OGTT Metrics Surpass Continuous Glucose Monitoring Data for T1D Prediction in Multiple-Autoantibody–Positive Individuals

Abstract

Context

The value of continuous glucose monitoring (CGM) for monitoring autoantibody (AAB)-positive individuals in clinical trials for progression of type 1 diabetes (T1D) is unknown.

Objective

Compare CGM with oral glucose tolerance test (OGTT)–based metrics in prediction of T1D.

Methods

At academic centers, OGTT and CGM data from multiple-AAB relatives were evaluated for associations with T1D diagnosis. Participants were multiple-AAB–positive individuals in a TrialNet Pathway to Prevention (TN01) CGM ancillary study (n = 93). The intervention was CGM for 1 week at baseline, 6 months, and 12 months. Receiver operating characteristic (ROC) curves of CGM and OGTT metrics for prediction of T1D were analyzed.

Results

Five of 7 OGTT metrics and 29/48 CGM metrics but not HbA1c differed between those who subsequently did or did not develop T1D. ROC area under the curve (AUC) of individual CGM values ranged from 50% to 69% and increased when adjusted for age and AABs. However, the highest-ranking metrics were derived from OGTT: 4/7 with AUC ∼80%. Compared with adjusted multivariable models using CGM data, OGTT-derived variables, Index60 and DPTRS (Diabetes Prevention Trial-Type 1 Risk Score), had higher discriminative ability (higher ROC AUC and positive predictive value with similar negative predictive value).

Conclusion

Every 6-month CGM measures in multiple-AAB–positive individuals are predictive of subsequent T1D, but less so than OGTT-derived variables. CGM may have feasibility advantages and be useful in some settings. However, our data suggest there is insufficient evidence to replace OGTT measures with CGM in the context of clinical trials.

It is accepted that most individuals with multiple diabetes-related autoantibodies (AABs) will eventually progress to clinical type 1 diabetes (T1D). This concept is codified in a description of disease stages: multiple AABs with normal glucose tolerance is considered stage 1 disease, development of abnormal glucose tolerance defines stage 2 diabetes, and the clinical onset of disease is stage 3 (1). This categorization depends on data obtained from a formal 2-hour oral glucose tolerance test (OGTT) in which AAB positive individuals present in a fasted state, undergo intravenous (IV) placement, and consume a weight-based volume of Glucola over a 5-minute period; blood samples for glucose, insulin, and C-peptide are generally obtained at 30-minute intervals. In clinical trials of multiple-AAB–positive individuals, OGTT testing is generally performed at 6-month intervals. In addition to providing glucose values to categorize the glucose tolerance status of the individual, derived data from OGTT measures have been examined as predictors of disease progression. These include measures of secretion from early (0-30 minute) or integrated (area under the curve; AUC) insulin or C-peptide calculations, combinations of variables for risk scores including Diabetes Prevention Trial-Type 1 (DPT-1) Risk Score (DPTRS), and Index60, as well as estimates of insulin resistance (2-4). DPTRS and Index60 can be used to further refine time dependent risk in those with multiple AABs.

With diabetes defined by both the American Diabetes Association and the World Health Organization according to dysglycemia that crosses glucose thresholds from OGTT or HbA1c measures, such tests are likely to continue to be the key outcome measures for trials to delay or prevent T1D.

With rapid innovation in technology used for diabetes care, exploration of whether continuous glucose monitoring (CGM) data can be leveraged to assess disease risks and progression has become an area of intense investigation. Indeed, CGM-based metrics, like percent time above 140 or 160 mg/dL, have been associated with disease progression in several studies (5, 6). It is not currently known whether CGM data may be appropriate for entry or outcome criteria for clinical intervention trials or could provide additional insights as to the trajectory of glucose abnormalities during disease progression.

We analyzed longitudinal CGM data from the TrialNet Pathway to Prevention (TN01) CGM Metrics and Dysglycemia ancillary study to compare the ability of CGM and OGTT metrics to predict T1D in multiple-AAB–positive individuals.

Materials and Methods

Study Population

Data were obtained from the TrialNet Pathway to Prevention (TN01) CGM Metrics and Dysglycemia ancillary study initiated in 2015. Participants and/or their guardians provided informed consent (and assent, if applicable) to enroll in this study at participating TrialNet sites and undergo up to 3 CGM assessment periods every 6 months. This was in accordance with the protocol approved by the TrialNet Institutional Review Board, applicable across all participating TrialNet sites. As previously described (6), 105 first- or second-degree relatives were asked to wear a Dexcom G4 Platinum CGM system (Dexcom, San Diego, CA) for up to 7 days on 3 occasions 6 months apart. While only participants with multiple AABs at enrollment were included (n = 93) in our analysis, by the time of their baseline visit 7 individuals reverted to a single AAB. The median (range) of time from enrollment to baseline was 6.2 (1.7-18.9) months. These individuals (including those who reverted to a single AAB) were further classified as stage 1 (n = 58, normal glucose tolerance) and stage 2 T1D (n = 35, abnormal glucose tolerance) at their baseline visit. Visits occurred at baseline, 6 months, and 12 months, and participants discontinued follow-up if diagnosed with T1D during the follow-up period. Each CGM assessment period had an associated OGTT visit that took place during, or approximately 7 days before or after sensor placement. HbA1c was also measured at each visit. Additional OGTT and/or T1D status was obtained from ongoing participation in TN01. The median length of follow-up time from baseline visit until T1D diagnosis or last available data from TN01 was 3.05 years (range 0.04-7.4 years).

CGM Data Processing

As previously described, CGM glucose data that were complete for 4 full days were used for analysis; data from the first 12 hours of sensor wear were excluded and data were truncated at 06:00 hours on the final day of CGM wear (6). CGM data were analyzed in 3 timeframes separately: 24 hours, daytime (06:00-00:00 hours), and nighttime (00:00-06:00 hours) (7). Implausible CGM data were excluded, particularly removing compression lows where sensor glucose was <50 mg/dL during nighttime hours and the individual was likely to be lying on the sensor; strings of consecutively low readings that were flanked with readings ≥80 mg/dL within 10 minutes before and after the string; strings of >200 mg/dL that were flanked with readings <70 mg/dL (8).

Statistical Analysis

CGM metrics were computed using the R package “iglu” (9, 10) at each visit using the all_metrics() function, which outputs 48 metrics; metrics were computed for 24-hour, daytime (06:00-00:00 hours), and nighttime (00:00-06:00 hours) timeframes, individually. Both CGM- and OGTT-derived metrics, and HbA1c were assessed for predictive ability of T1D. T1D outcomes included T1D diagnosed during follow-up and T1D diagnosed within 1 year (±10 weeks) from final CGM. C-peptide AUC mean was determined at each OGTT visit using the trapezoidal rule; other OGTT-derived metrics included early C-peptide (change in C-peptide from 0 to 30 minutes), fasting glucose and fasting C-peptide (mean of −10 and 0 minute), 2-hour glucose, Index60 (4), and DPTRS (2). Definitions of CGM metrics in the R package “iglu” have been previously published (9). All glucose metrics are in mg/dL and OGTT-derived C-peptide metrics are in pmol/mL.

To evaluate longitudinal CGM data, change scores between visits were computed for continuous CGM and OGTT variables. Spearman correlations were computed for associations of CGM metrics between different timeframes, and for associations of CGM and OGTT metrics with age. Chi-square tests were done for associations of categorical CGM and OGTT data with T1D diagnosis, or Fisher's exact tests where expected counts were less than 5. Associations of continuous data with T1D were determined using Wilcoxon rank sum tests for comparison of metrics between T1D-diagnosed and nondiagnosed groups at each timepoint, and between stages of T1D. The kappa statistic was used to assess agreement between OGTT categorization definitions and CGM categorization definitions.

Logistic regression models used CGM and OGTT metrics as predictors of T1D diagnosis, also including baseline age and total AABs as covariates. Receiver operating characteristic (ROC) curves were generated, and AUC was determined. Specificity, sensitivity, positive predictive value (PPV), and negative predictive value (NPV) were calculated based on the optimal cut-point as identified by Youden's J index (11). Lastly, variable clustering, employing principal components analysis, was done on all 48 CGM metrics for variable reduction. These most representative variables were assessed for their combined predictive ability of T1D in logistic regression models, both unadjusted and adjusted for age and autoantibodies; similarly, expert-identified CGM variables were combined in logistic regression models adjusted for age and autoantibodies and assessed for predictive ability of T1D. Comparisons of ROC curves were done between data-driven CGM multivariable models, expert identified CGM multivariable models, and OGTT-derived variables. All analyses were performed using SAS 9.4 (The SAS Institute; Cary, NC, USA), JMP Pro 16 (SAS Institute; Cary, NC, USA), and R version 4.2.2 (12). Statistical significance was defined as P < .05.

Results

Characterization of Individuals Who Progressed to Stage 3 T1D

In this cohort of 93 at-risk individuals, 75 underwent a second visit, and 48 completed all 3 visits (Fig. 1). Thirty-four of 93 (37%) individuals progressed to stage 3 T1D; 20 of these were diagnosed within 1.2 years from their last CGM visit. Characteristics of those with and without T1D overall (Table 1) and within 1.2 years from last CGM visit (Table S1 (13)) indicate that those who developed T1D were younger and more likely to have more than 2 autoantibodies. Eight individuals from the entire cohort had protective human leukocyte antigen (HLA), with 1 of these participants progressing to stage 3 T1D.

Figure 1.

CONSORT diagram. Ninety-four participants from TrialNet Pathway to Prevention study (TN01) enrolled in TrialNet ancillary CGM study with follow up through December 2022 in TN01.

Table 1.

Baseline characteristics of participants with and without subsequent T1D

	No T1D (n = 59)	Subsequent T1D (n = 34)	P value
Age at baseline CGM visit			.003a
Mean (SD)	26.6 (15.2)	18.2 (13.3)
Median	21.8	12.8
Q1, Q3	14.2, 43.6	10.1, 24.8
Sex			.065b
Female	36 (61.0%)	14 (41.2%)
Male	23 (39.0%)	20 (58.8%)
Race			.312b
Asian	0 (0.0%)	1 (2.9%)
White	55 (93.2%)	32 (94.1%)
Unknown or not reported	4 (6.8%)	1 (2.9%)
Total antibodies at baseline visit			.012b
1	5 (8.5%)	2 (5.9%)
2	22 (37.3%)	2 (5.9%)
3	17 (28.8%)	13 (38.2%)
4	11 (18.6%)	13 (38.2%)
5	4 (6.8%)	4 (11.8%)
DR3			.416b
Absent	38 (64.4%)	19 (55.9%)
Present	21 (35.6%)	15 (44.1%)
DR4			.962b
Absent	24 (40.7%)	14 (41.2%)
Present	35 (59.3%)	20 (58.8%)
Protective HLA			.139b
Absent	52 (88.1%)	33 (97.1%)
Present	7 (11.9%)	1 (2.9%)

Data are n (%) unless otherwise noted.

Abbreviations: CGM, continuous glucose monitoring; HLA, human leukocyte antigen; T1D, type 1 diabetes.

^a Wilcoxon.

^b Chi-square.

Characteristics of CGM Data in At-Risk Individuals

All CGM metrics for 24 hours were highly correlated with daytime values (Spearman correlation range 0.61 for CVsd, standard deviation [SD] of daily coefficient of variation [CV], to 0.97 for minimum glucose); weaker correlations were noted for the 24-hour and nighttime only data (correlation 0.20 for CVsd to 0.88 for median glucose; Fig. S1 (13)). Since the CGM nighttime data had lower variability than daytime and given the strong association between 24-hour values and daytime only data, analyses were conducted using daytime CGM metrics.

Examining data from the baseline visit, 29 of the 48 (60.4%) daytime CGM metrics showed statistically significant differences between those who did and did not eventually develop T1D (Table 2). Thirty-three of the 48 (68.8%) CGM metrics were different between those who did and did not develop T1D within 1.2 years of final CGM visit (Table S2 (13)). Twenty-six of these were the same as the metrics different at baseline for subsequent T1D overall; 7 variables, all associated with hyperglycemia, were different only for those who developed T1D within 1.2 years of assessment. The converse was true for 3 variables, all measures of glycemic variability (average daily risk range (ADRR; 14), SD of the rate of change (SD.roc; 15), average SD of glucose hour-long intervals (SDwsh; 16)).

Table 2.

Median and IQR values for CGM metrics that differed statistically significantly at baseline in those with or without subsequent T1D

Daytime CGM metric	No T1D (n = 59)	Subsequent T1D (n= 34)	P value
% above 140	2.46 (0.7, 6.45)	5.45 (2.16, 11.02)	.026
% above 180	0 (0, 0)	0.16 (0, 1.33)	.003
% in range 60-140	96.51 (93.27, 98.52)	93.38 (88.23, 97.02)	.003
3rd quartile	108 (102, 116)	113.88 (107, 121)	.016
ADRR	11.38 (8.11, 14.83)	12.74 (10.56, 17.95)	.041
Conga	21.95 (19.58, 26.28)	26.03 (22.68, 30.41)	.006
CV	18.26 (15.14, 21.91)	20.88 (17.95, 22.52)	.009
CVmean	16.12 (13.34, 18.97)	17.36 (15.56, 19.6)	.034
GRADE	1.27 (0.94, 1.68)	1.66 (1.31, 2.17)	.003
GRADE eugly	67.74 (54.03, 84.29)	58.72 (50.73, 68.87)	.015
GRADE hyper	14.46 (5.19, 27.65)	23.68 (10.7, 44.26)	.037
HBGI	0.13 (0.06, 0.33)	0.28 (0.14, 0.61)	.011
Hyper-index	0.01 (0, 0.04)	0.04 (0.01, 0.1)	.014
IQR	22 (18.5, 28)	25 (22, 31)	.002
J-index	13.83 (11.69, 15.96)	15.36 (13.36, 17.65)	.014
MAD	14.83 (13.34, 20.76)	17.79 (16.31, 22.24)	.005
MAG	12.86 (9.95, 15)	14.67 (12.94, 17.27)	.003
MAGE	43.13 (36.98, 53.31)	51.52 (42.39, 59.49)	.017
Max	166 (154, 174)	183 (164, 206)	.004
MODD	17.06 (14.98, 20.52)	19.87 (17.74, 22.79)	.005
M-value	1.05 (0.79, 1.8)	1.73 (1.06, 2.68)	.002
Range	108 (96, 120)	121 (104, 148)	.005
SD	17.83 (14.73, 21.67)	21.11 (17.32, 24.21)	.003
SD.roc	0.66 (0.55, 0.75)	0.71 (0.64, 0.84)	.029
SDb	15.28 (13.21, 17.85)	17.89 (14.91, 20.56)	.003
SDb//dm	13.82 (11.67, 15.99)	15.94 (13.46, 19.19)	.002
SDhh:mm	9.11 (7.14, 11.71)	10.74 (9.02, 13.32)	.031
SDw	16.12 (13.09, 18.82)	18.73 (15.13, 20.77)	.009
SDwsh	6.7 (5.56, 7.65)	7.44 (6.51, 8.44)	.010

Abbreviations: ADRR, average daily risk range; CGM, continuous glucose monitoring; CV, coefficient of variation; GRADE, glycemic risk assessment diabetes equation; HBGI, high blood glucose index; IQR, interquartile range; MAD, median absolute deviation; MAG, mean absoluteglucose; MAGE, mean amplitude of glycemic excursions; MODD, mean of daily differences; SD, standard deviation; SD.roc, SD of rate of change; SDb, SD between days for each timepoint; SDb//dm, SD between days for each timepoint corrected for changes in daily means; SDhh:mm, SD between timepoints; SDw, SD within days; SDwsh, average SD of hour-long intervals; T1D, type 1 diabetes.

Since age is well-known to be associated with T1D progression, how each of these CGM metrics associated with age was explored. Many metrics were strongly correlated with age (Fig. 2). Interestingly, measures reflective of hypoglycemia (percentage of time below 54 mg/dL, percentage of time below 70 mg/dL, glycemic risk assessment in diabetes equation [GRADE] hypo, low blood glucose index [LBGI], hypo-index) had a strong inverse association with age. Those who were younger tended to experience more hypoglycemia. Hypo-index was one of the most strongly associated with age: daytime correlation of −0.51 (P = 1.48 × 10^–7) at baseline. When stratifying by stage of T1D, this relationship was preserved; correlations were around −0.50 at baseline with P < .01, in both stage 1 and stage 2 (Fig. 2).

Figure 2.

CGM metrics are associated with age. Heatmap shows Spearman correlations of daytime CGM metrics with age. Many CGM metrics are strongly associated in all participants (n = 93) and associations are consistent regardless of disease stage (n = 58 stage 1, n = 35 stage 2). Many metrics indicative of hypoglycemia are inversely correlated with age.

Hypoglycemia was fairly rare at baseline. While 32% (n = 30) of individuals experienced glucose levels below 54 mg/dL, the median (interquartile range, IQR) percentage time below 54 mg/dL among these individuals was only 0.61 (0.30-1.31). More individuals (83 of 93) experienced glucose levels below 70 mg/dL at some point during the study, though the median time below 70 mg/dL among these individuals was low: median percent time (IQR) of 3.38 (0.89-8.05) (Fig. 3). Thus, hyperglycemia was contributing more to time out of a target range, as opposed to hypoglycemia.

Figure 3.

Distribution of percent of time with hypoglycemia at baseline, and 6- and 12-month visits. Metrics summarizing hypoglycemia are shown as percent of time (A) below 54 mg/dL and (B) below 70 mg/dL. Hypoglycemia was rare, especially for percent of time below 54 mg/dL (note different y-axis scales). Violin plots are overlaid with boxplots to show distributions. Pink indicates participants diagnosed with subsequent T1D (n = 34).

Characteristics of OGTT Data in At-Risk Individuals

At baseline, 5 OGTT derived metrics (2-hour glucose, DPTRS, Index60, C-peptide AUC mean, and early C-peptide) demonstrated differences between clinical T1D and non-T1D groups (Table 3). Fasting glucose, fasting C-peptide, and HbA1c showed no statistically significant difference between T1D and non-T1D at any study visit.

Table 3.

Median and interquartile range values for OGTT-derived/other metrics in those with or without subsequent T1D

OGTT metric	No T1D (n = 59)	Subsequent T1D (n = 34)	P value
2-hour glucose	108 (93, 134)	147.5 (112, 184)	1.18 × 10–04
C-peptide AUC mean	2.15 (1.83, 2.75)	1.69 (1.29, 2.24)	.003
DPTRS	5.59 (4.41, 6.23)	7.03 (6.08, 7.69)	1.44 × 10–05
Early C-peptide	1.41 (0.93, 1.86)	0.78 (0.54, 1.49)	9.12 × 10–05
Fasting C-peptide	0.62 (0.47, 0.81)	0.54 (0.36, 0.75)	.128
Fasting glucose	92.5 (85, 100)	92.25 (87.5, 98)	.848
HbA1c	5.2 (5, 5.3)	5.3 (4.9, 5.5)	.270
Index60	−0.43 (−0.99, 0.24)	0.92 (0.18, 1.33)	3.06 × 10–06

Abbreviations: OGTT, oral glucose tolerance test; T1D, type 1 diabetes.

Individual CGM and OGTT Metrics for Prediction of T1D

Changes from baseline to 6 months were then computed for each of the CGM and OGTT metrics. Longitudinal analyses focused on the change from baseline to 6 months (n = 75) since approximately half (51.6%) of participants did not have a 12-month visit. When comparing change from baseline to 6 months, there were no statistically significant differences in change scores for any of the CGM nor any of the OGTT metrics between those who progressed and did not progress to T1D.

ROC curves were generated to determine how individual CGM and OGTT metrics obtained at baseline predicted subsequent T1D diagnosis. AUC values for CGM metrics ranged from 50.2% to 69.3%; age adjustment had a marked impact on AUC values; after adjusting for age, all 48 daytime CGM metrics had AUC exceeding 65% (Fig. 4A). However, OGTT-derived metrics (Index60, DPTRS, 2-hour glucose, and early C-peptide) ranked highest in predictive ability of T1D in both unadjusted and adjusted analyses.

Figure 4.

Discrimination ability of baseline metrics for (A) subsequent development of T1D and (B) T1D within 1.2 years of final CGM. ROC AUC values are plotted, where blue markers indicate OGTT-derived metrics or HbA1c, black indicates CGM daytime (06:00-00:00 hours) metrics. ROC curves were determined from logistic regression models; metrics were individually assessed for (A) prediction of subsequent T1D, and (B) T1D within 1.2 years of final CGM visit in unadjusted (circle), age-adjusted (triangle), and age and autoantibody-adjusted (cross) models. Metrics are ranked by mean AUC of the 3 models.

Similar analysis was done comparing the final CGM values for 20 participants diagnosed with T1D within 1 year (±10 weeks: 0.8-1.2 years) of their final CGM visit to 69 individuals without T1D during that time. Unadjusted for age or AAB status, SDw (variability defined as average of daily glucose SD) (16) and median absolute deviation had AUC values close to 80%, with adjustments for age and AAB, many CGM metrics reflecting hyperglycemia as well as variability had ROC ≥75%. However, OGTT metrics remained the strongest predictors with the AUC for DPTRS at 91% and Index60 at 87% without requiring adjustment for any covariates (Fig. 4B).

The distribution of values for the top ranking OGTT and CGM metrics between those who developed subsequent T1D (Index60 ROC with AUC 79.8% and SDb//dm (SD between days for each timepoint corrected for changes in daily means; 15), a measure of variability between days, with AUC 69.3%) or T1D within 1.2 years of their final CGM (DPTRS with AUC 91.0% and SDw, average SD of daily glucose, with AUC 77.5%) is shown in Fig. 5.

Figure 5.

Baseline distributions for OGTT and CGM metrics with greatest discrimination ability for prediction of T1D outcomes. Violin plots overlaid with boxplots show distributions of top ranking OGTT and CGM metrics for prediction of subsequent T1D (A, B) and T1D within 1.2 years of final CGM (C, D). OGTT-derived Index60 and DPTRS had the highest AUC in unadjusted logistic regression models for prediction of subsequent T1D (A; AUC 80%) and T1D within 1.2 years (C; AUC 90%), respectively. Daytime CGM metrics SDb//dm and SDw, both measures of variability, had the highest AUC in unadjusted logistic regression models for prediction of subsequent T1D (B; AUC 69%) and T1D within 1.2 years (D; AUC 78%), respectively. All metrics tend to be higher in those diagnosed with T1D than those not diagnosed with greater differences seen closer to diagnosis (C, D).

Multivariable Models for Prediction of T1D

Clustering of the 48 baseline CGM metrics identified 6 clusters. Within each cluster the most representative variables were glucose coefficient of variation (CV), mean glucose, hypoglycemic index, IQR, hyperglycemic index, and SDwsh (average SD of glucose in hour-long intervals) (16) (Table S3 (13)).

We then asked how well these 6 representative CGM variables discriminate between T1D and non-T1D individuals. The unadjusted AUC for these combined CGM variables is 69.1% (Fig. 6). As noted for the individual CGM variables, age and antibody adjustment improved the AUC (76.6%). The ROC curve generated from the age and AAB-adjusted model identified an optimal threshold corresponding to PPV of 57.8% and NPV of 83.3% (Table S4 (13)). We then performed similar calculations using each of the OGTT metrics. The OGTT-derived metrics Index60 and DPTRS performed best in discrimination between subsequent T1D and non-T1D individuals, without including any additional covariates: Index60 AUC 79.8%, PPV 67.6%, NPV 84.2%; DPTRS AUC 78.9%, PPV 70.4%, NPV 82.5% (Fig. 6; Table S4 (13)).

Figure 6.

Discrimination ability for prediction of subsequent T1D demonstrated by ROC curves. ROC curve of CGM model for prediction of T1D using most representative variables from clustering as predictors (CV, mean, hypo-index, hyper-index, IQR, SDwsh) is indicated by teal line. ROC curves of OGTT-derived variables DPTRS and Index60 for T1D prediction are indicated by the pink and purple lines, respectively. DPTRS and Index60 were consistently identified as top predictors of T1D outcomes. The dashed diagonal line is representative of a random classifier.

CGM metrics identified from clustering are not necessarily the most readily recognized by clinicians. Thus, we evaluated a combination of clinician recommended CGM metrics in a multivariable model: percent time in range 60 to 140 mg/dL (target range used for individuals who have not developed stage 3 T1D (17)), percent time above 140 mg/dL, glucose CV, and mean glucose (Table S4 (13)). After analyzing ROC curves from this multivariable model, predictive ability was not improved in comparison to the model including only CGM metrics from the data-driven clustering approach. For instance, the unadjusted clinician identified model had an AUC of 69.4% vs data-driven clustering AUC of 69.1% (ROC contrast estimate 0.003; 95% CI −0.076 to 0.082, P = .93).

OGTT and CGM Metrics Fluctuate Over Time

Thresholds are often used in the context of T1D to categorize individuals in terms of disease risk, progression, and diagnosis. One challenge with the use of OGTT-defined thresholds for normal, abnormal, and diabetic glucose tolerance is that it is not uncommon for an individual to demonstrate variation over time such that they change from one category to another. In the context of clinical trials, this has often resulted in requiring sequential confirmation of glucose tolerance status when used as entry or outcome criteria. Therefore, it was of particular interest to evaluate how categorization of individuals changed over time when using either CGM or OGTT-derived definitions.

As previously reported (6), individuals whose CGM at baseline demonstrated at least 5% or 8% of time above 140 mg/dL had a higher likelihood of subsequent T1D. At baseline, 35 (37.6%) individuals had ≥5% of time above 140 mg/dL, of which 18 (51.4%) progressed to T1D. There were 22 (23.7%) individuals at baseline with ≥8% of time above 140 mg/dL, of which 12 (54.5%) progressed to T1D. The threshold of ≥10% of time above 140 mg/dL was not informative as only 4 individuals did not also meet the ≥8% category (ie, were at 9% time above 140 mg/dL), and only 1 of these was subsequently diagnosed with T1D. As expected, fewer individuals met these definitions with increasing cutoffs; a 15% of time above 140 mg/dL was demonstrated in only 6 (6.5%) participants.

We then evaluated how many individuals changed categories according to the 5% and 8% thresholds, and whether this related to subsequent T1D. Changing categories from baseline to 6 months was less common if defined as ≥8% of time >140 mg/dL compared with ≥5% of time (Fig. S2A and S2B (13)). Using the 8% cutoff, 16 individuals changed categories (10 abnormal to normal, 6 normal to abnormal) while 22 changed using the 5% cutoff (14 abnormal to normal, 8 normal to abnormal).

There were only 4 individuals, all of whom developed T1D, who had ≥8% of time >140 mg/dL at both the baseline and the 6-month visit. Of the 16 who changed CGM categories, 4 (25%) developed T1D. However, 55 individuals did not meet this criterion at either timepoint; nearly 30% of these individuals (n = 15) progressed to eventual T1D.

We then evaluated OGTT-defined glucose tolerance at baseline and 6 months (Fig. S2C (13)). There were 21 individuals who had OGTT defined abnormal glucose tolerance at both baseline and 6 months. Of these 13 (61.9%) developed T1D. As seen with CGM categorizations, individuals shifted between categories (16 total: 6 abnormal to normal, 10 normal to abnormal). Five (31%) of these developed T1D. Thirty-eight individuals had normal glucose tolerance at both timepoints, and only 13% (n = 5) of these individuals progressed to clinical T1D.

Lastly, we assessed agreement between OGTT categories and CGM categories at all visits using the kappa statistic, which evaluates the concordance between the results. OGTT glucose tolerance classification had only fair agreement with CGM categories when categorizing CGM data by time above 140 mg/dL ≥5% of time (kappa = 0.23, 95% CI 0.10-0.37; ≥8% of time kappa = 0.23, 95% CI 0.10-0.36). For ≥5% of time spent above 140 mg/dL and OGTT-determined glucose tolerance, the categories were concordant in 63 (74%) baseline visits and discordant in 30, but when assessing concordance at all visits, only 65% agreement was seen (139 of 214 total visits). There was, however, a statistically significant relationship observed at baseline between OGTT and CGM categories. For instance, 15 of the 22 (68%) with abnormal glucose tolerance also met the abnormal threshold of ≥8% of time above 140 mg/dL (chi-square P < .001; Fig. S3 (13)).

Discussion

While T1D is due to impaired insulin secretion, the definition of early stages and the clinical diagnosis relies on standardized measures of glucose dysregulation as determined by oral glucose tolerance testing. As such, glucose tolerance categories from an OGTT for those with multiple AABs are used as entry or outcome criteria for clinical trials of disease modifying therapies in early-stage disease (18-20) and OGTT glucose values are used to define the clinical onset of T1D. Further, derived values from OGTT including measures of insulin secretion (early C-peptide, C-peptide AUC, DPTRS, Index60) have all been shown to be predictive for development of T1D in those with multiple AABs (2, 3, 21). Despite the importance of OGTT measures, a standard OGTT requires a visit to a research or clinical facility while fasting and either multiple blood draws or placement of an IV to conduct the test. The OGTT assesses disease progression at limited time periods (typically every 6 months). With the widespread use of CGM in those with stage 3 T1D, consideration has been given to leveraging this technology in early stages of T1D either as an alternate, or complementary method, to OGTT-based measures. While glycemic dysregulation identified by various CGM measures has been associated with disease progression (5, 6), a comparison of the predictive value of CGM and OGTT measures for subsequent T1D has not been explored.

Our data demonstrate that individual baseline metrics obtained from both CGM and OGTT are associated with disease progression, but to differing degrees. The highest ROC values for prediction of subsequent T1D were for the OGTT derived metrics Index60 (80%) and DPTRS (79%). In contrast, none of the individual CGM metrics exceeded an AUC of 70% for prediction of T1D. The CGM metrics with the highest AUC for prediction of subsequent T1D were glucose measures including mean average glucose, percent time in range (60-140 mg/dL), and GRADE, and 1 measure of variability (SDb//dm). In the present analysis none of these individual CGM metrics exceeds the predictive value for subsequent T1D of data derived from OGTT metrics. Interestingly, HbA1c was not significantly different between those who subsequently did or did not develop T1D at any study visit.

Not surprisingly, using only CGM or OGTT data obtained within about a year of the onset of T1D increased the strength of prediction of T1D. For CGM, both median absolute deviation and SD within days (SDw, the average daily SD) had AUC values close to 80%. However, the OGTT metrics remained stronger predictors with DPTRS exceeding an AUC of 90%.

When looking at either subsequent T1D or T1D within ∼1 year from final CGM visit, adjusting for age notably improved predictive ability, particularly for CGM metrics. It is important to note that age is incorporated in the DPTRS metric. Including AABs as an additional covariate also increased predictive ability for many variables, but to a lesser degree.

To further explore whether combining CGM metrics would improve the prediction of T1D, we used a data-driven approach that generated 6 clusters from the 48 CGM variables. For each cluster we selected the most representative variable and then determined the AUC for these 6 variables together. When adjusted for age and number of AABs, the 6 combined CGM variables had an AUC for prediction of T1D of 76.6%; similar to, but less than the AUCs for OGTT metrics, DPTRS and Index60. Using the optimal threshold for these adjusted combined CGM variables, the NPV was 83%; essentially identical to the NPV for both DPTRS and Index60. However, in contrast to PPV values of 70% for DPTRS and 68% for Index60, the PPV for covariate adjusted combined CGM variables was only 58%, indicating the superiority of OGTT metrics.

OGTT-derived parameters have demonstrated variability over time, causing an individual's status to alternate between normoglycemic and dysglycemic states. We thus aimed to determine whether a CGM-defined threshold led to more consistent, stable classification over time. Since ≥8% of time with glucose >140 mg/dL has been identified as an important classifier of risk, this threshold was applied to determine how many individuals changed between risk classes and whether this was related to development of T1D. Little difference in fluctuation around this CGM threshold and the OGTT classifications regardless of subsequent development of T1D was noted. Thus, the use of this CGM category as an entry criteria or outcome measure would not have an advantage over OGTT. We assessed thresholds higher than 8% but found that few participants met these criteria—it would be of interest to evaluate these higher thresholds in larger studies. Using the available longitudinal data, changes over time in CGM parameters did not provide added predictive value for T1D development.

Hypoglycemia was not observed as a risk factor for subsequent T1D. While 32% of individuals had glucose levels below 54 mg/dL, the median percentage of time with this degree of hypoglycemia was less than 1%. Glucose values below 70 mg/dL were observed for 83 of the 93 individuals at some point during the study with the median of 3.38%. A median time of 1.1% (15 minutes/day) <70 mg/dL has been previously reported in healthy individuals (8). Nevertheless, the presence of occasional hypoglycemia in at-risk individuals is consistent with previous work from the DPT-1, which periodically assessed 4 daily capillary glucose values and reported asymptomatic hypoglycemia in both the control and parenteral insulin groups (22), suggesting that the ailing beta cell results in mismatch in timing of insulin release to physiological demands.

There are several caveats to our study. While 93 multiple-AAB–positive participants were enrolled, only 34 (37%) developed clinical T1D during the period of follow-up. It is expected that essentially all individuals with multiple AABs will eventually progress to clinical disease; longer follow-up may find that CGM metrics or changes in HbA1c are beneficial in predicting subsequent T1D or prove informative about predicting time to T1D. Further, though the study design was longitudinal, almost half did not complete all 3 CGM assessments as planned. The sample size also limited our ability to explore time to T1D by stages. At the time of the conduct of this ancillary study, the Dexcom G4 was used. With newer CGM technology, the need for fingerstick calibrations has been alleviated, devices have a disposable transmitter, and the size of the device has been vastly reduced (23). It is possible that a newer device would have improved compliance and accuracy in our ancillary study. Additionally, this study involves a predominantly white cohort: all participants who reported race were of white racial background with the exception of 1 Asian participant.

Our study was not designed to address questions about the role of CGM in AAB-positive individuals in the context of clinical practice or to compare practical issues between CGM and OGTT such as cost and patient acceptance, including the psychological impact of wearing a device potentially years prior to clinical T1D. CGM was worn in 6-month intervals in this study, and, thus, further work is needed to determine whether CGM on a more frequent basis could add value in a surveillance program. The present analysis found that the glucose value at the 2-hour time point during the OGTT conducted under standardized conditions and when adjusted for age and number of AABs had a high ROC AUC for prediction of T1D. Like the use of CGM, whether this observation could translate into clinical recommendations for periodic standardized home postprandial glucose testing in those with multiple AABs requires additional study.

In summary, CGM measures in multiple-AAB–positive individuals are predictive of subsequent T1D, but less so than established variables including age, AABs, and OGTT glucose, C-peptide, or other OGTT-derived variables. While CGM can be done at home, it too carries burdens on participants, and like OGTT categories, these variables fluctuate over time. As screening for T1D risk transitions to clinical practice, CGM may have an important role in AAB-positive individuals in monitoring for presymptomatic glucose decompensation. However, our data suggest there is insufficient evidence to replace OGTT measures with CGM in the context of clinical trials.

Abbreviations

AAB

autoantibody

AUC

area under the curve

CGM

continuous glucose monitoring

CV

coefficient of variation

HLA

human leukocyte antigen

IV

intravenous

OGTT

oral glucose tolerance test

PPV

positive predictive value

NPV

negative predictive value

ROC

receiver operating characteristic

SD

standard deviation

T1D

type 1 diabetes

Funding

This ancillary study was supported by the National Institutes of Health (NIH) and National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) (DP3DK101083), Juvenile Diabetes Research Foundation (JDRF) (SRA-2019-763-A-N), and the McNair Medical Institute at The Robert and Janice McNair Foundation (awarded to Massimo Pietropaolo MD). The Type 1 Diabetes TrialNet Study Group is a clinical trials network funded through a cooperative agreement by the National Institutes of Health (NIH) through the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK), the National Institute of Allergy and Infectious Diseases (NIAID), the Eunice Kennedy Shriver National Institute of Child Health and Human Development, through UC4 DK106993, and JDRF.

Disclosures

C.B., C.J.G., S.L., S.L.P., A.K.S., A.Y. have nothing to declare. J.L.G. consults for Vertex Pharmaceuticals, Avotres Inc., Imcyse; has previously consulted for Dompé; has equity interest in Vertex Pharmaceuticals; has research funding from Avotres Inc., Dompé, and Imcyse. J.L.S. serves, or has served, on advisory panels for Bigfoot Biomedical, Cecelia Health, Insulet Corporation, Medtronic Diabetes, StartUp Health Diabetes Moonshot, and Vertex Pharmaceuticals; has served as a consultant to Abbott Diabetes, Bigfoot Biomedical, Insulet, Medtronic Diabetes, and Zealand. Yale School of Medicine has received research support for J.L.S. from Abbott Diabetes, JAEB Center for Health Research, JDRF, Insulet, Medtronic, NIH, and Provention Bio. C.S. has previously consulted for Vertex Pharmaceuticals. D.M.W. serves on advisory board for Enable Biosciences.

Data Availability

Some or all datasets generated during and/or analyzed during the current study are not publicly available but are available from the corresponding author on reasonable request.