Continuous glucose monitoring evidence of celiac disease in type 1 diabetes

Study cohort

We performed a retrospective analysis of a matched cohort drawn from a study cohort of children aged 2–21 years followed in the Boston Children’s Hospital (BCH) T1D program. This study was conducted in accordance with the Declaration of Helsinki. This study protocol was reviewed and received Ethical Approval from the Boston Children’s Hospital Institutional Review Board, approval number IRB-P00032591 on July 12, 2019. The need for written informed consent was waived for this retrospective study by the Institutional Review Board.

The study cohort included individuals diagnosed with T1D from January 1, 2015 (beginning of BCH CGM database) to December 30, 2022. Eligible subjects did not have a diagnosis of CD at the time of diabetes diagnosis, underwent celiac serologies <12 months following diabetes diagnosis, used Dexcom CGM (DexCom, Inc., San Diego, CA, USA) in the month prior to serologies and shared CGM data with BCH via the cloud-based Dexcom Clarity Portal. We excluded children using automated insulin delivery technology including predictive low glucose suspend features and/or with gastrointestinal comorbidities (e.g. inflammatory bowel disease).

Exposure definition

The exposure of interest was undiagnosed CD. The exposure window was the 30 days prior to a documented celiac antibody testing date (i.e., index date) (Figure 1A). The index date for one individual was the date of follow-up serologies after initial positive serologies, as they started CGM in the interim but were advised to continue consuming gluten with documented gluten intake between serologies. Confirmed CD (CD+) was defined using the gold standard, i.e., duodenal histologic findings of crypt hyperplasia, villous atrophy and intraepithelial lymphocytosis plus elevated tissue transglutaminase IgA [23]. Those unexposed to CD (CD−) were defined as those with normal CD serologies.

Figure 1:

Study design and prandial outcome definitions. (A) Schematic of cohort characteristics and timing of celiac screening test (measurement of exposure) relative to the window for CGM outcome analysis in the matched cohort. (B) Schematic depicting definitions of prandial glucose outcomes. Meals were inferred by first identifying a glucose peak with a subsequent glucose trough. The meal glucose start was defined as an inflection from downtrending to uptrending glucose preceding the peak. Meal glucose rise (rise) was calculated as peak minus start and rise velocity was calculated as rise divided by time from start to peak. Meal glucose fall (fall) was calculated as the absolute value of trough minus peak and fall velocity was calculated as fall divided by time from peak to trough. CD−, no celiac disease; CD+, undiagnosed celiac disease; CGM, continuous glucose monitoring; AUC, area under curve. Created in BioRender. Ruiz, J. (2025) https://BioRender.com/2x4rz0m.

Matched cohort criteria

Demographic and clinical data were collected via retrospective chart review. Potential subjects were only eligible for matching if their CGM data met sufficiency criteria. CD+ children were randomly matched with up to four CD−, using five clinical covariates to adjust for confounding from clinically-relevant physiologic states. Covariates were dichotomized to perform exact matching: sex, pubertal age (males ≥11.5 years [median age to attain testicular volume ≥4 mL [24]]; females ≥10 years [median age to observe Tanner stage II breast development in girls with BMI 10–84th percentile [25]]), longer diabetes duration (≥6 months), overweight (BMI z-score ≥1.0, corresponding to the overweight definition ≥85th percentile [5]) and above target hemoglobin HbA1c ≥7.5 % (accepted target for children who do not utilize automated insulin delivery [5]). Subjects with identical combinations of matching covariates were combined into the same strata for more statistically efficient analyses. Combined race and ethnicity categories were recorded but not used for matching. Individuals with Hispanic race or ethnicity recorded were categorized as Hispanic and all other race/ethnicity categories represent non-Hispanic individuals.

Prandial and summative glucose outcomes

Our prandial outcomes quantify prandial dynamics, identified as periods of transient glucose rise, peak (maximal excursion), fall and trough (post-prandial nadir) most likely to represent meals. The primary outcome was meal trough glucose, selected to operationalize the evaluation of post-prandial hypoglycemia. Secondary meal outcomes (Figure 1B) included starting glucose, peak glucose, meal glucose rise (peak minus start), rise velocity (rise divided by time from start to peak), meal glucose fall (absolute value of trough minus peak) and fall velocity (fall divided by time from peak to trough). Area under curve (AUC) was calculated from meal start to trough using the trapezoidal rule (Python sklearn.metrics.auc). All continuous meal outcomes were log-transformed due to skewness. Meal hypoglycemia was also assessed as a dichotomous outcome, defined as meal trough glucose <70 mg/dL. Meal outcomes were analyzed for all meals and then repeated analyzing a subset restricted to the first meal per day, to reduce meal-to-meal heterogeneity.

Summative CGM outcomes were mean glucose, glucose management indicator (GMI=3.31+[0.02392 × mean glucose]), standard deviation and percent time in the following ranges: 70–180 mg/dL (time in range), 70–140 mg/dL (time in tight range), >180 mg/dL, >250 mg/dL, <70 mg/dL and <54 mg/dL. Summative CGM outcomes were computed for the matched cohort using their raw, 30-day CGM data without interpolation or smoothing (as might be available in a clinical setting), via R package iGlu (4.0.1). These CGM metrics were analyzed for the full 24 h per day, then repeated limited to typical waking hours (6–12AM [26]), a proxy for the mealtime period.

Meal annotation algorithm

Broadly, the meal annotation algorithm takes smoothed CGM data; identifies meals by searching for a trio of meal peak, trough, then preceding starting glucose; and extracts key CGM interstitial glucose (herein glucose) values. We extracted the CGM glucose values from the 30-day windows prior to each index date from the BCH Dexcom Clarity dataset (downloaded November 11, 2023) using R (version 3.6.2) on the BCH Enkefalos-v2 (E2) high-performance computing cluster (Linux operating system). Meal annotation was conducted using Python (version 3.9.13) with packages scipy (1.9.1), numpy (1.21.5), pandas (1.4.4) and sklearn (1.0.2).

First, for more reliable meal detection while minimizing non-physiologic fluctuations, missing glucose data were imputed using linear interpolation and high-frequency fluctuations >0.0005 Hz [27] were smoothed using a low pass filter. Then, peaks were identified during typical waking hours (6–12 AM) using scipy.signal.find_peaks to search the glucose data for features matching the following parameters: peak height ≥100 mg/dL, ≥2 h between peaks, rise ≥50 mg/dL above baseline and with elevated glucose ≥30 min duration. Similarly, troughs were identified as follows: trough glucose ≥15 min to ≤4 h after peak, fall ≥15 mg/dL below baseline for ≥15 min duration. Troughs were restricted to 6–3 AM to allow for a trough following a late-night meal while filtering out overnight fasting lows. The lowest trough glucose following each peak was selected. Last, the start of each meal was defined as the nearest glucose inflection from downtrending to uptrending glucose during waking hours ≤4 h before meal peak and with ≥50 mg/dL glucose increase from meal start to peak.

Data sufficiency for inclusion

To be eligible for matching, individuals were required to have at least 14 days of CGM data with ≥0.5 qualifying meals per day on average. Meals were disqualified if glucose values (prior to imputation) were missing ≤15 min before or after meal start, peak or trough; or if, from meal start to trough, more than 30 % of expected glucose values were missing or the glucose rate of change was >6 mg/dL/min [28].

Statistical analyses

Patient characteristics were compared between groups using Wilcoxon rank-sum or Fisher’s exact tests. Linear mixed models were used to compare meal outcomes between CD groups in the matched cohort with a fixed effect for matched strata (adjusting for confounding from covariates) and a random effect for subject (accounting for correlation between repeated measures from individuals). For the log-transformed outcomes, effect estimates represent the percent change in the CD+ relative to the CD− group, calculated by exponentiating the coefficient (beta), subtracting 1 and multiplying by 100. Mixed effects logistic regression was used to compare meal hypoglycemia between CD groups, using the same fixed and random effects. Rank-based van Elteren tests were used to compare summative CGM metrics between the CD groups, adjusting for matched strata. All prandial and summative CGM outcome calculations and statistical analyses were conducted using R (version 4.3.2) using two-sided 0.05 level tests.

Data availability and reporting guidelines

The datasets and code generated and/or analyzed in this study are available from the corresponding author upon reasonable request. We used the STROBE reporting guideline [29] to draft this manuscript, and the STROBE reporting checklist [30] when editing.

Study cohort characteristics

During the study period, 1,531 individuals were diagnosed with T1D and 921 had recorded CD serologies <1 year of diabetes diagnosis. Of these, 850 were age 2–21 years with recently recorded BMI and HbA1c required for matching. Three hundred individuals also had CGM data available in the 30-day exposure window prior to CD serologies, and 288 met data sufficiency criteria, leaving 16 CD+ and 272 CD− individuals for matching.

The matched cohort comprised 16 CD+ and 60 CD− individuals: 13 CD+ matched exactly with four CD− individuals, two matched with three CD− individuals and one matched with two CD− individuals. The combined matched cohort was 70 % female and 75 % non-Hispanic White with median age 11.1 years (interquartile range 8.9–13.3), BMI z-score 0.23 (−0.15 to 0.61), diabetes duration 3.2 months (1.5–5.0) and HbA1c 7.3 % (6.3–8.4). There were 2,067 days of CGM data analyzed (median 30 days/subject [25.8–30.0]) and 3,659 meals analyzed (median 48.5 meals/subject [34.5–62.5]). There were no statistically significant differences with respect to these characteristics between CD+ and CD− groups (Table 1).

Prandial glucose dynamics

Our primary outcome, meal trough glucose, broadly operationalized the hypothesis that undiagnosed CD is associated with post-prandial hypoglycemia, evaluating for any change in trough glucose even if not meeting criteria for hypoglycemia. Using a linear mixed model, there was no significant difference in meal trough glucose between the as-yet undiagnosed CD+ group and the unexposed CD−group (−8.2 % effect estimate, 95 % CI, −20.1–5.5 %, p=0.23; Table 2). Undiagnosed CD was also not associated with meal hypoglycemia, defined as meal trough glucose <70 mg/dL (17.1 % CD+ vs. 13.9 % CD−meals, odds ratio 1.32, 95 % CI, 0.72 to 2.45, p=0.36).

We then assessed for other changes in prandial glucose dynamics in the setting of T1D and undiagnosed CD. The meal glucose rise (increase from start to peak) was significantly lower among CD+ compared to CD− (−8.9 % effect estimate, 95 % CI, −14.9–−2.5 %, p=0.009). The meal glucose fall (decrease from peak to trough) was also lower in the CD+ group, though not reaching statistical significance (−6.7 % effect estimate, 95 % CI, −13.3–0.5 %, p=0.07, Figure 2A). Similarly, other related meal outcomes including peak glucose, rise velocity, fall velocity and AUC had negative effect estimates but did not reach statistical significance (Table 2).

Figure 2:

Blunted meal glucose excursions in youth with undiagnosed celiac disease and type 1 diabetes. Data from the CD−group are shown in orange and data from the CD+ group are shown in blue. (A–B) comparison of distribution of meal glucose rise, fall and trough for (A) all meals and (B) a subset consisting of the first meal per day (first meals). Rise was calculated as meal glucose peak minus start and fall was calculated as the absolute value of meal glucose trough minus peak. Boxes represent the interquartile range (IQR, 25th to 75th percentile). Horizontal lines inside the boxes denote median values, whiskers outside the boxes extend to 1.5 × IQR and dots denote outliers outside 1.5 × IQR. (C–D) mean of mean meal glucose per individual (solid circles) and 95 % confidence intervals (shaded regions) for (C) all meals and (D) the subset of first meals. Timepoint-wise mean glucose was calculated at each 5-min increment following meal start, first by individual, then across the CD groups. Meals were of variable durations and the number of individuals with data at each time increment is represented by circle size. To generate representative illustrations of meal glucose patterns, the longest 5 % of meals per individual were excluded, then the longest 5 % of all remaining meals were excluded. CGM, continuous glucose monitoring; CD−, no celiac disease; CD+, undiagnosed celiac disease; *p<0.05, **p<0.01. Created in BioRender. Ruiz, J. (2025) https://BioRender.com/ulrxr00.

We then repeated the analysis with CGM meal data limited to the first meal of each day to reduce the impact of heterogeneity from inter-prandial glycemia and potential diurnal variation in insulin sensitivity and/or meal glycemic index. Within this subset of first meals, undiagnosed CD was again associated with a significantly reduced meal glucose rise (−10.5 % effect estimate, 95 % CI, −17.8–−2.4 %, p=0.01) and a significantly reduced meal glucose fall (−9.3 % effect estimate, 95 % CI, −16.5–−1.5 %, p=0.02, Figure 2B). The fall velocity was also slower in the CD+ group, though not reaching statistical significance (−7.7 % effect estimate, 95 % CI, −15.3–0.7 %, p=0.08). There were no significant differences between groups in trough glucose or the other secondary outcomes, though effect estimates were all negative (Table 2). Figure 2 illustrates representative CGM meal glucose data for all meals (Figure 2C) and the subset of first meals (Figure 2D).

Summative glucose metrics

Last, we analyzed differences in 30-day summative CGM metrics (that might be available at a clinical visit) in relation to celiac status. Among individuals in the cohort, there was a median of 95.3 % (91.9–98.7 %) of expected CGM values available, with only one individual with <70 % of expected CGM values (56.6 %). The summative CGM metrics, including time below 70 mg/dL and time below 54 mg/dL, did not significantly differ between groups when calculated over the full 24 h per day or when limited to waking hours (Table 3).