Food-i-Sense Analytics: Integrating AI Into Continuous Glucose Monitoring Data Analysis for Precision Nutrition. (Food_i Sense)

The Food_iSense Analytics (FiS) study is an observational, cross-sectional protocol designed to advance precision nutrition through the integration of continuous glucose monitoring (CGM), artificial intelligence (AI), and comprehensive phenotyping. The project builds upon preliminary work using data from the ENSATI and TEMPUS studies, where the research team developed GLIA, an AI-driven algorithm capable of generating individualized glucotypes-patterns of glycemic behavior that reflect the dynamic response of glucose to daily living conditions and meal intake.

Scientific Background and Rationale Although individuals without diagnosed diabetes may exhibit blood glucose values within standard reference intervals, the shape, duration, and variability of glucose excursions reflect underlying physiological regulation and may reveal early signs of metabolic dysfunction. Research has demonstrated high inter-individual variability in glycemic responses to identical meals, suggesting that dietary guidelines must move toward personalization.

The introduction of CGM devices (FreeStyle Libre 3) allows for high-resolution temporal data capturing minute-to-minute changes in interstitial glucose. However, traditional CGM metrics (mean glucose, time in range, coefficient of variation) do not sufficiently capture the full complexity of glucose dynamics.

GLIA addresses this limitation by extracting multidimensional features that quantify:

Peak morphology: slope, amplitude, recovery time, decay kinetics. Variability features: short- and long-term variability indexes, glycemic volatility, post-prandial oscillation density.

Chrononutrition-related features: differences in glycemic control across circadian windows (morning/afternoon/evening), alignment with habitual eating patterns.

Derived metrics: composite indexes generated via principal component analysis (PCA) and clustering.

Using machine learning and unsupervised clustering with bootstrapping, GLIA identifies stable glucose response phenotypes. These glucotypes are then examined in relation to health indicators, dietary patterns, and predictive models of individual glycemic responses.

Study Structure and Workflow Overview

The study consists of three in-person visits across approximately 21 days, during which participants undergo:

Initial assessment (demographics, anthropometrics, medical history, baseline health measures).

Continuous 14-day CGM period with detailed dietary monitoring using:

Two structured 3-day dietary records Automated AI-based food image recognition Mediterranean diet and ultraprocessed food questionnaires Physical activity questionnaires

Final assessment including biological sample collection (fasting blood and first-morning urine), updated anthropometry, and final quality check of all dietary records.

The final dataset incorporates more than 140 nutritional variables per food item, combined with high-resolution glucose time-series data, clinical phenotype, and multiple molecular biomarkers.

Registry-Related Quality Procedures and Data Governance Although this study is not a patient registry in the classical sense, the research team implements registry-grade data management procedures due to the scale, multidimensionality, and long-term value of the dataset. The following subsections reflect the registry-quality framework.

1. Quality Assurance Plan

   A comprehensive quality assurance (QA) plan governs all activities from recruitment to data analysis. Key components include:

   Standardized training of all personnel (dietitians, research nurses, data managers).

   Calibration schedules for anthropometric devices (stadiometer, scale, bioimpedance instruments) and for blood pressure monitors.

   Daily consistency checks of CGM data uploads. Protocol deviation logs documenting missing measurements, device issues, and participant non-compliance.

   Internal monthly audits performed by the IMDEA Quality Office to verify protocol adherence.

   Independent external audit capability is maintained, though audits are not routinely scheduled unless required by funders or ethics committees.

2. Data Validation and Automated Data Checks

   Incoming data are processed through a multi-stage validation pipeline:

   1. Range and plausibility checks Automatic filtering flags values outside expected biological ranges (e.g., impossible BMI, extreme macronutrient percentages, duplicate CGM timestamps).

   2. Internal consistency checks

      Cross-field validation identifies inconsistencies, such as:

      caloric intake mismatching macronutrient totals, dietary patterns incompatible with photographed meals, anthropometric values inconsistent across visits.

   3. Technical consistency

   CGM streams are checked for:

   signal dropouts longer than 15 minutes, abrupt shifts indicating sensor displacement, unrealistic glucose kinetics (rise/decay rates).

   Problematic sections are annotated but not deleted, preserving data integrity for sensitivity analyses.

3. Source Data Verification (SDV)

   To ensure data accuracy and representativeness:

   Dietary records are cross-validated against food photographs and questionnaire responses.

   Medical history and medication data are verified against documents participants bring to Visit 2 (e.g., lab reports ≤ 6 months old).

   Blood pressure and anthropometry undergo dual measurement with two assessors performing random checks on 10% of sessions.

   CGM data are compared with participant logs describing sensor issues, physical activity peaks, and atypical meals.

   All verification steps follow Good Clinical Practice (GCP) documentation practices.

4. Data Dictionary and Variable Coding

The study uses an extensive data dictionary organized into modules:

Sociodemographic module

Definitions, coding schemes (e.g., ISCED for education), universe and skip patterns.

Clinical phenotype module

Standard coding for medical conditions using ICD-10 when applicable. Definitions of metabolic syndrome markers and derived variables.

Anthropometry and body composition

Measurement rules, rounding conventions, and device-specific calibrations.

CGM module

Time-series format, sampling interval, sensor metadata, derived features, preprocessing rules.

Dietary variables

Food items mapped to validated food composition databases for Spain. Nutrient coding includes standardized units and upper/lower expected ranges.

Molecular biomarkers

Units, assay platforms, storage conditions, and laboratory reference values.

The dictionary includes more than 500 entries and is version-controlled.

SOPs govern all operational and analytic tasks, including:

Recruitment and informed consent CGM application, monitoring, and removal Anthropometric and BIA measurement Blood/urine sample collection, processing, aliquoting, and storage Data entry and double-entry verification for paper forms Food image capture guidelines and data upload protocol Use of the AI food-recognition system and manual override procedures Documentation of adverse events related to CGM or blood collection Dataset freeze procedures and change control for analysis scripts Secure storage, encryption, and pseudonymization workflows

SOPs are reviewed periodically and updated as needed.

6. Sample Size Assessment

Using synthetic data generated during algorithm development, investigators determined that approximately 392 complete datasets provide adequate clustering stability:

Maximization of the Silhouette score (~0.67 at n=392). Near-minimal Davies-Bouldin index at ~392-588 samples. Robust PCA component stability at ≥350 participants.

Accounting for a 20% potential data-loss rate, 471 participants are targeted for recruitment.

7. Missing Data Plan Missingness may arise from incomplete dietary records, sensor failures, lost images, or participant withdrawal.

The plan includes:

Classification of missingness as MCAR, MAR, or MNAR. Short-gap interpolation for CGM data gaps ≤ 20 minutes using cubic spline or Kalman smoothing.

Exclusion of unusable segments (e.g., sensor warm-up periods). Multiple imputation (MI) for covariates such as weight or blood pressure if only one visit yields valid measures.

Complete-case sensitivity analyses to confirm robustness of imputed models.

Retention strategies minimize missing data, but high technical data density ensures analytic viability even with partial loss.

8. Statistical Analysis Plan (SAP) The SAP covers both algorithm validation and hypothesis-generating analyses. 8.1 Preprocessing

CGM normalization to individual baselines. PCA for dimensionality reduction. Unsupervised clustering (k-means and hierarchical) with bootstrapping. Feature engineering for peak characteristics and chrononutrition variables.

8.2 Glucotype generation and validation

Internal validation with Silhouette, Calinski-Harabasz, and Davies-Bouldin metrics.

Stability assessment through subsampling and perturbation tests.

8.3 Association analyses

ANOVA/Kruskal-Wallis for between-glucotype comparisons. Logistic and linear regression adjusting for key confounders. MANOVA for overall cardiometabolic profiles.

8.4 Dietary modelling

Multiple Correspondence Analysis (MCA) for dietary pattern extraction. Regression models linking nutrient intake to glucotype. Machine learning models (Random Forest, SVM, XGBoost) to predict meal-level glucose responses.

8.5 Predictive simulations Agent-based modelling is used to simulate hypothetical diet changes and predict personalized glucose responses under various nutritional scenarios.

8.6 Software and reproducibility Analyses are performed in Python and R, with scripts managed through version control, containerization (Docker), and reproducibility logs.

Conclusion The FiS protocol integrates advanced CGM analytics, AI-driven phenotyping, and detailed dietary and biochemical profiling to map the diversity of glucose regulation in adults without diabetes. The study applies rigorous registry-level quality procedures, including automated data validation, structured data dictionaries, SOPs, and a comprehensive statistical analysis plan. The resulting dataset and validated algorithm will support future work in personalized nutrition, early risk detection, and tailored dietary interventions.