Continuous Glucose Monitor Metrics Are Associated with Emergency Department Visits and Hospitalizations for Hypoglycemia and Hyperglycemia,but Have Low Predictive Value

Abstract

Objective:

Determine whether continuous glucose monitor (CGM) metrics can provide actionable advance warning of an emergency department (ED) visit or hospitalization for hypoglycemic or hyperglycemic (dysglycemic) events.

Research Design and Methods:

Two nested case–control studies were conducted among insulin-treated diabetes patients at Kaiser Permanente, who shared their CGM data with their providers. Cases included dysglycemic events identified from ED and hospital records (2016–2021). Controls were selected using incidence density sampling. Multiple CGM metrics were calculated among patients using CGM >70% of the time, using CGM data from two lookback periods (0–7 and 8–14 days) before each event. Generalized estimating equations were specified to estimate odds ratios and C-statistics.

Results:

Among 3626 CGM users, 108 patients had 154 hypoglycemic events and 165 patients had 335 hyperglycemic events. Approximately 25% of patients had no CGM data during either lookback; these patients had >2 × the odds of a hypoglycemic event and 3–4 × the odds of a hyperglycemic event. While several metrics were strongly associated with a dysglycemic event, none had good discrimination.

Conclusion:

Several CGM metrics were strongly associated with risk of dysglycemic events, and these can be used to identify higher risk patients. Also, patients who are not using their CGM device may be at elevated risk of adverse outcomes. However, no CGM metric or absence of CGM data had adequate discrimination to reliably provide actionable advance warning of an event and thus justify a rapid intervention.

Twitter Summary

While some CGM metrics were significantly associated with the risk of hypoglycemia or hyperglycemia, they did not reliably predict which patients would have an event.

Introduction

Despite decades of improvements in diabetes care and glucose monitoring devices, emergency department (ED) visits and hospitalizations for acute hypoglycemic and hyperglycemic (dysglycemic) events remain common and costly.^1

–4 The current gold standard for glycemic control relies on the hemoglobin A1c test, since it has been well established that A1c control leads to reductions in clinically significant long-term microvascular and macrovascular complications.^5,6

However, A1c levels are not strongly predictive of short-term dysglycemic events, particularly hypoglycemia.^7
–9 We and others have shown that continuous glucose monitor (CGM) initiation is associated with lower rates of ED visits or hospitalizations for hypoglycemia,¹⁰ as well as for hyperglycemia and all-cause hospitalizations.¹¹ In addition, CGM-derived metrics have been shown to better identify individuals at higher risk for dysglycemic events.^12
–14 Thus, remote monitoring of CGM data may open up new opportunities for population health management approaches in the outpatient setting to prevent imminent severe dysglycemic events. This would require a highly discriminatory metric that reliably predicts such events within a timeframe that provides sufficient advance warning for health care providers to mobilize and intervene.

The use of real-time CGM data for population health management to identify patients who would benefit from timely clinical intervention is a potential opportunity to improve diabetes care, but applications have thus far been limited. One such application is the timely intervention for diabetes excellence algorithm, which was developed based on clinical targets from the 2019 consensus report on CGM metrics, and was designed to flag patients with deteriorating glycemic control, which would merit asynchronous, but timely clinical intervention.¹⁵

This algorithm was tested in a pediatric type 1 diabetes clinic and was found to improve time in range (TIR), while reducing staff burden in reviewing CGM data.¹⁶ This proof-of-concept study demonstrates the potential for remote monitoring of CGM data to improve glycemic control. However, the use of remote monitoring of CGM data to predict clinically serious impending dysglycemic events (e.g., ED or hospitalization for hypoglycemia or hyperglycemia) has not been evaluated.

We conducted a nested case–control study of insulin-treated diabetes patients to determine whether remote review of CGM data by a provider in the outpatient setting might provide actionable and sufficient advanced warning (within 8–14 days) before an ED visit or hospitalization for hypoglycemia or hyperglycemia, with enough reliability to justify allocating population-management resources.

Research Design and Methods

Study population

This study was conducted among Kaiser Permanente Northern California members who were insulin-treated patients with diabetes, used a CGM device between January 1, 2016, and January 31, 2021, and shared their time-stamped CGM data with a provider through the Dexcom Clarity diabetes management application. CGM data were linked to individual patients by unique medical record numbers or through a fuzzy matching algorithm if the medical record numbers were missing or incomplete (Supplementary Data). Because this study involved secondary research of identifiable private information for which consent is not required, the Kaiser Permanente Northern California Institutional Review Board determined this study to be exempt.

Case–control definitions

We conducted two nested case–control studies (one for hypoglycemic events and the other for hyperglycemic events). Hypoglycemic “events” were defined for the purposes of this study as having a primary diagnosis of hypoglycemia in the ED or principal diagnosis of hypoglycemia in the hospital using a validated International Classification of Diseases 10th Revision (ICD-10)-based algorithm.¹⁷ Hyperglycemic “events” were defined as having a primary diagnosis of hyperglycemia in the ED or principal diagnosis of hyperglycemia in the hospital using ICD-10 diagnosis codes for diabetic ketoacidosis or hyperglycemic hyperosmolar nonketotic syndrome. Hypoglycemia or hyperglycemia event(s) were analyzed separately.

Event dates were recorded, and patients could contribute multiple events to each case–control study. In each analysis, a patient could be a control until becoming a case, but after that was permitted to be only a case.¹⁸ For each case event, we used incidence density matching with random sampling, with replacement to select multiple observations for control patients, matching on year and month of each case's event date (“index date”). This matching strategy yields unbiased estimates of risk and ameliorates bias due to temporal trends.¹⁹

CGM data acquisition

Patients used smartphones or CGM receivers to collect their blood sugar readings. For smartphone users who gave permission to share their data with the diabetes clinic, glucose data were automatically and continually uploaded to the cloud-based platform that was shared with the provider. For CGM receiver users, CGM data were available only when patients presented to the diabetes clinic for a device upload, or when they performed an upload at home to the cloud platform.

CGM metrics

The CGM devices in this study recorded glucose readings every 5 min, up to 2016 readings per 7 days (12 readings/h × 24 h/day × 7 days = 2016 readings). The CGM metrics were calculated for two separate lookback periods: 0–7 days before each index date (to identify any signal in the days immediately preceding an event, for possible intervention by the patient) and 8–14 days before each index date (to identify any signal that could provide sufficient advanced warning for possible intervention by a provider).

CGM data capture during each lookback period was quantified both by number of days CGM data were present, and by “active time,” defined as the number of readings divided by the maximum possible number of readings. Active time was classified as <70% or ≥70%, since ≥70% use of CGM over a 2-week period has been established as a standard threshold, indicating adequate CGM data for reliable evaluation.^20
–22 Missing CGM data (i.e., active time <100%) could be attributable to device failure, to the patient not uploading data, or to the patient not using the CGM device.

Among those with ≥70% active time, CGM metrics were calculated from the raw, time-stamped CGM data using iglu (Interpreting Glucose Data from Continuous Glucose Monitors, v.3.3.0), an open-source package in R.²³ iglu calculates over 30 peer-reviewed CGM metrics for assessing glucose control and variability, including the standard metrics recommended by the consensus reports.

The metrics examined included those standard metrics (shown in Tables 2 and 3) as well as other, nonstandard metrics (shown in Supplementary Tables S2 and S3): Measures of central tendency (area under the curve, CGM index, calculated A1c, glucose management indicator [GMI], mean GRADE score, percentage of GRADE score attributable to target range, percent TIR, index of glycemic control, J-index, mean absolute deviation, mean absolute glucose, M-value, and mean, median, range, and 25th and 75th percentiles of glucose); measures of glycemic variability (coefficient of variation [CV], mean of the CV, standard deviation (SD) of the CV, continuous overall net glycemic action, glucose variability percentage, SD of glucose values, SD of the rate of change, vertical SD within days, SD between time points, SD within series, horizontal SD, SD between days within time points, SD between days within time points and corrected for changes in daily means, mean amplitude of glycemic excursions, mean difference between glucose values obtained at the same time of day, and average daily risk range); hypoglycemia-specific measures (percentage of GRADE score attributable to hypoglycemia, hypoglycemia index, level 1 time below range (<70 mg/dL), level 2 time below range (<54 mg/dL), level 1 time below range during nighttime, level 2 time below range during nighttime, and low blood glucose index [LBGI]); and hyperglycemia-specific measures (percentage of GRADE score attributable to hyperglycemia, hyperglycemia index, percentage of glucose measures >140 mg/dL, level 1 time above range (>180 mg/dL), level 2 time above range (>250 mg/dL), and high blood glucose index) (Supplementary Table S1).

Statistical analysis

For each subject, the CGM metrics were calculated using data from each of their lookback period (0–7 and 8–14 days) and used to calculate summary statistics (e.g., mean of a CGM metric) among cases and controls. For the primary analysis of standard CGM metrics that are commonly reported on the ambulatory glucose profile (AGP), we used cut points recommended by the consensus reports to define exposure and reference categories.

In secondary analyses of both the standard metrics and nonstandard metrics, we modeled each CGM metric as a continuous variable since recommended cut points for nonstandard metrics have not been established. Odds ratios (OR) and 95% confidence intervals (CI) were estimated from generalized estimating equations using a binomial distribution and logit link function and an exchangeable correlation matrix to properly adjust for nonindependence of the residual error in the repeated measures design.²⁴ Given the rarity of both the outcomes, the OR are a reasonable estimate of relative risk.²³ C-statistics were estimated as a measure of discrimination (i.e., how reliably each metric predicted who will vs. will not have an event); good discrimination was defined as a C-statistic ≥0.7. Data management and analyses were conducted using R version 4.0.2 and SAS^® 9.4.

Results

The study included 3626 CGM users and 343,092,477 time-stamped glucose readings. The mean (SD) number of readings per patient was 94,620 (96,689) with a median of 61,509 readings and an interquartile range of 21,859–136,143 readings. The mean (SD) number of days of readings per patient was 329 (336); the median was 214 days and the interquartile range was 76–473 days.

We identified 154 hypoglycemia events (i.e., cases) observed in 108 patients and incidence density matched them to 3296 observations (controls) in 2101 patients for a total of 3450 hypoglycemia events+matched observations in 2102 patients (Supplementary Figure S1a). All patients with a hypoglycemia event, except for one, also served as controls. As shown in Table 1, the mean age at the end of the study period (January 1, 2021) was 49.3 years, 51.8% were women; 74.9% had type 1 diabetes and 25.1% had type 2 diabetes. The sample included white (66.8%), African American (7.3%), Latino (12.7%), Asian (8.1%), or other (5.0%) race or ethnicity patients. The only significant difference in demographic characteristics between cases and controls was that hypoglycemia cases were more likely to be African American (15.9%) compared to controls (7.3%), P < 0.0001.

Table 1.

Characteristics of 2102 Insulin-Treated Diabetes Patients Who Used a Continuous Glucose Monitor in the Hypoglycemia Case–Control Study

Characteristic	Category	ALL (n = 2102)	Hypoglycemia cases ^a (n = 108)	Controls (n = 2101)	P^b
Age, mean (SD)		49.3 (18.1)	52.8 (18.3)	49.3 (18)	0.06
Sex, n (%)	Women	1089 (51.8)	52 (48.1)	1088 (51.8)	0.39
Sex, n (%)	Men	1013 (48.2)	56 (51.9)	1013 (48.2)
Diabetes type, n (%)	Type 1	1574 (74.9)	78 (72.2)	1574 (74.9)	0.61
Diabetes type, n (%)	Type 2	528 (25.1)	30 (27.8)	527 (25.1)
Race/ethnicity, n (%)	White	1378 (66.8)	68 (63.6)	1377 (66.8)	0.52
	African American	151 (7.3)	17 (15.9)	151 (7.3)	<0.01
	Latino	262 (12.7)	10 (9.3)	262 (12.7)	0.38
	Asian	166 (8.1)	8 (7.5)	166 (8.1)	0.87
	Other^c	105 (5.0)	4 (3.7)	105 (5.0)	0.60

Cases include subjects who experience a hypoglycemic event within the observation window. A patient could be a control until becoming a case, but after that was permitted to be only a case, such that cases and controls are not mutually exclusive.

P-values from generalized estimating equations to account for nonindependence between the cases and controls.

Includes Native American, Pacific Islander and multiracial. Missing race (n = 41) not shown in table.

SD, standard deviation.

We identified 335 hyperglycemic events in 165 patients and matched them to 3577 observations in 1971 control patients for a total of hyperglycemia events plus matched observations in 1971 patients (Supplementary Figure S1b). All patients with a hyperglycemia event also served as controls. Cases were slightly younger than controls (44 ± 19.9 years vs. 48.6 ± 17.7 years, respectively) (Table 2). Cases were more likely to be women (62.4% vs. 50.7% of controls) and to have type 1 diabetes (89.1% vs. 76.1% of controls). Similar to patients in the hypoglycemia study, most patients in the hyperglycemia cohort were white (67.3%), but there was no notable difference in race/ethnicity distribution between cases and controls.

Table 2.

Characteristics of 1971 Insulin-Treated Diabetes Patients Who Used a Continuous Glucose Monitor in the Hyperglycemia Case–Control Study

Characteristic	Category	ALL (n = 1971)	Hyperglycemia cases ^a (n = 165)	Controls (n = 1971)	P ^b
Age, years, mean (SD)		48.6 (17.7)	44 (19.9)	48.6 (17.7)	<0.01
Sex, n (%)	Women	999 (50.7)	103 (62.4)	999 (50.7)	<0.01
Sex, n (%)	Men	972 (49.3)	62 (37.6)	972 (49.3)
Diabetes type, n (%)	Type 1	1500 (76.1)	147 (89.1)	1500 (76.1)	<0.01
Diabetes type, n (%)	Type 2	471 (23.9)	18 (10.9)	471 (23.9)
Race/ethnicity, n (%)	White	1306 (67.3)	114 (69.1)	1306 (67.3)	0.44
	African American	128 (6.6)	14 (8.5)	128 (6.6)	0.22
	Latino	248 (12.8)	17 (10.3)	248 (12.8)	0.40
	Asian	166 (8.6)	8 (4.8)	166 (8.5)	0.20
	Other^c	94 (4.8)	12 (7.3)	94 (4.8)	0.06

P-values from generalized estimating equations to account for nonindependence between the cases and controls.

Includes Native American, Pacific Islander and multiracial. Missing race (n = 41) not shown in table.

Hypoglycemia events

Having no CGM data more than doubled the odds of a hypoglycemia event in the 0–7-day lookback (OR 2.77; 95% CI 1.76–4.36) and the 8–14-day lookback (OR 2.05; 95% CI 1.36–3.10) compared to having any CGM data (Table 3). Similarly, CGM active time <70% in the 0–7-day lookback was associated with greater odds of hypoglycemia (OR 1.67; 95% CI 1.11–2.5) compared to CGM active time ≥70%. There was also a trend, although nonsignificant, toward increased odds of hypoglycemia in patients with CGM active time <70% in the 8–14-day lookback (OR 1.40; NS).

Table 3.

Associations Between Nine Categorical Standardized Continuous Glucose Monitor Metrics and Severe Hypoglycemia Events for Two Lookback Periods

Metric	Description	Lookback period 0–7 days				Lookback period 8–14 days
Metric	Description	Cases, N (%) ^a	Controls, N (%) ^a	Odds ratio (95% CI)	C-statistic	Cases, N (%) ^a	Controls, N (%) ^a	Odds ratio (95% CI)	C-statistic
All patients in hypoglycemia case–control cohort
Days with CGM data during lookback period	0 days	53 (34.4)	553 (16.8)	2.77 (1.76–4.36)^***	0.59	64 (41.6)	883 (26.8)	2.05 (1.36–3.10)^**	0.57
Days with CGM data during lookback period	1+ days (ref)	101 (65.6)	2743 (83.2)	—	—	90 (58.4)	2413 (73.2)	—	—
Percentage of time CGM is active during lookback period	<70%	82 (53.2)	1312 (39.8)	1.67 (1.11–2.50)^*	0.57	78 (50.6)	1376 (41.7)	1.40 (0.92–2.12)^ns	0.55
Percentage of time CGM is active during lookback period	≥70% (ref)	72 (46.8)	1984 (60.2)	—	—	76 (49.4)	1920 (58.3)	—	—
Subset of patients in hypoglycemia case–control cohort who have CGM data available > = 70% of the time
GMI	<7	28 (38.9)	654 (33.0)	1.37 (0.81–2.33)^ns	0.53	29 (38.2)	604 (31.5)	1.36 (0.84–2.19)^ns	0.53
GMI	≥7 (ref)	44 (61.1)	1330 (67.0)	—	—	47 (61.8)	1316 (68.5)	—	—
Glycemic variability: %CV	High (>36%)	42 (58.3)	798 (40.2)	2.09 (1.11–3.95)^*	0.59	45 (59.2)	741 (38.6)	2.31 (1.49–3.58)^***	0.60
Glycemic variability: %CV	Target ≤36% (ref)	30 (41.7)	1186 (59.8)	—	—	31 (40.8)	1179 (61.4)	—	—
TAR: % of readings >250 mg/dL (level 2)	TAR ≥5%	47 (65.3)	1334 (67.2)	0.85 (0.47–1.52)^ns	0.51	54 (71.1)	1301 (67.8)	1.17 (0.71–1.91)^ns	0.52
TAR: % of readings >250 mg/dL (level 2)	TAR <5% (ref)	25 (34.7)	650 (32.8)	—	—	22 (29.0)	619 (32.2)	—	—
TAR: % of readings >180 mg/dL (level 1)	TAR ≥25%	51 (70.8)	1431 (72.1)	0.87 (0.52–1.46)^ns	0.51	50 (65.8)	1401 (73)	0.68 (0.42–1.09)^ns	0.54
TAR: % of readings >180 mg/dL (level 1)	TAR <25% (ref)	21 (29.2)	553 (27.9)	—	—	26 (34.2)	519 (27.0)	—	—
TIR: % of readings 70–180 mg/dL (in range)	TIR <70%	57 (79.2)	1414 (71.3)	1.46 (0.74–2.91)^ns	0.54	57 (75.0)	1380 (71.9)	1.16 (0.72–1.86)^ns	0.52
TIR: % of readings 70–180 mg/dL (in range)	TIR ≥70% (ref)	15 (20.8)	570 (28.7)	—	—	19 (25.0)	540 (28.1)	—	—
TBR: % of readings <70 mg/dL (level 1)	TBR ≥4%	27 (37.5)	430 (21.7)	2.18 (1.16–4.11)^*	0.58	26 (34.2)	417 (21.7)	1.87 (1.15–3.04)^*	0.56
TBR: % of readings <70 mg/dL (level 1)	TBR <4% (ref)	45 (62.5)	1554 (78.3)	—	—	50 (65.8)	1503 (78.3)	—	—
TBR: % of readings <54 mg/dL (level 2)	TBR ≥1%	24 (33.3)	354 (17.8)	2.12 (1.12–4.02)^*	0.58	27 (35.5)	357 (18.6)	2.40 (1.44–3.99)^**	0.59
TBR: % of readings <54 mg/dL (level 2)	TBR <1% (ref)	48 (66.7)	1630 (82.2)	—	—	49 (64.5)	1563 (81.4)	—	—

See Supplementary Table S1 for a complete description of each metric.

P < 0.05, ^** P < 0.01, ^*** P < 0.001.

Number of observations (not patients) with each metric (column 2). Column percentages were calculated using the total number of hypoglycemia or control events (not patients) as the denominator. This repeated measures design allows patients to be selected as a case (i.e., experiencing a hypoglycemic event) or control multiple times. In each analysis, a patient could be a control until becoming a case, but after that were permitted to be only a case.

CGM, continuous glucose monitor; CI, confidence interval; CV, coefficient of variation; GMI, glucose management indicator; TAR, time above range; TBR, time below range; TIR, time in range.

In the primary analysis, several standard CGM metrics were strongly and significantly associated with higher odds of hypoglycemia (Table 3). Hypoglycemia events were associated with these standard CGM metrics in the 0–7-day lookback: ≥4% of readings <70 mg/dL (level 1 time below range) (OR 2.18; 95% CI 1.16–4.11), ≥1% of readings <54 mg/dL (level 2 time below range) (OR 2.12; 95% CI 1.12–4.02), and glycemic variability (%CV) >36% (OR 2.09; 95% CI 1.11–3.95). Hypoglycemic events had similar associations with the same metrics in the 8–14-day lookback. Among the nonstandard metrics, LBGI and hypoglycemia had the strongest association with hypoglycemic events in both lookback periods (Supplementary Table S2).

Despite the significant associations, all standard and nonstandard CGM metrics had poor discrimination (all C-statistics <0.7) (Table 3 and Supplementary Table S2), and thus could not reliably predict hypoglycemic events in either lookback period.

Hyperglycemia events

Having no CGM data more than tripled the odds of a hyperglycemia event compared to having any CGM data; this was true for both the 0–7-day lookback (OR 4.16; 95% CI 2.59–6.70) and the 8–14-day lookback (OR 3.27; 95% CI 2.06–5.20) (Table 4). CGM active time <70% was associated with approximately three times the odds of hyperglycemia compared to CGM active time ≥70% for both the 0–7-day lookback (OR 3.01; 95% CI 1.95–4.66) and the 8–14-day lookback (OR 2.62; 95% CI 1.69–4.08).

Table 4.

Associations Between Nine Categorical Standardized Continuous Glucose Monitor Metrics and Severe Hyperglycemia Events for Two Lookback Periods

Metric	Description	Lookback period 0–7 days				Lookback period 8–14 days
Metric	Description	Cases, N (%) ^a	Controls, N (%) ^a	Odds ratio (95% CI)	C-statistic	Cases, N (%) ^a	Controls, N (%) ^a	Odds ratio (95% CI)	C-statistic
All patients in hyperglycemia case–control cohort
Days with CGM data during lookback period	0 days	185 (55.2)	777 (21.7)	4.16 (2.59–6.70)^***	0.67	195 (58.2)	1050 (29.4)	3.27 (2.06–5.20)^***	0.64
Days with CGM data during lookback period	1+ days (ref)	150 (44.8)	2800 (78.3)	—	—	140 (41.8)	2527 (70.7)	—	—
Percentage of time CGM is active during lookback period	<70%	239 (71.3)	1550 (43.3)	3.01 (1.95–4.66)^***	0.64	233 (69.6)	1601 (44.8)	2.62 (1.69–4.08)^***	0.62
Percentage of time CGM is active during lookback period	≥70% (ref)	96 (28.7)	2027 (56.7)	—	—	102 (30.5)	1976 (55.2)	—	—
Subset of patients in hyperglycemia case–control cohort, with CGM data available ≥70% of the time
GMI	≥9	37 (38.5)	245 (12.1)	4.68 (2.79–7.83)^***	0.63	29 (28.4)	220 (11.1)	3.27 (1.92–5.58)^***	0.59
GMI	<9 (ref)	59 (61.5)	1782 (87.9)	—	—	73 (71.6)	1756 (88.9)	—	—
Glycemic variability: %CV	High (>36%)	49 (51.0)	718 (35.4)	2.11 (1.28–3.48)^***	0.58	42 (41.2)	734 (37.1)	1.16 (0.71–1.89)^ns	0.52
Glycemic variability: %CV	Target ≤36% (ref)	47 (49.0)	1309 (64.6)	—	—	60 (58.8)	1242 (62.9)	—	—
TAR: % of readings >250 mg/dL (level 2)	TAR ≥5%	89 (92.7)	1501 (74.1)	5.20 (2.02–13.36)^**	0.59	88 (86.3)	1474 (74.6)	2.12 (1.15–3.91)^*	0.56
TAR: % of readings >250 mg/dL (level 2)	TAR <5% (ref)	7 (7.3)	526 (26.0)	—	—	14 (13.7)	502 (25.4)	—	—
TAR: % of readings >180 mg/dL (level 1)	TAR ≥25%	90 (93.8)	1627 (80.3)	4.20 (1.54–11.44)^**	0.57	89 (87.3)	1576 (79.8)	1.66 (0.80–3.44)^ns	0.54
TAR: % of readings >180 mg/dL (level 1)	TAR <25% (ref)	6 (6.3)	400 (19.7)	—	—	13 (12.8)	400 (20.2)	—	—
TIR: % of readings 70–180 mg/dL (in range)	TIR <70%	91 (94.8)	1574 (77.7)	5.98 (2.12–16.83)^**	0.59	87 (85.3)	1532 (77.5)	1.69 (0.88–3.25)^ns	0.54
TIR: % of readings 70–180 mg/dL (in range)	TIR ≥70% (ref)	5 (5.2)	453 (22.4)	—	—	15 (14.7)	444 (22.5)	—	—
TBR: % of readings <70 mg/dL (level 1)	TBR ≥4%	11 (11.5)	320 (15.8)	0.65 (0.31–1.39)^ns	0.52	16 (15.7)	319 (16.1)	0.88 (0.47–1.67)^ns	0.50
TBR: % of readings <70 mg/dL (level 1)	TBR <4% (ref)	85 (88.5)	1707 (84.2)	—	—	86 (84.3)	1657 (83.9)	—	—
TBR: % of readings <54 mg/dL (level 2)	TBR ≥1%	10 (10.4)	275 (13.6)	0.67 (0.31–1.47)^ns	0.52	17 (16.7)	259 (13.1)	1.27 (0.64–2.53)^ns	0.52
TBR: % of readings <54 mg/dL (level 2)	TBR <1% (ref)	86 (89.6)	1752 (86.4)	—	—	85 (83.3)	1717 (86.9)	—	—

See Supplementary Table S1 for a complete description of each metric.

P < 0.05, ^** P < 0.01, ^*** P < 0.001.

Number of observations (not patients) with each metric (column 2). Column percentages were calculated using the total number of hypoglycemia or control events (not patients) as the denominator. The repeated measures design allows patients to be selected as a case (i.e., experiencing a hypoglycemic event) or control multiple times. In each analysis, a patient could be a control until becoming a case, but after that were permitted to be only a case.

Several CGM metrics were strongly and significantly associated with higher odds of hyperglycemia. In the 0–7-day lookback, all standard CGM metrics designed to measure high glycemia, including GMI >9 (OR 4.68; 95% CI 2.79–7.83), ≥5% of readings >250 mg/dL (level 2 time above range) (OR 5.20; 95% CI 2.02–13.36), ≥25% of readings >180 mg/dL (level 1 time above range) (OR 4.20; 95% CI 1.54–11.44), and <70% of readings 70–180 mg/dL (TIR) (OR 5.98; 95% CI 2.12–16.83), were associated with a hyperglycemic event (Table 4).

In the 8–14-day lookback, only GMI >9 (OR 3.27; 95% CI 1.92–5.58) and ≥5% of readings >250 mg/dL (level 2 time above range) (OR 2.12; 95% CI 1.15–3.91) were significantly associated with a hyperglycemic event. Compared to patients with low glycemic variability, having higher glycemic variability (>36% CV) was associated with twice the odds of experiencing a hyperglycemic event in the 0–7-day lookback (OR 2.11; 95% CI 1.28–3.48), but not in the 8–14-day lookback (OR 1.16; 95% CI 0.71–1.89). Among the nonstandard metrics, hyperglycemia index had the strongest association with hyperglycemic events in both lookback periods (Supplementary Table S3).

Using cutoffs determined by the CGM consensus panel, none of the standard metrics had good discrimination (i.e., a C-statistic ≥0.7).

Secondary analyses: CGM metrics (standard and nonstandard) as continuous variables

In secondary analyses, several standard and nonstandard CGM metrics as continuous variables were significantly associated with a dysglycemic event (hypoglycemia or hyperglycemia, online Supplementary Tables S2 and S3). Of note, some metrics defined as continuous variables achieved good discrimination (C-statistic >0.7). However, continuous metrics are not practical to use in an implementation strategy due to the lack of defined thresholds at which clinical action should be taken by the clinician.

Discussion

In this study of patients with insulin-treated diabetes, several CGM metrics in two lookback periods (0–7 or 8–14 days before the event) were significantly associated with the risk of ED visits or hospitalizations for hypoglycemia or hyperglycemia. The most notable finding of this study was that, for both hypoglycemia and hyperglycemia events, having no CGM data in the lookback period or having CGM active time ≤70% was a factor most strongly associated with an increased risk of these events.

As might be expected, hypoglycemic events were associated with hypoglycemia-related metrics (e.g., level 1 and level 2 time below range), as well as high glycemic variability (%CV) in both the 0–7- and 8–14-day lookback periods. The metrics most strongly associated with a hypoglycemic event in the 0–7 days before the event included level 1 time below range ≥4%, level 2 time below range ≥1%, and %CV >36%.

The finding that these metrics were associated in both lookback periods suggests that the value of these hypoglycemia metrics in assigning risk may be stable over time, reflecting a “high-risk” CGM profile, rather than acute near-event changes in metrics associated with the imminent events themselves. None of the hyperglycemia-related metrics or CGM metrics that summarize average glycemia (e.g., GMI) was associated with hypoglycemic events, which is consistent with the lack of association between specific A1c levels and hypoglycemia risk.²⁵ Similarly, TIR was also not associated with hypoglycemia risk.

CGM metrics that summarize average glycemia (e.g., GMI, TIR) as well as glycemic variability (%CV) were associated with hyperglycemic events in the 0–7-day lookback period, in addition to the time above range metrics. However, only GMI >9% and level 2 time above range (>250 mg/dL) ≥5% in the 8–14-day lookback were associated with imminent hyperglycemic events. Thus, while higher glycemic variability and less TIR may only have utility as red flags just before (i.e., <7 days) the occurrence of a hyperglycemic event, GMI and time above range are more consistently informative.

While there were significant associations of CGM hypoglycemia or hyperglycemia metrics with ED visits and hospitalizations, none of the standardized CGM metrics that is commonly reported on the AGP had adequate discrimination (i.e., able to differentiate between those who will versus will not experience an event) to reliably predict either outcome with sufficient warning to mobilize an intervention (i.e., 8–14 days in advance).²⁶ Therefore, while some CGM metrics are strongly associated with risk, an outpatient, population-based care management strategy using an “early warning” signal based on single CGM metrics to identify patients at imminent risk of hypoglycemic or hyperglycemic events would likely not be an effective or efficient use of resources.

In other words, among patients who end up in the ED or hospital for a dysglycemic event, ∼100% would be expected to have hypoglycemia or hyperglycemia detected on their CGM. However, among patients with hypoglycemia or hyperglycemia detected on CGM, few end up in the ED or hospital. Thus, CGM hypoglycemia and hyperglycemia metrics are strongly associated, but not sufficiently predictive of events. Whether or not a patient ends up at the ED is certainly due to other factors that cannot be gleaned from CGM data alone (e.g., unforeseeable behavioral changes such as skipping meals, patient support at home, whether the patient has glucagon, and utilization preferences of the patient).

While risk signals in the 0–7-day lookback period would unlikely provide adequate advance warning for an outpatient provider to intervene on rapidly deteriorating glycemia or immediate causes of dysglycemia (e.g., meal skipping or missed insulin) to avert an event, this time frame may work well for averting glycemic deterioration in settings in which closer CGM monitoring is possible, for example, in an inpatient ward or a nursing home.

This study is one of the largest studies linking CGM metrics to risk of ED visits or hospitalizations for hypoglycemia or hyperglycemia and included a racially and ethnically diverse sample of ambulatory diabetes patients. That said, our study has some limitations worth noting. We do not know whether missing CGM data during a given lookback period were due to nonuse of the CGM device, device failure, or simply failure by the patient to upload the data. However, only nonuse or device failure provides a likely explanation for the adverse outcomes associated with missing data. In addition, data was more likely to be missing among patients who used reader devices rather than smartphones (which more easily provide automatic, continuous uploads). This could lead to confounding by socioeconomic status or digital literacy, which we were unable to assess.

The thresholds defined by the CGM consensus groups, used in this study, were determined largely based on expert opinion, and different (evidence-based) thresholds might change our findings. As recommended for observational studies of this type, we did not adjust P-values for multiple comparisons, and thus findings should be considered exploratory.²⁷ While the study is large and does include some racial and ethnic diversity, the frequency of racial or ethnic minorities in this study population is considerably lower than in the underlying patient population, which is consistent with disparities shown in other studies of CGM initiation and use.^28,29 We noted that significantly more African Americans were cases than controls (15.9% vs. 7.3%) in the hypoglycemia study, consistent with our previous work, which found higher rates of hypoglycemia-related ED visits and hospitalizations among African Americans compared to other racial/ethnic groups.³⁰

More exploration is needed to establish empirically optimized metric thresholds and whether these would provide better discrimination than thresholds set by the CGM consensus group. Future research is needed to explore whether machine learning models that incorporate some of the strongest predictors mentioned here and their interactions (e.g., high glycemic variability × time below range) or other newer nonstandard metrics such as the glycemia risk index,³¹ along with other demographic, clinical, and utilization variables, would be able to accurately predict hypoglycemic or hyperglycemic events with adequate discrimination to justify implementing an advanced warning system.

In conclusion, this study of patients with insulin-treated diabetes found that missing (or sparse) CGM data and several CGM metrics were strongly associated with increased risk of ED visits or hospitalizations for hypoglycemia or hyperglycemia. These metrics could be used to identify patients who may be at elevated risk of dysglycemic events and need further clinical attention. However, discrimination of individual metrics was poor (C-statistics <0.7). Thus, single CGM metrics available in the AGP alone are not sufficiently reliable to support a population-based advance warning strategy of targeting patients at high risk of hypoglycemia or hyperglycemia.

Footnotes

Author Disclosure Statement

No competing financial interests exist.

Funding Information

This study was funded by Kaiser Permanente Northern California Community Health. National Institute of Aging (R01 AG063391) also provided additional funding for Andrew J. Karter, Melissa M. Parker, and Howard H. Moffet, and National Institute of Diabetes and Digestive and Kidney Diseases (P30 DK092924) provided additional funding for Andrew J. Karter.

Supplementary Material

Supplementary Data

Supplementary Figure S1

Supplementary Table S1

Supplementary Table S2

Supplementary Table S3

References

Stratton

, Adler

, Neil

, et al. Association of glycaemia with macrovascular and microvascular complications of type 2 diabetes (UKPDS 35): Prospective observational study. BMJ, 2000; 321(7258):405–412.

Diabetes Control and Complications Trial Research Group; Nathan

, Genuth

, et al. The effect of intensive treatment of diabetes on the development and progression of long-term complications in insulin-dependent diabetes mellitus. N Engl J Med, 1993; 329(14):977–86; doi: 10.1056/NEJM199309303291401

The Diabetes Control and Complications Trial (DCCT). Design and methodologic considerations for the feasibility phase. The DCCT Research Group. Diabetes, 1986; 35:530–545.

McCoy

, Galindo

, Swarna

, et al. Sociodemographic, clinical, and treatment-related factors associated with hyperglycemic crises among adults with type 1 or type 2 diabetes in the US from 2014 to 2020. JAMA Netw Open, 2021; 4(9):e2123471; doi: 10.1001/jamanetworkopen.2021.23471

Intensive blood-glucose control with sulphonylureas or insulin compared with conventional treatment and risk of complications in patients with type 2 diabetes (UKPDS 33). UK Prospective Diabetes Study (UKPDS) Group. Lancet, 1998; 352(9131):837–853.

Barr

. Retinopathy and nephropathy in patients with type 1 diabetes four years after a trial of intensive therapy. The Diabetes Control and Complications Trial/Epidemiology of Diabetes Interventions and Complications Research Group. N Engl J Med, 2000; 342:381–389.

Hypoglycemia in the Diabetes Control and Complications Trial. The Diabetes Control and Complications Trial Research Group. Diabetes, 1997; 46(2):271–86.

Huang

, Liu

, Moffet

, et al. Glycemic control, complications, and death in older diabetic patients: the diabetes and aging study. Research Support, N.I.H., Extramural. Diabetes Care, 2011; 34(6):1329–36; doi: 10.2337/dc10-2377

McCoy

, Lipska

, Yao

, et al. Intensive treatment and severe hypoglycemia among adults with type 2 diabetes. JAMA Intern Med, 2016; 176(7):969–78; doi: 10.1001/jamainternmed.2016.2275

10.

Karter

, Parker

, Moffet

, et al. Association of real-time continuous glucose monitoring with glycemic control and acute metabolic events among patients with insulin-treated diabetes. JAMA, 2021; 325(22):2273–2284; doi: 10.1001/jama.2021.6530

11.

Reaven

, Newell

, Rivas

, et al. Initiation of continuous glucose monitoring is linked to improved glycemic control and fewer clinical events in type 1 and type 2 diabetes in the Veterans Health Administration. Diabetes Care, 2023; 46(4):854–863; doi: 10.2337/dc22-2189

12.

Oliver

, Gimenez

, Calhoun

, et al. Continuous glucose monitoring in people with type 1 diabetes on multiple-dose injection therapy: The relationship between glycemic control and hypoglycemia. Diabetes Care, 2020; 43(1):53–58; doi: 10.2337/dc19-0977

13.

Monnier

, Wojtusciszyn

, Molinari

, et al. Respective contributions of glycemic variability and mean daily glucose as predictors of hypoglycemia in type 1 diabetes: Are they equivalent?. Diabetes Care, 2020; 43(4):821–827; doi: 10.2337/dc19-1549

14.

Toschi

, Slyne

, Sifre

, et al. The relationship between CGM-derived metrics, A1C, and risk of hypoglycemia in older adults with type 1 diabetes. Diabetes Care, 2020; 43(10):2349–2354; doi: 10.2337/dc20-0016

15.

Scheinker

, Gu

, Grossman

, et al. Algorithm-enabled, personalized glucose management for type 1 diabetes at the population scale: Prospective evaluation in clinical practice. JMIR Diabetes, 2022; 7(2):e27284; doi: 10.2196/27284

16.

Ferstad

, Vallon

, Jun

, et al. Population-level management of type 1 diabetes via continuous glucose monitoring and algorithm-enabled patient prioritization: Precision health meets population health. Pediatr Diabetes, 2021; 22(7):982–991; doi: 10.1111/pedi.13256

17.

Karter

, Warton

, Lipska

, et al. Development and validation of a tool to identify patients with type 2 diabetes at high risk of hypoglycemia-related emergency department or hospital use. Multicenter study validation studies. JAMA Intern Med, 2017; 177(10):1461–1470; doi: 10.1001/jamainternmed.2017.3844

18.

Schlesselman

JJ.

Case-Control Studies: Design, Conduct, Analysis. Monographs in Epidemiology and Biostatistics. Oxford University Press: New York: Oxford, 1982.

19.

Robins

, Gail

, Lubin

. More on “Biased selection of controls for case-control analyses of cohort studies.”. Biometrics, 1986; 42(2):293–299.

20.

Battelino

, Danne

, Bergenstal

, et al. Clinical targets for continuous glucose monitoring data interpretation: Recommendations from the International Consensus on Time in Range. Diabetes Care, 2019; 42(8):1593–1603; doi: 10.2337/dci19-0028

21.

Danne

, Nimri

, Battelino

, et al. International consensus on use of continuous glucose monitoring. Diabetes Care, 2017; 40(12):1631–1640; doi: 10.2337/dc17-1600

22.

Xing

, Kollman

, Beck

, et al. Optimal sampling intervals to assess long-term glycemic control using continuous glucose monitoring. Diabetes Technol Ther, 2011; 13(3):351–358; doi: 10.1089/dia.2010.0156

23.

Broll

, Urbanek

, Buchanan

, et al. Interpreting blood GLUcose data with R package iglu. PLoS One, 2021; 16(4):e0248560; doi: 10.1371/journal.pone.0248560

24.

Diggle

, Liang

, Zeger

. Analysis of Longitudinal Data. Oxford University Press: Oxford, 1994.

25.

Lipska

, Warton

, Huang

, et al. HbA1c and risk of severe hypoglycemia in type 2 diabetes: The Diabetes and Aging Study. Research Support, N.I.H., Extramural. Diabetes Care, 2013; 36(11):3535–42; doi: 10.2337/dc13-0610

26.

Varga

, Niss

, Estampador

, et al. Association is not prediction: A landscape of confused reporting in diabetes—A systematic review. Diabetes Res Clin Pract, 2020; 170:108497; doi: 10.1016/j.diabres.2020.108497

27.

Rothman

. No adjustments are needed for multiple comparisons. Epidemiology, 1990; 1:43–46.

28.

Lai

, Lipman

, Willi

, et al. Early racial/ethnic disparities in continuous glucose monitor use in pediatric type 1 diabetes. Diabetes Technol Ther, 2021; 23(11):763–767; doi: 10.1089/dia.2021.0120

29.

Lai

, Lipman

, Willi

, et al. Racial and ethnic disparities in rates of continuous glucose monitor initiation and continued use in children with type 1 diabetes. Diabetes Care, 2021; 44(1):255–257; doi: 10.2337/dc20-1663

30.

Karter

, Lipska

, O'Connor

, et al. High rates of severe hypoglycemia among African American patients with diabetes: the surveillance, prevention, and Management of Diabetes Mellitus (SUPREME-DM) network. J Diabetes Complications, 2017; 31(5):869–873; doi: 10.1016/j.jdiacomp.2017.02.009

31.

Klonoff

, Wang

, Rodbard

, et al. A glycemia risk index (GRI) of hypoglycemia and hyperglycemia for continuous glucose monitoring validated by clinician ratings. J Diabetes Sci Technol, 2023; 17(5):1226–1242; doi: 10.1177/19322968221085273

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

1.12 MB