Agreement Between Fingerstick Blood Glucose and Continuous Glucose Monitor Measures Among Long-Term Care Facility Residents

Abstract

Background:

Long-term care facility (LTCF) residents with diabetes are at high risk of hypoglycemia. Continuous glucose monitoring (CGM), which measures interstitial glucose at 5-min intervals over 10–14 days, and fingerstick blood glucose (FBG) which analyzes glucose from a drop of blood, are both used to monitor glucose levels. Observational studies using electronic health record (EHR) data containing FBG measures could help to identify ways to reduce hypoglycemia risk. We first need to understand the validity of such data. Our objective was to compare EHR-based FBG measures against reference-standard CGM measures of hypoglycemia.

Methods:

We studied two cohorts of residents with diabetes in parallel. In Cohort 1, we analyzed linked CGM and Long-Term Care Data Cooperative EHR-based FBG data collected in 2023. In Cohort 2, we analyzed linked CGM and EHR-based FBG data obtained directly from LTCFs between 2022 and 2023. We defined hypoglycemia as glucose <70 mg/dL and assessed the sensitivity and specificity of FBG versus CGM measures to detect hypoglycemia. The unit of analysis was each pair of contemporaneous FBG-CGM measures.

Results:

In Cohort 1, two White female residents with a mean (standard deviation [SD]) age of 81 [12.7] years generated 25 daily hypoglycemia measurements. The sensitivity and specificity were 14% and 100%, respectively, for FBG-measured hypoglycemia. Cohort 2 included 40 residents (mean [SD] age 68 [11] years, 45% females, 60% White race) who generated 425 daily measurements of hypoglycemia. The sensitivity and specificity were 13% and 99%, respectively.

Conclusion:

EHR FBG measures of hypoglycemia had high specificity but failed to identify four out of every five hypoglycemic events among LTCF residents. Researchers and healthcare providers should assume hypoglycemia is measured with substantial errors in EHRs and account for this in their research and clinical practice.

1 |. Introduction

Hypoglycemia is clinically defined as blood glucose levels <70 mg per deciliter (mg/dL) [1]. It is a highly prevalent and clinically significant condition among long-term care facility (LTCF) residents with type 2 diabetes [2, 3]. In an observational study comprising 1409 LTCF residents, 42% experienced at least one episode of blood glucose <70 mg/dL, and 7% experienced blood glucose of <40 mg/dL [4]. These findings, along with other studies, indicate that intensive glucose control is common in LTCFs, increasing hypoglycemia risk, particularly among those treated with high-risk glucose-lowering agents such as insulin and sulfonylureas [4–6]. Individuals with diabetes and cognitive impairments [7], such as Alzheimer’s disease and related dementias (ADRD), face an even greater risk of hypoglycemia because they may be unable to recognize symptoms or report them to caregivers [2]. Hypoglycemia has been linked to falls, cardiovascular events, emergency department visits, hospitalizations [8], and even death [9], underscoring the urgent need for improved glucose monitoring strategies in LTCFs.

Fingerstick blood glucose (FBG) measurements, a type of point-of-care capillary blood glucose measure, are the standard method used to monitor glucose levels among LTCF residents [2]. FBG involves pricking a finger with a lancet, collecting a small blood sample, and analyzing it using a reflectance photometer [10]. These measurements are incorporated into LTCF electronic health records (EHRs), making them widely available for clinical and research use. Typically performed before meals and at bedtime, FBG aligns with meal intake and insulin administration but may be conducted as often as every 4–6 h for patients not consuming food or receiving enteral feeding [11]. However, FBG has several limitations: (1) it captures only single time points, potentially missing asymptomatic or nocturnal hypoglycemia [12], (2) it requires frequent finger pricks, which may be painful for residents [13, 14] and burdensome for staff [14, 15], and (3) it provides limited insight into overall glycemic patterns, particularly in frail older adults at high risk of hypoglycemia.

Given these limitations, continuous glucose monitoring (CGM) devices have emerged as a more comprehensive tool for assessing glycemic patterns, particularly in those with comorbid diabetes and dementia [16]. CGM devices measure interstitial glucose levels in real-time, typically every 5–15 min and over 10–14 days, using a sensor worn on the body [17–19]. This continuous data collection enables the detection of nocturnal and asymptomatic hypoglycemia, providing a more complete understanding of glucose fluctuations than FBG [20, 21]. CGM may be particularly beneficial for older adults with diabetes and impaired cognition who require caregiver support for glucose monitoring [22]. Moreover, CGMs reduce patient burden by eliminating frequent fingersticks and lessen staff workload by automating glucose data collection [23].

Despite its widespread use in LTCFs, FBG may underestimate the true frequency and severity of hypoglycemia compared with CGM, which can be considered a more sensitive reference standard for detecting glucose fluctuations. Although some studies have examined the differences in the rate of hypoglycemia detection comparing FBG to CGM [24–26], more research is needed to fully understand how much FBG underestimates hypoglycemia relative to CGM, particularly in LTCFs. Given the unique health challenges of LTCF residents with multiple comorbidities and polypharmacy, it is critical to ensure that current monitoring practices accurately capture the true burden of hypoglycemia. Therefore, our objective was to compare FBG and CGM in detecting hypoglycemia among LTCF residents and to assess the extent to which FBG may misclassify or fail to capture clinically significant hypoglycemia events.

2 |. Methods

2.1 |. Study Design, Population, and Follow-Up Period

We studied two cohorts of LTCF residents in parallel. Residents or their legally authorized representatives provided consent to undergo continuous glucose monitoring for approximately 2 weeks using the Dexcom G6 Pro professional CGM. The Dexcom G6 Pro consists of a thread-like sensor that was applied in a minimally invasive fashion on the resident’s upper arm or abdomen by our research team. The sensor measured glucose values of the interstitial fluid every 5 min [27], and these data were stored on the waterproof transmitter sitting atop the residents’ skin. The masked sensor does not require any input from residents and thus is ideal for individuals with cognitive impairment. The sensor remained on the residents’ skin like an adhesive bandage and was removed by peeling it off at the end of the assessment period. The data from the sensor was later uploaded to a cloud-based HIPAA-compliant platform by a member of our research team. Bluetooth was not enabled while the residents were using the devices. Throughout CGM use, residents and their providers were blinded to the CGM readings, so they were unaware of the glucose levels measured by the CGM.

Cohort 1 included residents with a diagnosis of Alzheimer’s Disease/Alzheimer’s Disease and Related Dementia (AD/ADRD) based on the Minimum Data Set (MDS) version 3.0 active diagnoses checkboxes I4200 (Alzheimer’s Disease) or I4800 (Non-Alzheimer’s Dementia) and the Cognitive Function Scale. In addition, residents were required to have a diagnosis of diabetes based on MDS active diagnosis checkbox I2900 (Diabetes Mellitus) [28]. Finally, residents had to have CGM and FBG readings generated on the same days. All CGM and FBG measures were collected in 2023 for Cohort 1.

Residents in Cohort 2 had a clinical diagnosis of diabetes in their medical records and used a glucose-lowering medication of any kind. They needed to have CGM and FBG readings on the same day. Cohort 2 residents were not required to have ADRD, unlike Cohort 1. All CGM and FBG measures were collected in 2022 and 2023 for Cohort 2.

For both cohorts, LTCFs were sampled from the same chain.

2.2 |. Data Sources

CGM data from the residents in Cohort 1 were linked to EHR data in the Long-Term Care Data Cooperative based on strict matching criteria comprising date of birth, sex, LTCF name, zip code, state, and calendar day of glucose measurements. We obtained information on LTCF residents’ demographics and FBG levels from the EHR and comorbid conditions from the MDS records. The MDS clinical assessments are performed on all LTCF residents upon admission to the LTCF, periodically, and upon discharge [29].

CGM and EHR-FBG records were obtained directly from the LTCFs in Cohort 2. Demographic information was collected from the MDS and EHR records.

Cohorts 1 and 2 were approved by the Institutional Review Boards at Advarra and the Joslin Diabetes Center, respectively. The statistical analysis code and supporting documentation are available on GitHub: https://doi.org/10.5281/zenodo.15058677.

2.3 |. Statistical Analysis

Hypoglycemia was defined as a glucose level below 70 mg/dL. We assessed the sensitivity, specificity, and positive and negative predictive values of FBG-measured hypoglycemia using CGM-measured hypoglycemia as the reference standard. We also examined the direction and magnitude of concordance between FBG and CGM measures using phi coefficients with associated 95% confidence intervals (CI) based on Fisher’s transformation. We interpreted the strength of the phi correlation coefficients using guidelines commonly applied to Pearson correlation coefficients, given the mathematical equivalence of the two when applied to binary variables. Specifically, we used the following thresholds: 0.00–0.10 (negligible), 0.10–0.39 (weak), 0.40–0.69 (moderate), 0.70–0.89 (strong), and 0.90–1.00 (very strong) [30]. We performed our validation analyses using 3 units of analysis: resident-level, calendar day-level, and 8-h interval-level.

For the resident-level analysis, we classified LTCF residents as having a hypoglycemic event if they had at least one glucose reading < 70 mg/dL during follow-up. Because Cohort 1 included only two residents, resident-level analyses were conducted for Cohort 2 only to ensure meaningful and interpretable estimates. In addition, we examined a scatterplot showing the total number of hypoglycemic events per resident detected by CGM (x-axis) versus those detected by FBG (y-axis) during follow-up.

For the day-level analyses, we classified residents as experiencing a hypoglycemic event on a given day if their minimum glucose value was below 70 mg/dL on that day.

We conducted two separate analyses using different 8-h interval windows. In the first 8-h interval analysis, we divided each day into three consecutive intervals of 12:00 a.m.—7:59 a.m., 8:00 a.m.—3:59 p.m., and 4:00 p.m.—11:59 p.m. and classified residents as hypoglycemic if the lowest glucose value during that interval was < 70 mg/dL.

We performed a stratified analysis, examining the performance of FBG and CGM within 8-h intervals of 10:00 p.m.—5:59 a.m., 6:00 a.m.—1:59 p.m., and 2:00 p.m.—9:59 p.m. to align with typical FBG assessment windows in Cohort 2. The intervals were constructed using the timing of FBG readings. Within each interval, the minimum FBG glucose value was selected, and the corresponding nearest CGM glucose value (regardless of whether it was the minimum in that interval or not) was identified. These paired values were used to determine whether a resident experienced any hypoglycemic event during a given interval.

For the day-and overall interval-level analyses, we reported statistics on the difference in time between pairs of CGM-FBG measures in two ways: (1) the difference in time between the CGM and FBG measures with the lowest blood glucose values on a given calendar day or interval, and (2) the difference in time between the FBG measure with the lowest glucose value on a given calendar day or interval and the nearest CGM measure in time (regardless of the CGM glucose value, i.e., not necessarily the lowest).

2.4 |. Software

Analyses were conducted using R version 4.3.3 (R Foundation for Statistical Computing, Vienna, Austria) and SAS version 9.4 (SAS Institute, Cary, NC, USA).

3 |. Results

3.1 |. Study Sample Characteristics

In Cohort 1, two White female residents (mean [standard deviation{SD}] age: 81 [12.7] years) contributed 25 daily and 35 8-h interval glucose measurements. In Cohort 2, 40 residents (mean [SD] age: 68 [11] years, 45% females, 60% White race) contributed 425 daily and 792 8-h interval measurements of glucose.

The mean (SD) number of FBG readings per resident was 28.1 (13), ranging from 9 to 59, compared to 2739.5 (230.8) CGM readings per resident, ranging from 1825 to 2856. In addition, the mean (SD) number of FBG readings per resident per day was 2.6 (1.3), with a range of 1–10, whereas the mean (SD) number of CGM readings per resident per day was 257.8 (64.9), ranging from 37 to 288.

3.2 |. Hypoglycemia Prevalence

Overall, the prevalence of hypoglycemia was 50% according to CGM and 10% according to FBG measurements in Cohort 2. At the daily measurement level, hypoglycemia was detected in 8% of FBG measures in Cohort 1, compared to 56% of CGM measures. Similarly, in Cohort 2, FBG detected hypoglycemia in 2% of daily measures, whereas CGM identified 9%. At the 8-h interval level, hypoglycemia was detected in 6% of FBG measures in Cohort 1 compared to 29% of CGM measures. In Cohort 2, FBG detected hypoglycemia in 1% of 8-h intervals, whereas CGM detected 5%.

3.3 |. Comparison of Resident-Level Hypoglycemia Detection

In Cohort 2, we observed a phi coefficient of 0.33 (95% CI 0.16–0.49), suggesting a weak correlation between resident-level EHR and CGM measures of hypoglycemia. The sensitivity of EHR-measured hypoglycemia was just 20%, whereas specificity was high at 100%. The overall agreement between the two glucose monitoring methods was 60% (Table 1). Figure 1 depicts a scatterplot of the total number of hypoglycemic events detected by CGM versus those detected by FBG at the resident level during follow-up. The majority of the points, each representing a unique resident, are clustered along the x-axis, illustrating that CGM identified multiple hypoglycemic events when FBG identified a couple or none.

TABLE 1 |.

Characteristics of fingerstick blood glucose measures of hypoglycemia versus continuous glucose monitoring measures in Cohort 2, resident-level measures (N = 40).

		CGM hypoglycemia
		Yes	No
FBG hypoglycemia	Yes	4	0	PPV 100%
	No	16	20	NPV 55.6%
		Sensitivity 20%	Specificity 100%	Percent overall agreement 60%

Abbreviations: CGM, continuous glucose monitoring; FBG, fingerstick blood glucose; N, number; NPV, negative predictive value; PPV, positive predictive value.

FIGURE 1 |.

Hypoglycemia event counts per resident: Fingerstick blood glucose versus continuous glucose monitor approaches. Abbreviations: CGM, continuous glucose monitoring; FBG, fingerstick blood glucose.

3.4 |. Comparison of Day-Level Hypoglycemia Detection

In Cohort 1, we observed a phi coefficient of 0.26 (95% CI 0.06–0.44), suggesting a weak correlation between daily EHR and CGM measures of hypoglycemia. The sensitivity of EHR-measured hypoglycemia was just 14%, but specificity was high at 100%. The overall agreement between the two glucose monitoring methods was 52% (Table 2).

TABLE 2 |.

Characteristics of fingerstick blood glucose measures of hypoglycemia versus continuous glucose monitoring measures in Cohort 1, day-level measures (N = 25).

		CGM hypoglycemia
		Yes	No
FBG hypoglycemia	Yes	2	0	PPV 100%
	No	12	11	NPV 48%
		Sensitivity 14%	Specificity 100%	Percent overall agreement 52%

Abbreviations: CGM, continuous glucose monitoring; FBG, fingerstick blood glucose; N, number; NPV, negative predictive value; PPV, positive predictive value.

In Cohort 2, we observed a phi coefficient of 0.27 (95% CI 0.10–0.44), suggesting a weak correlation between EHR-measured hypoglycemia and CGM-measured hypoglycemia at the day level. The sensitivity of EHR-measured hypoglycemia was just 13%, but specificity was high at 99%. The overall agreement was 91% (Table 3). The mean (SD) difference in time between pairs of the CGM and FBG measures with the lowest blood glucose values on a given calendar day was 338.6 (310.7) minutes. The mean (SD) difference in time between pairs of the FBG measure with the lowest glucose value and the nearest CGM measure in time (regardless of the glucose value, i.e., not necessarily the lowest) was 18.4 (68.1) minutes (Figure S1).

TABLE 3 |.

Characteristics of fingerstick blood glucose measures of hypoglycemia versus continuous glucose monitoring measures in Cohort 2, day-level measures (N = 425).

		CGM hypoglycemia
		Yes	No
FBG hypoglycemia	Yes	5	2	PPV 71%
	No	35	383	NPV 92%
		Sensitivity 13%	Specificity 99%	Percent overall agreement 91%

Abbreviations: CGM, continuous glucose monitoring; FBG, fingerstick blood glucose; N, number; NPV, negative predictive value; PPV, positive predictive value.

3.5 |. Comparison of 8-h Interval-Level Hypoglycemia Detection

3.5.1 |. Overall Analysis

Conducting analyses every 8 h as the discrete-time unit in Cohort 1, we observed a phi coefficient of 0.39 (95% CI 0.11–0.61), suggesting a weak correlation between EHR-measured hypoglycemia and CGM-measured hypoglycemia. The sensitivity was slightly higher than the day-level analysis, at 20%, whereas specificity remained 100% (Table S1). In Cohort 2, the phi correlation was 0.30 (95% CI 0.11–0.46), suggesting a weak correlation between EHR-measured hypoglycemia and CGM-measured hypoglycemia when using 8-h intervals as the unit of analysis. The sensitivity and specificity of EHR-measured hypoglycemia were 14% and 100%, respectively (Table S2). The mean (SD) difference in time between the lowest CGM and lowest FBG glucose measures within each 8-h interval was 158.9 (119.4) min. The mean (SD) difference in time between pairs of the FBG measure with the lowest glucose value and the nearest CGM measure in time (regardless of the glucose value, i.e., not necessarily the lowest) was 18.9 (71.6) min (Figure S1).

3.5.2 |. Stratified Analysis

The sensitivity and specificity of EHR-measured hypoglycemia were 67% and 99%, respectively, during the time interval 10:00 p.m.—5:59 a.m., with a phi correlation of 0.66 (95% CI 0.01–0.92). During the time interval 6:00 a.m.—1:59 p.m., the sensitivity, specificity, and phi correlation were 50%, 99%, and 0.49 (95% CI −0.02–0.80). FBG did not detect any hypoglycemic events between 2:00 p.m. and 9:59 p.m., whereas CGM identified two events during this interval. The sensitivity of FBG was 0% and the specificity was 100% (Table 4). The phi correlation was not possible to compute for the 2:00 p.m. to 9:59 p.m. interval. The mean (SD) difference in time between pairs of the FBG measure with the lowest glucose value and the nearest CGM measure in time in the intervals 10:00 p.m.—5:59 a.m., 6:00 a.m.—1:59 p.m., and 2:00 p.m.—9:59 p.m. was 6.7 (47.2), 19.1 (68.1) and 24.2 (82.3) min, respectively (Figure S2).

TABLE 4 |.

Characteristics of fingerstick blood glucose measures of hypoglycemia versus continuous glucose monitoring measures in Cohort 2 stratified by 8-h intervals.

		CGM hypoglycemia
		Yes	No
10:00 p.m.–5:59 a.m. (N = 151)
FBG Hypoglycemia	Yes	2	1	PPV 67%
	No	1	147	NPV 99%
		Sensitivity 67%	Specificity 99%	Percent overall agreement 99%
6:00 a.m.—1:59 p.m. (N = 327)
FBG hypoglycemia	Yes	2	2	PPV 50%
	No	2	321	NPV 99%
		Sensitivity 50%	Specificity 99%	Percent overall agreement 99%
2:00 p.m.—9:59 p.m. (N = 341)
FBG hypoglycemia	Yes	0	0	PPV 0%
	No	2	339	NPV 99%
		Sensitivity 0%	Specificity 100%	Percent overall agreement 99%

Abbreviations: CGM, continuous glucose monitoring; FBG, fingerstick blood glucose; N, number; NPV, negative predictive value; PPV, positive predictive value.

4 |. Discussion

Our results indicate that routinely collected EHR FBG measures of hypoglycemia have near-perfect specificity but low sensitivity among LTCF residents. The correlations between EHR FBG measures of hypoglycemia and CGM measures of hypoglycemia were slightly higher at the resident level and the 8-h interval level compared to the day level. FBG measures fail to identify approximately four out of every five hypoglycemic events among LTCF residents. When stratified by time of day, there were very few hypoglycemia events identified by either measure, rendering calculated diagnostic metrics unstable and less interpretable. Although EHR data improve upon Medicare claims data, which lack direct glycemic measures, our study highlights a critical limitation of EHR data—its low sensitivity in identifying hypoglycemia events. This limitation suggests that LTCF residents experience substantial undiagnosed and unaddressed hypoglycemia.

Two key factors likely explain the under-ascertainment of hypoglycemia by FBG. First, FBG tests are performed relatively infrequently compared to CGM, leading to missed events. This difference in sampling frequency is apparent in Cohort 2. Second, FBG monitoring does not capture nocturnal hypoglycemia, a common occurrence in older adults, because glucose levels are not typically measured while residents are asleep [31]. These factors contribute to an incomplete understanding of hypoglycemia risk in LTCF populations, with potential implications for clinical management and medication adjustments.

Our findings align with prior investigations among LTCF residents, which demonstrated significant underdiagnosis of hypoglycemia events by FBG monitoring compared to CGM [24–26]. Other studies [8, 32] similarly highlight the limitations of FBG-based monitoring for detecting hypoglycemia in older adults. These findings reinforce the need for caution when using EHR-based FBG data to assess hypoglycemia risk in LTCF residents. Relying exclusively on FBG data may lead to many missed hypoglycemia events among LTCF residents and opportunities to de-intensify glucose-lowering therapies. Our findings support the need to integrate more sensitive glucose monitoring methods in the LTCF setting, such as the unblinded, real-time CGM, which may help reduce the risk of hypoglycemia and allow appropriate treatment revisions and behavioral changes in a timely manner.

A major strength of our study is that LTCF residents and their providers were blinded to the CGM values, reducing the likelihood that FBG testing behaviors (e.g., timing or frequency) were influenced by glucose fluctuations. This allowed us to analyze natural trends and fluctuations in glucose levels. Additionally, we presented the results from two cohorts, which yielded nearly identical conclusions about the sensitivity and specificity of FBG-measured hypoglycemia, thereby bolstering the robustness of our inferences.

The primary limitation of our study is the relatively small sample size, particularly in Cohort 1, where only 2 of 9 residents with CGM had overlapping CGM and FBG data on the same days. However, given the consistency of our findings across both cohorts and the fact that this represents one of the largest validation studies of FBG against CGM in LTCFs to date, this limitation may not be of practical concern.

5 |. Conclusion

Clinicians and researchers should exercise caution when using EHR data alone to guide clinical decisions regarding hypoglycemia management in LTCF residents. Given the potential for substantial under-detection, with approximately four out of every five hypoglycemic events missed by FBG tests, reliance on EHR-based FBG measures of hypoglycemia may lead to missed opportunities for de-intensifying glucose-lowering therapy, potentially exposing residents to unnecessary harms. Alternative approaches to hypoglycemia surveillance may be needed to ensure appropriate medication adjustments and patient safety.

Supplementary Material

Supporting Information

Additional supporting information can be found online in the Supporting Information section. Figure S1: Difference in time (minutes) between CGM-FBG measures at the day and interval levels. Figure S2: Difference in time (minutes) between CGM-FBG measures within each 8-h interval. Table S1: Characteristics of fingerstick blood glucose measures of hypoglycemia versus continuous glucose monitoring measures in Cohort 1, every 8 h measures (N = 35). Table S2: Characteristics of fingerstick blood glucose measures of hypoglycemia versus continuous glucose monitoring measures in Cohort 2, every 8 h measures (N = 792).

Summary.

• Key points

  • In two cohorts of long-term care facility residents with diabetes, routinely collected fingerstick blood glucose measures available in electronic health records detected hypoglycemia with 13%–20% sensitivity and 99%–100% specificity compared to reference-standard continuous glucose monitor measures.

  • The correlation between fingerstick blood glucose-measured hypoglycemia and continuous glucose monitor-measured hypoglycemia was weak (phi correlation in Cohort 1 = 0.26 [95% CI 0.06–0.44] and Cohort 2 = 0.27 [95% CI 0.10–0.44]) when analyzing day-level data, but slightly higher when using more granular 8-h interval data (Cohort 1: 0.39 [95% CI 0.11–0.61] and Cohort 2 = 0.30 [95% CI 0.1–0.46]).

• Why does this paper matter?

  • Our findings inform both clinical and research practice for long-term care facility (i.e., nursing home) residents with diabetes by validating electronic health record-based fingerstick blood glucose measures of hypoglycemia against reference-standard continuous glucose monitoring-identified hypoglycemia, ultimately demonstrating that fingerstick blood glucose has comparatively poor sensitivity, with approximately 4 out of every 5 hypoglycemic events missed. Healthcare professionals should remain cognizant of the limited sensitivity of fingerstick blood glucose monitoring and account for this in their clinical evaluations and interventions, whereas researchers should remain aware of the substantial under-ascertainment of hypoglycemia when using long-term care facility electronic health record data and either employ approaches to account for this potential bias or interpret the results of their studies judiciously.