PaperlayerView exact revision on arXiv

Source-linked AI summary

Reliability-Aware Short-Term Glycemic Risk Prediction Using an Anchor-Based Comparative Framework of Sequence, Multimodal, and State-Updating Digital-Twin Models

Pir Bakhsh Khokhar, Carmine Gravino, Fabio Palomba

arXiv:2610.36363v1cs.LGcs.AI

TL;DR

Short-term glycemic-risk research lacks unified multimodal evidence and consistent reliability-oriented evaluation. This study compares sequence and digital-twin models using standardized prediction anchors and a common patient-level protocol; predictive performance was comparable, while model formulations showed distinct reliability benefits.

  • Problem

    Prior work provides limited unified multimodal prediction of both glycemic risks across horizons, and evidence about digital-twin reliability advantages over sequence predictors under common protocols remains limited.

    Source: §1 · p. 2

  • Method

    The study compares three sequence and digital-twin architectures on standardized longitudinal and multimodal prediction anchors under common patient-level validation, including gap-aware state updating.

    Source: §1 · pp. 2–3, State-Updating Digital Twin · p. 8

  • Results

    The three models showed comparable predictive performance, with distinct benefits in calibration, explainability, and longitudinal representation.

    Source: §5 · p. 23, §6 · p. 25

  • Takeaways & Limitations

    The findings support evaluating digital-twin formulations beyond predictive accuracy, while not treating architectural complexity alone as a guarantee of improved prediction.

    Source: §5 · p. 23

    The five-fold evaluation has limited power to detect small performance differences, so nonsignificant results do not establish model equivalence.

    Source: §5.6 · p. 25

Abstract

from arXiv · show

Anticipating short-term hypoglycemic and hyperglycemic risk is essential for proactive diabetes management. However, existing machine learning approaches often use inconsistent evaluation protocols and emphasize predictive performance over reliability. This study introduces an anchor-based benchmarking framework integrating longitudinal glucose sequences and multimodal patient information. Using the T1DiabetesGranada dataset, 5,503,441 standardized prediction anchors were constructed from 736 individuals with type 1 diabetes to evaluate hypoglycemia and hyperglycemia prediction at 30-, 60-, and 120-minute horizons. Three neural architectures, a Hybrid LSTM-Transformer, a Multimodal Digital Twin, and a State-Updating Digital Twin with gap-aware longitudinal state progression, were compared under identical patient-level validation protocols. Beyond predictive performance, the evaluation examined calibration, classification thresholds, explanation reliability, cross-model disagreement, and robustness to controlled perturbations. No statistically significant predictive differences were observed after Holm correction, although the models exhibited complementary reliability characteristics. The Multimodal Digital Twin achieved the lowest mean Brier score, while the State-Updating Digital Twin demonstrated the lowest mean expected calibration error (ECE-10). The Hybrid LSTM-Transformer showed strong explanation faithfulness and stability under its temporal attribution protocol. These findings highlight the importance of evaluating digital-twin models beyond predictive accuracy and emphasize the need for reliability assessment and external validation before clinical deployment.

1 Introduction

Prior glycemic-risk studies offer limited unified multimodal prediction across event types and horizons, and inconsistent evaluation complicates comparisons. This study proposes a standardized comparison of sequence and digital-twin models that evaluates reliability alongside predictive performance.

1 Introduction

  • Existing studies provide limited evidence for unified multimodal prediction of hypoglycemia and hyperglycemia across multiple horizons, and often underuse patient-specific information.Evaluation also commonly emphasizes discrimination over calibration, operating thresholds, explanation reliability, and controlled robustness assessment.Source: §1 · p. 2

2 Related Work

  • Heterogeneous datasets, preprocessing, prediction definitions, observation windows, and validation protocols make fair comparisons between glycemic prediction models difficult.The T1DiabetesGranada dataset combines longitudinal glucose with demographic, biochemical, and diagnostic information.Source: §1 · p. 2, §2 · p. 4

1 Introduction

  • This study compares a Hybrid LSTM–Transformer, Multimodal Digital Twin, and State-Updating Digital Twin under identical anchor construction, preprocessing, prediction targets, and patient-level validation.Its six-task benchmark covers both glycemic risks at 30-, 60-, and 120-minute horizons.Source: §1 · pp. 2–3

2 Related Work

  • Digital-twin evidence remains limited on reliability advantages over sequence predictors under common evaluation protocols, despite varied diabetes applications including augmentation, constrained modeling, forecasting, and intervention optimization [41] [3] [28].The reviewed studies use substantially different objectives and evaluation settings.Source: §2 · p. 5
  • The comparison examines whether multimodal fusion and longitudinal state updating offer measurable reliability differences rather than presuming digital-twin superiority.The evaluation includes calibration, explanation behavior, model disagreement, and robustness.Source: §1 · p. 3, §2 · p. 5

Position of This Study

The study positions its work as a controlled, anchor-based comparison of sequence and digital-twin models, with evaluation extending beyond discrimination to reliability behaviors. Its framework standardizes task construction and patient-level validation for six short-term glycemic-risk outcomes.

Position of This Study

  • The framework compares three model families using identical anchors, six-hour observation windows, prediction targets, preprocessing, and patient-level validation.It evaluates calibration, threshold behavior, explanation reliability, disagreement, error patterns, and controlled perturbations in addition to discrimination.Source: Position of This Study · p. 5

3 Materials and Methods

  • Three research questions address unified task formulation, controlled model comparison, and reliability evaluation beyond predictive discrimination.Together, they target four identified limitations in prior glycemic-prediction research.Source: §3 · pp. 5–6

3.1 Overview of the Proposed Anchor-Based Comparative Framework

  • The four-component framework covers anchor-based task construction, multimodal patient representation, model development, and reliability-aware evaluation for both risks at three horizons.It is intended for retrospective risk assessment, not autonomous insulin dosing, treatment recommendation, or direct clinical intervention.Source: §3.1 · p. 6

3.2 Dataset and Anchor-Based Task Formulation

  • T1DiabetesGranada records from 736 people with type 1 diabetes were converted into 5,503,441 anchors, each linking a six-hour glucose history and structured variables to six future-risk outcomes.Anchors were generated at a nominal one-hour stride and retained with at least five glucose measurements in the history and an observation at anchor time.Source: §3.2 · p. 6

3.3 Input Modalities and Preprocessing

  • Each glucose history contains 24 fifteen-minute steps with glucose, observation-mask, elapsed-time, and freshness channels; structured inputs include static, biochemical, and contextual groups.Missing modality information is retained through availability indicators, and imputation and normalization parameters are estimated from training folds.Source: §3.2 · p. 6, §3.3 · pp. 7–8

3.4 Model Families

  • The primary comparison includes a target-specific Hybrid LSTM–Transformer baseline and multitask Multimodal and State-Updating Digital Twins.The architectures share anchor representation, preprocessing, and patient-level validation; preliminary XGBoost and LightGBM models are not part of the primary comparison.Source: §3.4 · p. 8

Target-Specific Sequence Baseline.

The section describes a target-specific sequence baseline and a multitask multimodal digital twin. The former models each risk separately, while the latter fuses temporal and structured inputs for six task-specific outputs.

Target-Specific Sequence Baseline.

  • The Hybrid LSTM–Transformer combines bidirectional LSTM encoding with Transformer attention and trains a separate classification model for each of six prediction targets.It serves as the sequence-based reference for comparison with unified multitask digital-twin architectures.Source: Target-Specific Sequence Baseline · p. 8

Multitask Multimodal Digital Twin.

  • The Multimodal Digital Twin encodes temporal inputs recurrently and with a Transformer, processes static, biochemical, and contextual inputs through dedicated MLPs, then fuses representations.A multitask prediction layer maps the shared representation to six task-specific outputs.Source: Multitask Multimodal Digital Twin · p. 8

State-Updating Digital Twin.

The State-Updating Digital Twin carries a gap-adjusted latent state across ordered anchors and uses validity masking. The section also specifies patient-level testing, reliability analyses, and reporting boundaries.

State-Updating Digital Twin.

  • The State-Updating Digital Twin updates a longitudinal latent state from each fused anchor embedding and the gap-decayed previous state.Gap-dependent decay reduces older observations' influence, while validity masking prevents padded positions from affecting state updates or optimization.Source: State-Updating Digital Twin · p. 8

3.5 Training and Patient-Level Validation

  • Outer-fold testing holds out patients, but chronological anchor-level splitting means patients may appear in both training and validation; thresholds are selected on validation data.The authors caution that validation estimates are not independent patient-level generalization estimates and thresholds are not clinically validated alert thresholds; prospective and independent clinical validation are needed before deployment.Source: §3.5 · pp. 8–9

3.6 Predictive, Calibration, and Statistical Evaluation

  • The evaluation measures AUROC, AUPRC, F1-score, balanced accuracy, Brier score, ECE, and MCE; ECE and MCE use ten equal-width probability bins.Pairwise comparisons use five paired outer-fold estimates and Holm correction; nonsignificance is treated as insufficient evidence of superiority, not equivalence.Source: §3.6 · p. 9

3.7 Reliability, Explainability, and Behavioral Evaluation

  • Explanation analysis uses temporal and channel attribution for sequence models, modality attribution for the multimodal twin, and historical-anchor attribution for the state-updating twin.Faithfulness is measured by probability change after altering highly ranked components, while stability tracks dominant-component consistency across folds and targets; perturbation sensitivities are not causal or clinically actionable counterfactuals.Source: §3.7 · pp. 9–10

3.8 Reproducibility and Sofware Considerations

  • The report does not fully specify feature definitions, fold assignments, preprocessing settings, hyperparameters, and evaluation code needed for exact computational reproducibility.Computational efficiency also cannot be compared because training time, memory use, and inference latency were not measured consistently.Source: §3.8 · p. 10

4 Analysis of the Results

The analysis establishes a shared multimodal anchor representation for short-term hypoglycemia and hyperglycemia prediction, preserving patient coverage and irregular monitoring patterns. It defines the cohort, anchor eligibility, and outcome scope used for subsequent comparisons.

Cohort Coverage and Multimodal Availability.

  • 736 patients contributed longitudinal glucose records; biochemical data were available for 723 (98.23%), diagnostic data for 511 (69.43%), and both for 498.The harmonized cohort contained 22,649,679 glucose measurements, 87,470 biochemical records, and 1,733 diagnostic records.Source: Cohort Coverage and Multimodal… · pp. 10–11, Table 1 · p. 11
  • Monitoring volume varied widely, with a median of 24,993.5 glucose measurements per patient (IQR 10,668.75–45,684.75; range 401–137,275), motivating density and missingness descriptors.Source: Cohort Coverage and Multimodal… · p. 11

Anchor-Level Dataset Construction.

  • 5,503,441 anchors from 736 patients represented six-hour glucose histories and six future-risk outcomes across 30-, 60-, and 120-minute horizons.Anchors required at least five glucose measurements in the retrospective window and an observed value at the anchor time; the representation included 43 variables.Source: Anchor-Level Dataset Construction · p. 11
  • Patient-level folds kept each patient’s anchors together, preventing overlapping histories and future intervals from crossing development and held-out test sets.The five outer folds contained 147–148 patients and approximately 1.03–1.18 million held-out anchors each; the anchors are repeated longitudinal observations, not independent samples.Source: Anchor-Level Dataset Construction · p. 11

Temporal Irregularity and Intermeasurement Gaps.

Outcome Distribution Across Prediction Horizons.

  • Hypoglycemia prevalence rose from 5.76% at 30 minutes to 11.07% at 120 minutes, while hyperglycemia rose from 37.98% to 50.23%.Hypoglycemia was more imbalanced at every horizon, informing use of AUPRC, F1, balanced accuracy, calibration, and threshold-dependent evaluation.Source: Outcome Distribution Across Prediction… · p. 13, Table 3 · p. 13

Overall and Horizon-Dependent Performance.

Across the six targets, model-family differences were small relative to fold variability, and no Holm-corrected predictive advantage was established. Numerical leaders varied by target and horizon, while predictive performance declined as horizons lengthened.

Overall and Horizon-Dependent Performance.

  • Across five patient-level folds, all three architectures performed best on 30-minute outcomes, with progressive reductions at 60 and 120 minutes and differences generally small relative to fold variability.The spread between highest and lowest mean F1 was approximately 0.005 for 30-minute hypoglycemia and less than 0.001 for 60-minute hyperglycemia.Source: Overall and Horizon-Dependent… · pp. 13–14, Table 4 · p. 14
  • Across all three model families, mean F1 declined with horizon: hypoglycemia fell from approximately 0.78–0.79 at 30 minutes to 0.60–0.61 at 120 minutes, and hyperglycemia from approximately 0.932 to 0.832–0.834.The 120-minute hypoglycemia outcome was the most difficult task; Fig. 3 shows a larger reduction for hypoglycemia than hyperglycemia.Source: Overall and Horizon-Dependent… · p. 14, Figure 3 · p. 15

Target-Specific Rankings and Statistical Comparison.

  • The Multimodal Digital Twin achieved the highest mean values for the 30-minute outcomes, while no architecture achieved the highest mean across all outcomes.Source: Target-Specific Rankings and… · p. 14
  • No model family had a Holm-corrected significant advantage on any of six targets; inference was limited by only five paired outer-fold estimates.The results do not establish statistical equivalence or similarity between architectures.Source: Target-Specific Rankings and… · pp. 15–16
  • The two digital-twin models had numerically similar predictive performance to the Hybrid LSTM–Transformer while estimating all six risks in a unified multitask framework.The study presents them as competitive alternatives, not as predictively superior or equivalent models.Source: Target-Specific Rankings and… · p. 16

4.3 RQ3: Reliability, Explainability, and Deployment-Oriented Behavior

  • The reliability analyses characterize complementary model properties rather than establishing architectural superiority or clinical deployment readiness.Source: §4.3 · p. 16

Calibration and Threshold-Dependent Performance.

Probabilistic reliability differed across models, with Brier score and ECE-10 favoring different digital twins; threshold-based rankings also varied by target and horizon. Explanation patterns differed by architecture and were evaluated using model-specific units.

Calibration and Threshold-Dependent Performance.

  • The Multimodal Digital Twin had the lowest mean Brier score (0.0702), while the State-Updating Digital Twin had the lowest mean ECE-10 (0.0611).The metrics capture related but distinct aspects: Brier score combines calibration and discrimination, while ECE-10 measures probability-frequency discrepancy using 10 bins.Source: Calibration and Threshold-Dependent… · p. 16, Table 6 · p. 16
  • At 120-minute hypoglycemia, ECE-10 was 0.1845 for the Hybrid LSTM–Transformer, 0.1715 for the Multimodal Digital Twin, and 0.1633 for the State-Updating Digital Twin.Across hyperglycemia targets and model families, ECE-10 remained below 0.026; descriptive differences do not establish significant or clinically meaningful calibration superiority.Source: Calibration and Threshold-Dependent… · pp. 16–17
  • Thresholds were selected to maximize validation F1 and fixed before patient-disjoint testing, so the resulting operating points are not clinically optimized alert settings.Prospective evaluation is needed to set thresholds against sensitivity, specificity, false-alert burden, and consequences of missed events.Source: Calibration and Threshold-Dependent… · p. 17
  • Balanced-accuracy leaders varied across targets: the Hybrid LSTM–Transformer led hypoglycemia at 30 minutes (0.8887) and 120 minutes (0.7699), while the State-Updating Digital Twin led at 60 minutes (0.8285).For hyperglycemia, the Multimodal Digital Twin led at 30 minutes (0.9440), the State-Updating Digital Twin at 60 minutes (0.9047), and the Hybrid LSTM–Transformer at 120 minutes (0.8405).Source: Calibration and Threshold-Dependent… · p. 17, Table 7 · p. 17

Global and Model-Specific Explanation Structure.

  • Explanation units differed across architectures, so their attribution values do not share a measurement scale and should not be directly ranked as explanation quality.Source: Global and Model-Specific Explanation… · pp. 17–18
  • Hybrid LSTM–Transformer attribution concentrated near recent positions, while some longer-horizon outcomes spread importance more broadly; observation-mask, elapsed-time, and freshness channels contributed non-zero amounts.Source: Global and Model-Specific Explanation… · p. 18, Figure 4 · p. 18
  • The Multimodal Digital Twin’s temporal branch dominated most hyperglycemia and imminent hypoglycemia outcomes, whereas contextual information contributed more to some longer-horizon hypoglycemia targets.State-Updating Digital Twin attribution spanned multiple historical anchors, with dominant anchors varying across folds and targets.Source: Global and Model-Specific Explanation… · p. 18

Explanation Faithfulness and Stability.

Faithfulness and stability results describe distinct model-specific explanation protocols rather than a common ranking. The architectures agreed on most binary decisions, though disagreement increased with horizon and subgroup findings remained exploratory.

Explanation Faithfulness and Stability.

  • These explanation measures use different units and definitions, so their numerical values do not establish a common ranking of explanation quality.The smaller individual-anchor effects and lower dominant-anchor consistency for the State-Updating Digital Twin may reflect information distributed across recurrent updates, but this remains descriptive.Source: Explanation Faithfulness and Stability · pp. 19–20
  • Under its temporal protocol, the Hybrid LSTM–Transformer had a mean perturbation effect of 0.3105 and dominant-component stability of 1.0000.The Multimodal Digital Twin had modality-level effect 0.1722 and stability 0.9667; the State-Updating Digital Twin had individual-anchor effect 0.0074 and stability 0.3667.Source: Explanation Faithfulness and Stability · p. 19, Table 9 · p. 20

Cross-Model Disagreement and Error Paterns.

  • Binary disagreement rose with prediction horizon, reaching 6.60% for 120-minute hyperglycemia; the largest mean probability spread was 0.1080 for 120-minute hypoglycemia.Mean disagreement across horizons was 2.37% for hypoglycemia and 3.72% for hyperglycemia; these measures describe prediction variation, not errors or calibrated uncertainty.Source: Cross-Model Disagreement and Error… · p. 20, Table 10 · p. 20
  • Exploratory subgroup analyses found no consistent error-rate elevation across age groups, while patients without biochemical records showed slightly higher error rates.The analyses are not a formal assessment of algorithmic fairness.Source: Cross-Model Disagreement and Error… · p. 20
  • Controlled perturbations examined sensitivity to recent glucose, missingness, structured variables, and historical anchors, not causal responses or physiological effects.Source: Cross-Model Disagreement and Error… · p. 20

Controlled Perturbation and State-History Behavior.

The models showed complementary reliability and architectural profiles rather than a clear overall winner. Perturbation analyses characterized model sensitivity, while deployment-oriented findings and limitations underscore the need for further validation.

Controlled Perturbation and State-History Behavior.

  • Lowering recent glucose increased predicted hypoglycemia risk, while raising it increased predicted hyperglycemia risk across architectures.Mean absolute probability-change ranges were 0.401–0.466 and 0.417–0.445 for the Hybrid LSTM–Transformer, 0.661–0.818 and 0.484–0.641 for the Multimodal Digital Twin, and 0.583–0.751 and 0.480–0.586 for the State-Updating Digital Twin, respectively.Source: Figure 6 · p. 21, Controlled Perturbation and… · p. 21
  • Removing individual historical anchors from the State-Updating Digital Twin produced smaller, non-monotonic changes, consistent with information distributed across recurrent state updates.Alternative explanations remain possible, and the result does not establish physiological validity of the learned state.Source: Controlled Perturbation and… · p. 21

Deployment-Oriented Synthesis.

  • The Multimodal Digital Twin combines joint six-task prediction with the lowest mean Brier score, while the State-Updating Digital Twin combines joint prediction, longitudinal updating, and the lowest mean ECE-10.The Hybrid LSTM–Transformer showed temporal-component perturbation effects and high dominant-component consistency under its model-specific explanation protocol, but required separate models for six targets.Source: Deployment-Oriented Synthesis · pp. 21–22, Table 11 · p. 22

5.1 Principal Findings

  • The three model families had broadly similar predictive performance but distinct reliability profiles, with no statistically established architectural superiority.The Multimodal Digital Twin had the lowest mean Brier score, while the State-Updating Digital Twin had the lowest mean ECE-10.Source: Deployment-Oriented Synthesis · p. 22, §5.1 · p. 23

5.2 Prediction Horizon and Model Complexity

  • Performance declined from 30 to 120 minutes across models, especially for hypoglycemia; unavailable insulin, food, activity, and behavioral data constrained long-horizon prediction.The digital-twin models did not show a consistent long-horizon advantage, although their components may support probability estimates, patient-specific representations, and continuous monitoring.Source: §5.2 · p. 23

5.3 Role of Multimodal Fusion and State Updating

  • Recent temporal information remained the Multimodal Digital Twin’s dominant contributor, while biochemical and static variables may chiefly provide contextual adjustment and refined calibration.The State-Updating Digital Twin uses gap-aware transitions and validity masking; its primary contribution is continuous patient-state representation under irregular observations, not guaranteed predictive gains.Source: §5.3 · pp. 23–24

5.4 Reliability, Explainability, and Behavioral Analysis

  • Reliability analyses found architecture-specific explanation behavior, greater cross-model disagreement at longer horizons, and sensitivity to perturbations; explanation protocols were not directly comparable.The Hybrid LSTM–Transformer had the strongest temporal attribution consistency, while the Multimodal Digital Twin had stable modality-level interpretation; future explanation analysis should examine trajectories, not only single components.Source: Deployment-Oriented Synthesis · p. 22, §5.4 · p. 24

5.6 Limitations and Future Work

  • The study supports further deployment-oriented investigation, not clinical readiness: evaluation used one retrospective dataset, and external and prospective validation remain necessary.The digital twins are data-driven operational representations, not physiological simulators or automated treatment systems; the five-fold comparisons also had limited power to detect small differences.Source: Deployment-Oriented Synthesis · p. 22, §5.6 · pp. 24–25

6 Conclusion

The study presents an anchor-based comparison of three model families for short-term glycemic-risk prediction. Their predictive performance was comparable, but reliability profiles differed; no statistically significant superiority was found, and external validation remains necessary.

6 Conclusion

  • The study transformed T1DiabetesGranada records into over 5.5 million standardized anchors from 736 patients for hypoglycemia and hyperglycemia prediction at 30-, 60-, and 120-minute horizons.The Hybrid LSTM–Transformer, Multimodal Digital Twin, and State-Updating Digital Twin were compared under patient-level validation.Source: §6 · p. 25
  • No statistically significant predictive superiority remained after Holm correction, while the models differed in calibration, explanation behavior, disagreement, and perturbation responses.The Multimodal Digital Twin had the lowest mean Brier score and the State-Updating Digital Twin the lowest mean ECE-10.Source: §6 · p. 25
  • The findings support multidimensional reliability evaluation, but external validation and prospective workflow assessment are needed before clinical translation.The State-Updating Digital Twin was selected for further investigation, not as a universally superior or deployment-ready model.Source: §6 · p. 25

A Supplementary Material

The supplementary material documents related-study comparisons, model configuration, dataset characteristics, and additional predictive-performance analyses. It reports broad cross-model patterns while cautioning that heterogeneous prior studies and descriptive rankings do not establish direct or statistically significant comparisons.

A.1 Comparison of Related Studies

  • The related-study table compares prediction objectives, modalities, modeling strategies, and evaluation protocols, but heterogeneous study settings limit direct performance comparisons.Source: §A.1 · p. 28, Table S1 · p. 28

A.2 Supplementary Model Configuration

  • The supplementary configuration table summarizes architectures and training procedures, with preprocessing and model selection confined to patient-level training and validation partitions.Source: §A.2 · p. 29, Table S2 · p. 29

A.3 Additional Results for RQ1: Anchor-Based Multimodal Representation

  • The anchor-based dataset analyses describe modality availability, fold distribution, glucose measurement gaps, and event prevalence.These supplementary analyses characterize the dataset representation used in the evaluation.Source: §A.3 · p. 29

A.3.1 Multimodal Coverage.

  • Glucose records were available for all 736 patients, while biochemical and diagnostic information had lower coverage; missing modalities were retained with availability indicators.Source: §A.3.1 · p. 29, Figure S1 · p. 30

A.3.2 Patient-Level Fold Distribution.

  • All of each patient’s anchors were assigned to one outer fold, and the figure displays held-out anchor distributions across five patient-level folds.Source: §A.3.2 · p. 29, Figure S2 · p. 30

A.3.3 Intermeasurement Gap Statistics.

  • Most consecutive glucose measurements followed the nominal 15-minute interval, but occasional prolonged gaps made the intermeasurement-gap distribution highly skewed.Source: §A.3.3 · p. 29, Table S3 · p. 31

A.3.4 Outcome Distribution.

  • Hypoglycemia events were substantially less prevalent than hyperglycemia events across prediction horizons, indicating different imbalance levels among the six tasks.Source: §A.3.4 · p. 29, Figure S3 · p. 31

A.4 Additional Results for RQ2: Predictive Performance

  • Supplementary AUPRC and F1-score figures show broadly similar model patterns, with performance generally declining at longer horizons, particularly for hypoglycemia.The detailed estimates and statistical comparisons are reported in the main manuscript.Source: §A.4.1 · p. 31, §A.4 · p. 31

A.4.2 Numerical Model Rankings.

Supplementary analyses show that numerical rankings vary across targets and that explanation, disagreement, subgroup-error, and perturbation results are descriptive or protocol-specific. They offer additional views of model behavior, not a common ranking of architectures or evidence of clinical validity.

A.4.2 Numerical Model Rankings.

  • No model family consistently achieved the highest numerical mean performance across targets and metrics, and the rankings do not establish statistically significant differences.Figures S4 and S5 show mean AUPRC and F1-score, while Figure S6 presents numerical rankings.Source: §A.4.2 · p. 31, Figure S4 · p. 32, Figure S5 · p. 32

A.5.1 Local Explanation Examples.

  • Representative SHAP examples cover true-positive, false-positive, false-negative, and true-negative hyperglycemia cases, for qualitative rather than population-level inspection.Source: §A.5.1 · pp. 32–33, Figure S7 · p. 34

A.5.2 Explanation Faithfulness and Stability.

  • Explanation faithfulness and stability are reported under model-specific protocols, so their values should not be ranked directly across architectures.Source: §A.5.2 · p. 33, Figure S8 · p. 35

A.5.3 Cross-Model Disagreement and Error Stratification.

A.5.4 Recent-Glucose Perturbation Analysis.

  • Recent-glucose perturbation results report ranges of mean probability changes across horizons and characterize directional sensitivity, not validated physiological counterfactual effects.Source: §A.5.4 · p. 33, Table S4 · p. 34

A.5.5 State-History Perturbation Analysis.

  • The State-Updating Digital Twin’s response to longitudinal state-history perturbations provides descriptive evidence about sensitivity to historical information.Source: §A.5.5 · p. 33, Figure S11 · p. 36
Loading 2610.36363v1…