YAML-configurable framework with a cross-dataset leaderboard [23]. Multimodal claims are especially fragile: in a benchmark over glucose, insulin and carbohydrate inputs, the benefit of adding modalities was conditioned on model complexity [9], and a consolidated 3,135-subject, 1,228-patient-year resource exists precisely because fragmented datasets reduce comparability [24]. Excursion-specific objectives can move the needle without new architectures — a hypo-hyper loss improved excursion-region performance by 46% over MSE across 125 patients, and federated training improved excursion detection by 35% over local models [8].

Conflicts and gaps

The zero-shot transfer claims of CGM-LSM [2] and GluFormer [6] sit uneasily with the unified empirical finding that zero-shot foundation models do not consistently beat Elastic Net or PatchTST [4]; the difference plausibly lies in what "zero-shot" is measured against, since the former compare to prior CGM-specific methods and the latter to tuned baselines including linear models. Postprandial performance is almost never reported as a primary stratified metric, so headline RMSE numbers conflate fasting, nocturnal and postprandial windows whose error structure differs [17]. Eating-context transfer remains nearly unmeasured: no retrieved paper evaluates a model trained on one dietary logging regime (weighed records, image-derived estimates, sparse recalls) and tested under another, despite evidence that logging completeness and modality both change the input distribution [4] [11] [21]. T2D and non-diabetes cohorts are underrepresented relative to T1D, and free-living versus standardised-meal contexts are not systematically crossed [19].

What would change the answer

A pre-registered cross-cohort benchmark that (i) stratifies error by postprandial window and meal state, (ii) includes a time-of-day mean and Elastic Net as mandatory baselines, and (iii) holds out entire cohorts and eating-context regimes rather than participants, would settle whether the transfer gains reported for sensor-filtering [14] and cross-view pretraining [10] survive context shift rather than only device shift. Paired multi-sensor recordings — participants wearing two CGM devices simultaneously across cohorts, as CGMacros partially provides — are the cleanest instrument for separating sensor from physiology, and the GlucoPRISM recalibration-on-half-the-participants design is the template to extend [14]. A factorial dietary-context experiment that crosses logging modality (image, text, weighed, none) with fusion architecture on a fixed encoder would test whether the ~15% postprandial gain from dietary fusion [4] is attributable to information or to inductive bias. Finally, meal-level rather than window-level splits, and prospective deployment against the exercise-related and excursion-detection targets already validated in the field [15] [8], would establish whether any of these models change glycemic outcomes.

References

[1] Meryem Altin Karagoz, Marc D. Breton, Anas El Fathi. "A Comparative Study of Transformer-Based Models for Multi-Horizon Blood Glucose Prediction". arxiv. 2025-05-15T00:00:00.000Z. URL: https://arxiv.org/abs/2505.08821

[2] Junjie Luo, Abhimanyu Kumbara, Mansur Shomali, Rui Han, Anand Iyer, Ritu Agarwal, Gordon Gao. "A Large Sensor Foundation Model Pretrained on Continuous Glucose Monitor Data for Diabetes Management". arxiv. 2025-08-04T00:00:00.000Z. URL: https://arxiv.org/abs/2412.09727

[3] Ebrahim Farahmand, Reza Rahimi Azghan, Nooshin Taheri Chatrudi, Eric Kim, Gautham Krishna Gudur, Edison Thomaz, Giulia Pedrielli, Pavan Turaga, Hassan Ghasemzadeh. "AttenGluco: Multimodal Transformer-Based Blood Glucose Forecasting on AI-READI Dataset". arxiv. 2025-02-17T00:00:00.000Z. URL: https://arxiv.org/abs/2502.09919

[4] Bowen Zhang, Hsiu-Wen Cheng, Hongyu Yang, Evie L. Shen, Joleen Vansomphone, Yuna Li, Kerry Zhou, Zitian Qu, Suning Zhao, Xiangning Deng, Hua Zhou, Jin J. Zhou. "Evaluating Time-Series Foundation Models and Multimodal Dietary Context for CGM Forecasting". arxiv. 2026-09-11T00:00:00.000Z. URL: https://arxiv.org/abs/2609.11872

[5] Saman Khamesian, Asiful Arefeen, Maria Adela Grando, Bithika M. Thompson, Hassan Ghasemzadeh. "Glycemic-Aware and Architecture-Agnostic Training Framework for Blood Glucose Forecasting in Type 1 Diabetes". arxiv. 2026-07-02T00:00:00.000Z. URL: https://doi.org/10.1186/s44398-026-00032-x

[6] Guy Lutsker, Gal Sapir, Smadar Shilo, Jordi Merino, Anastasia Godneva, Jerry R Greenfield, Dorit Samocha-Bonet, Raja Dhir, Francisco Gude, Shie Mannor, Eli Meirom, Gal Chechik, Hagai Rossman, Eran Segal. "From Glucose Patterns to Health Outcomes: A Generalizable Foundation Model for Continuous Glucose Monitor Data Analysis". arxiv. 2025-01-08T00:00:00.000Z. URL: https://arxiv.org/abs/2408.11876

[7] Willa Potosnak, Cristian Challu, Kin G. Olivares, Keith A. Dufendach, Artur Dubrawski. "Global Deep Forecasting with Patient-Specific Pharmacokinetics". arxiv. 2025-12-17T00:00:00.000Z. URL: https://arxiv.org/abs/2309.13135

[8] Darpit Dave, Kathan Vyas, Jagadish Kumaran Jayagopal, Alfredo Garcia, Madhav Erraguntla, Mark Lawley. "FedGlu: A personalized federated learning-based glucose forecasting algorithm for improved performance in glycemic excursion regions". arxiv. 2025-12-17T00:00:00.000Z. URL: https://doi.org/10.1038/s41598-025-22316-4

[9] Nathaniel Jeffries, Miriam Wolff, Sam Royston, Elizabeth Healey, Caleb Mayer, David Klonoff, Michael Snyder, Tao Wang. "MetaboNet-Bench: A Multi-modal Benchmark for Glucose Forecasting in Type 1 Diabetes". arxiv. 2026-06-26T00:00:00.000Z. URL: https://arxiv.org/abs/2606.1864

[10] Hada Melino Muhammad, Zechen Li, Flora Salim, Ahmed A. Metwally. "CGM-JEPA: Learning Consistent Continuous Glucose Monitor Representations via Predictive Self-Supervised Pretraining". arxiv. 2026-05-05T00:00:00.000Z. URL: https://arxiv.org/abs/2605.00933

[11] Lee, J., Yao, S., Tang, L., Yu, X., Cheney, M. C., Lin, H., Zhang, X., Mukherjee, D., Walker, M. E., Spartano, N. L., Cheng, H.. "Forecasting glucose from CGM and sparse meal logs with a residual-gated multimodal transformer". medRxiv. 2026-08-30T00:00:00.000Z. URL: https://doi.org/10.64898/2026.06.01.26354646

[12] Simon Lebech Cichosz, Stine Hangaard, Thomas Kronborg, Peter Vestergaard, Morten Hasselstrøm Jensen. "Personalized Forecasting of Glycemic Control in Type 1 and 2 Diabetes Using Foundational AI and Machine Learning Models". arxiv. 2026-01-05T00:00:00.000Z. URL: https://arxiv.org/abs/2601.00613

[13] Chengyuan Liu, Josep Vehi, Nick Oliver, Pantelis Georgiou, Pau Herrero. "Enhancing Blood Glucose Prediction with Meal Absorption and Physical Exercise Information". arxiv. 2024-12-20T00:00:00.000Z. URL: https://arxiv.org/abs/1901.07467

[14] Adinath Madhavrao Dukre, Abdullah Mohammad, Yifan Lu, Eran Segal, Ahmed Metwally, Imran Razzak. "GlucoPRISM: Protocol-Supervised Pretraining for Cross-Cohort Glucose Transfer". arxiv. 2026-10-12T00:00:00.000Z. URL: https://arxiv.org/abs/2610.12528

[15] Russon, C. L., Allen, M., Pulsford, R. M., Vaughan, N., Cockcroft, E., Pemberton, J. S., Frohock, A. M., Andrews, R. C.. "GlucoseGo: A Simple, User-Friendly, Machine Learning-Derived Tool for Predicting Exercise-Related Hypoglycaemia Risk in Type 1 Diabetes". medRxiv. 2025-04-25T00:00:00.000Z. URL: https://doi.org/10.1101/2025.04.24.25326242

[16] Yaowen Zhang, Shuhao Que, Ying Wang. "PMB-NN: Physiological-Model-Based Neural Network for Glucose Forecasting in Type 1 Diabetes". arxiv. 2026-10-16T00:00:00.000Z. URL: https://arxiv.org/abs/2610.15456

[17] seyedebrahimi, M., ojeda, c., Zarrintaj, P.. "A Leakage-Controlled Evaluation of Multimodal Sensor Fusion for Wrist-Worn Glucose Estimation". medRxiv. 2026-08-04T00:00:00.000Z. URL: https://doi.org/10.64898/2026.08.03.26359550

[18] Choudhry, N. K., Priyadarshini, S., Swamy, J., Mehta, M.. "PERSONALIZED GLYCEMIC RESPONSES TO FOOD AMONG INDIVIDUALS WITH TYPE 2 DIABETES IN INDIA: DEVELOPMENT OF A MACHINE LEARNING PREDICTION MODEL". medRxiv. 2024-10-22T00:00:00.000Z. URL: https://doi.org/10.1101/2024.10.20.24315560

[19] Toumi, M., Salathe, M.. "Short-term postprandial glucose monitoring reveals stable traits from noisy free-living meals". medRxiv. 2026-09-15T00:00:00.000Z. URL: https://doi.org/10.64898/2026.09.14.26362991

[20] Marcos Matabuena, Joe Sartini, Francisco Gude. "Beyond Scalar Metrics: Functional Data Analysis of Postprandial Continuous Glucose Monitoring in the AEGIS Study". arxiv. 2026-04-03T00:00:00.000Z. URL: https://doi.org/10.1186/s12874-025-02748-2

[21] Bou Ali, M. N., Lassoued, N., Ben Amor, B., Nasr, S., Aouididi, M. A.. "From meal text to carbohydrate to postprandial glucose rise: QLoRA fine-tuning of the Qwen3-4B language model and gradient-boosting prediction in type 2 diabetes". medRxiv. 2026-10-19T00:00:00.000Z. URL: https://doi.org/10.64898/2026.10.15.26365634

[22] Varvara Kondratyeva, Kamilia Zaripova, Nassir Navab, Azade Farshad. "Predicting Postprandial Glycemic Response from Meal Images, Clinical Variables, and Gut Microbiome Information". arxiv. 2026-09-22T00:00:00.000Z. URL: https://arxiv.org/abs/2609.24453

[23] Davide Marelli, Giorgia Rigamonti, Mirko Paolo Barbato, Paolo Napoletano. "GlucoTune: A Unified Framework for Blood Glucose Preprocessing, Forecasting, and Benchmarking in Diabetes". arxiv. 2026-07-24T00:00:00.000Z. URL: https://arxiv.org/abs/2607.21117

[24] Miriam K. Wolff, Peter Calhoun, Eleonora Maria Aiello, Yao Qin, Sam F. Royston. "MetaboNet: The Largest Publicly Available Consolidated Dataset for Type 1 Diabetes Management". arxiv. 2026-04-23T00:00:00.000Z. URL: https://arxiv.org/abs/2601.11505

[25] Pedersen, N. P. B., Bugajski, T. K., Vera-Valdes, J. E., Casper, S. H., Jensen, M. H., Vestergaard, P., Kronborg, T.. "Daily Rhythms in Blood Glucose: Time-of-Day Forecasts in Type 2 Diabetes". medRxiv. 2025-06-05T00:00:00.000Z. URL: https://doi.org/10.1101/2025.06.02.25328803

[26] Donat-Ergin, B., Mattishent, K., Minihane, A. M., Holt, R., Murphy, H., Dhatariya, K., Hornberger, M.. "CGM accuracy and reliability compared to point of care testing in older inpatients with comorbid type 2 diabetes and cognitive impairment". medRxiv. 2026-03-31T00:00:00.000Z. URL: https://doi.org/10.64898/2026.03.27.26349485

[27] Phat Tran, Neville Mehta, Clara Mosquera-Lopez, Robert H. Dodier, Lizhong Chen, Peter G. Jacobs. "PhysioSeq2Seq: A Hybrid Physiological Digital Twin and Sequence-to-Sequence LSTM for Long-Horizon Glucose Forecasting in Type 1 Diabetes". arxiv. 2026-05-19T00:00:00.000Z. URL: https://arxiv.org/abs/2605.1686

How this survey was made

This survey answers one question: “How do foundation models and deep sequence models for continuous glucose monitoring forecast postprandial glucose excursions, and how well do they transfer across cohorts and eating contexts?”

We keep our own collection of recent preprints and refresh it every day from bioRxiv, medRxiv, arXiv, PsyArXiv, SocArXiv and ChemRxiv. For each survey, we search that collection for recent work and keep the strongest matches: papers that use the same terms as the question, and papers on the same topic in different words.

The text was written from the titles and abstracts of those papers; the full texts were not read. Statements that draw on a particular paper are marked with a number in brackets. The reference list just above is the 27 papers whose work this text draws on, chosen from the 127 papers those searches returned; the rest of the shortlist is not printed. It is not a complete review of everything published on the question.

Treat this as a starting point and a reading list, not a peer-reviewed review. The searches behind this survey ran on October 22, 2026 and covered work from the previous two years.