Dynamic Relationships Among Continuous Glucose Metrics and Momentary Cognitive Performance in Diverse Adults With Type 1 Diabetes Free

Extensive research links diabetes with impaired performance across a variety of cognitive domains. Adults with type 1 diabetes (T1D) and type 2 diabetes (T2D) tend to have poorer global cognitive performance compared with those without diabetes, with processing speed, attention, and executive functioning particularly affected (1–3). However, studies typically use HbA_1c to measure glucose levels and traditional, in-person test batteries to measure cognition. HbA_1c cannot capture short-term glycemic variation or distinguish between exposure to hyperglycemia and hypoglycemia, which may be differentially associated with cognitive impairments (4,5). Further, office-based cognitive assessments are not designed for frequent readministration and are limited in their ability to detect variability in ecologically valid environments. These limitations make it unclear whether dysregulated glycemia precedes impaired cognitive performance, or vice versa, or whether both directions of influence are observable over the course of daily life with T1D. Better understanding these effects could inform interventions to protect cognitive function among adults with T1D.

Ecological momentary assessment (EMA) methods enable densely repeated cognitive testing to capture variability in performance over time in individuals’ everyday environments; performance can then be linked with other time-varying factors such as continuously monitored glucose levels. These temporally paired data can provide greater ecological validity, precision, and reliability than laboratory-based assessments (5–7). A few recent studies have examined nighttime glycemia measured using continuous glucose monitoring (CGM) as a predictor of daytime ambulatory cognitive performance (8,9), and one study evaluated CGM values within the prior 5 min in relation to ambulatory cognitive performance over the course of a day (10). In the Function and Emotion in Everyday Life with Type 1 Diabetes (FEEL-T1D) study, we found that greater overnight glycemic variability predicted worse next-day sustained attention, and more overnight time in hypoglycemia predicted worse next-day perceptual speed (8). Perceptual speed is a facet of processing speed (11), a cognitive domain that is consistently associated with diabetes (1,2,12,13). Similarly, preliminary results from 18 participants in the Glycemic Variability and Fluctuations in Cognitive Status in Adults with Type 1 Diabetes (GluCog) study indicate more time in nocturnal hypoglycemia predicted slower next-day processing speed within persons, whereas daytime hypoglycemia was not associated with same-day processing speed; neither nocturnal nor daytime hyperglycemia were associated with processing speed. No within-person associations were found between glycemia and sustained attention (9). Full sample results (n = 200) showed that both low and very high glucose levels assessed within minutes before cognitive assessments predicted slower processing speed within persons, with optimal processing speed observed at slightly higher glucose than is typical for each person. No within-person associations were observed between glucose and sustained attention performance. Lagged associations between cognitive performance and subsequent glycemic levels were not reported (10). The current study builds on these findings by using CGM to characterize glycemic regulation over periods throughout the day and evaluate lagged bidirectional relationships with ambulatory cognitive performance among diverse adults with T1D.

Our aims were to examine whether glucose metrics predicted subsequent ambulatory perceptual speed and sustained attention over 3-h time frames, as well as whether cognition predicted glycemia over the following 3 h. Further, we explored participant characteristics that may moderate relationships between glucose and cognitive performance. Figure 1 illustrates the different time frames and relationships examined. We hypothesized that higher time in range (TIR), lower time above and below range, and lower glycemic variability over 3 h would be associated with better cognitive performance at the end of that time frame, as well as 3 h later. Similarly, we hypothesized that better perceptual speed and sustained attention would predict subsequently better glycemic quality over the next 3 h given the cognitive demands of diabetes self-management (14–16). We hypothesized that older age, and indicators of more severe glycemic dysregulation (experiencing severe hypoglycemic episodes, complications, and fatigue), would strengthen associations between glycemic parameters and cognitive performance, based on prior evidence (8,10). We also examined gender as a moderator, given increased risk for cognitive decline among females (17). As CGM use can facilitate or decrease the need for certain health behaviors (i.e., self-monitoring of blood glucose), we hypothesized that CGM use would weaken associations, particularly those between cognitive performance and subsequent glycemia.

Participants used blinded CGM for 14 days and completed ambulatory cognitive tests five or six times daily at 3-h intervals. Participants also used wrist-worn accelerometers to measure physical activity. Study procedures were offered in English and Spanish. Participants were recruited from endocrinology clinics in Los Angeles and New York City based on the following inclusion criteria: 1) having a diagnosis of T1D for at least 1 year, 2) being age 18 years or older, 3) having adequate vision/fine motor skills, 4) being on stable T1D therapy for 3 months or more, and 5) having no planned medical procedures that could impact CGM readings. Participants were asked to begin their morning assessment soon after waking, and remaining surveys were scheduled 3 h apart from that initial start time. Details of the study protocol have been published (18). The Institutional Review Boards of the University of Southern California (Los Angeles, CA) and the Albert Einstein College of Medicine (Bronx, NY) approved the study protocol; all participants provided informed consent.

Demographic variables collected at baseline included age, gender, race and ethnicity, marital status, and socioeconomic variables (employment status, annual household income, level of education, and type of health insurance). Clinical information collected include duration of T1D, the presence of diabetes complications (diabetic eye, nerve, or kidney damage), and whether participants used CGM and/or an automated insulin delivery (AID) system as part of their usual care. Participants also reported whether they had experienced “severe low blood glucose that resulted in passing out, losing consciousness, or seizure” in the past 6 months.

Glycemia was measured continuously using the Abbott FreeStyle Libre Pro Flash Glucose Monitoring System (Alameda, CA). The device records interstitial glucose every 15 min. The CGM data were reprocessed with an algorithm equivalent to the FreeStyle Libre 2 system by the Abbott Diabetes Clinical Research group. Metrics extracted from CGM data included mean glucose, glycemic variability (coefficient of variation; CV), %TIR (70–180 mg/dL), and %time in high (181–250 mg/dL), %time in very high (>250 mg/dL), and %time in low (<70 mg/dL) glucose ranges (19,20). We collapsed time in low and very low ranges into one category given the low frequency of very low glucose. We also included the most contemporaneous measure of glucose to the time of cognitive assessment (within 15 min). We included 3-h summary metrics if >70% of data were available for that time frame (i.e., 9 of 12 observations). We only included participants with >0 variance in each CGM metric, leading to different N sizes for each metric (Table 2).

Cognitive functioning was assessed with two cognitive tasks that showed high reliability in previous studies (5,6,21,22).

Perceptual speed, a component of processing speed (11), was assessed using a Symbol Search task (5). Participants were asked to decide, as quickly as possible, which of two figures at the bottom of the screen match one of three options presented at the top of the screen (for 20 trials). The median response time of accurate trials (reliability = 0.57 within persons and 0.98 between persons) was calculated to indicate perceptual speed. Responses that were completed too quickly (<200 ms) or slowly (>5,000 ms; 3.32% of observations) (5) or indicated inattentive responding (sessions with <70% accuracy; 1% of observations) were removed (22). Response times were multiplied by −1 so that higher scores indicate better performance.

Sustained attention was measured with the TestMyBrain gradual onset continuous performance test (GradCPT), a go/no-go task in which participants responded to city images (go) and withheld responding to mountain images (no go) for 75 trials (21,23,24). The task was scored using the d′ metric (reliability 0.34 within person and 0.97 between persons) based on the proportions of hits (correct omissions to mountains) and false alarms (incorrect omissions to cities). Scores for sessions with omission errors >50% were excluded (3.1% of observations). Higher scores indicate better sustained attention.

Fatigue was assessed at every EMA prompt with, “at this moment, how tired do you feel?” Responses ranged from 1 to 100 (“not at all” to “extremely”) and were averaged across the full study period.

Daily step counts were calculated in 60-s epochs using the ActiGraph wGT3X-BT accelerometer (Pensacola, FL) and software. Sedentary behaviors were defined as vector magnitude counts <2,860 per minute, following recommended cut points (25).

We evaluated within-person concurrent and lagged bidirectional relationships between glucose and cognition using dynamic structural equation modeling (DSEM). DSEM combines time series analysis and multilevel modeling, accounting for the nonindependence in observations (e.g., multiple data points from the same individual) and allowing for the consideration of lagged effects (26–28). DSEM cross-lagged panel models can be used to show how two variables influence each other over time (29). We examined cross-lagged relationships between variables over sequential 3-h periods, as illustrated in Fig. 1.

Figure 1 also illustrates the statistical model structure and how this aligns with the data collected. A detailed schematic of the estimated DSEMs has been previously published (8). For each pairing of the two cognitive tests and seven glucose metrics (total of 14 pairings), a multilevel cross-lagged panel model estimated three within-person relationships of interest simultaneously. It tested whether 1) glycemia collected over 3 h predicted cognitive performance at the end of those 3 h (r), 2) glycemia collected over 3 h predicted cognitive performance 3 h later (β), and 3) cognitive performance predicted glycemic quality over the following 3 h (β). Between-person relationships were also estimated. All parameters were specified as correlated random effects, allowing for individual variation in their magnitudes and directions.

Given the unexpected pattern of association between very high glucose and perceptual speed over different time frames, we further explored post hoc mediation and moderation by variables that could help explain these findings. As experiencing very high glucose would be expected to initiate insulin administration, we explored whether AID use moderated the relationship between very high glucose and reduced perceptual speed 3 h later. We also examined whether glycemia over the 3-h time frame between cognitive assessments (e.g., from 9 a.m. to 12 p.m.) mediated the relationship between earlier glycemia (e.g., from 6 a.m. to 9 a.m.) and lagged cognitive performance (e.g., at 12 p.m.) (Fig. 1). Finally, we explored mediation by physical activity, which has been shown to be reduced following high glucose (8).

Further, we examined whether the three relationships tested by each model were moderated by gender, age, use of CGM, recent severe hypoglycemic episodes, presence of diabetes complications, and fatigue. We regressed the random effects of cross-lagged parameters and the residual correlation on a moderator (one at a time) at the between-person level. A significant Wald χ² test statistic was considered indicative of moderation.

The FEEL-T1D study was designed to have 80% power to detect an effect size of 0.10 for within-person dynamics between glucose and function (18). Missing data for included participants were handled using a Bayesian estimator. All analyses were completed using Mplus version 8.8 (30) via the R package MplusAutomation (31) in the statistical software R (32).

A total of 208 adults with T1D participated in the FEEL-T1D study, of which 182 participants had both CGM and ambulatory cognitive data available. Twelve participants did not complete the EMA portion of the study, and an additional 14 participants did not have CGM data (e.g., faulty or lost CGMs). Descriptives for the 182 participants are presented in Table 1. No significant differences in these characteristics were found between the 14 excluded participants and the analytic sample (data not shown). On average, the participants were in middle age, slightly more female than male, and socioeconomically and ethnically/racially diverse. Approximately 59% were using unblinded CGM as part of their care. The median EMA completion rate was 92%. Average glucose levels for participants were above recommended targets, with approximately half of their time spent in the target range, with low prevalence of time in hypoglycemia.

Table 2 provides a summary of bidirectional relationships. Perceptual speed had more pronounced within-person relationships with glycemia (Table 2, rows 1–3), while sustained attention had more pronounced between-person relationships (Table 2, row 8). Considering within-person relationships, glycemia and perceptual speed showed bidirectional relationships (Table 2, rows 1–3), while sustained attention and glycemia showed only a unidirectional association (Table 2, row 7).

Analyses modeling glycemia over 3 h and subsequent cognitive performance indicated that glycemia did not significantly predict attention scores at the end of the 3-h time frame (Table 2, row 5) or 3 h later (Table 2, row 6). The direction of the relationship between glycemia and subsequent perceptual speed depended on the time frame examined. Higher mean glucose (r = 0.12 [95% CI: 0.07, 0.16]), more time in very high glycemia (r = 0.15 [95% CI: 0.08, 0.21]), less time in low glycemia (r = −0.18 [95% CI: −0.24, −0.11]), and higher CV (r = 0.05 [95% CI: 0.02, 0.08]) predicted faster perceptual speed at the end of the 3-h time frame. The positive effect of glucose and perceptual speed was also observed when limited to immediate glucose collected within 15 min before the cognitive assessment (r = 0.11 [95% CI: 0.075, 0.14]) (Table 2, row 1). When examining the effects of glycemia on perceptual speed assessed 3 h later, higher mean glucose (β = −0.24 [95% CI: −0.38, −0.09]), lower TIR (β = 0.31 [95% CI: 0.06, 0.54]), and more time in very high glucose (β = −0.78 [95% CI: −1.26, −0.35]) predicted slower perceptual speed (Table 2, row 2). Low glucose, CV, and immediate glucose did not predict perceptual speed 3 h later.

While several glucose metrics only significantly predicted perceptual speed at one rather than both time points, this pattern of opposing relationships between glycemia and perceptual speed across different time frames was observed for every single metric including TIR, time spent in low glucose, CV, and immediate glucose (Table 2, rows 1 and 2).

When considering cognitive performance and subsequent glycemia, faster perceptual speed was associated with higher mean glucose (β = 0.01 [95% CI: 0.006, 0.020]), less TIR (β = −0.004 [95%CI: −0.007, −0.001]), and more time in very high glucose (β = 0.005 [95% CI: 0.002, 0.008]) over the next 3 h (Table 2, row 3). Better sustained attention predicted a lower mean glucose (β = −0.01 [95% CI: −0.01, −0.0001]), less time in very high glucose (β = −0.002 [95% CI: −0.004, −0.0004]), but higher CV (β = 0.001 [95% CI: 0.0, 0.001]) over the next 3 h; the effect of increased sustained attention on decreased glucose was also observed for glucose measured immediately after the cognitive assessment (β = −0.007 [95% CI: −0.012, −0.002]) (Table 2, row 7).

Exploratory analyses indicated that AID use did not moderate the relationship between time in very high glucose and reduced perceptual speed 3 h later (Wald χ² = 3.49, P = 0.06). Further, while people tended to have less time in very high glucose in the 3-h block after spending time in very high glucose, this did not mediate the relationship between time in very high glucose and lowered perceptual speed 3 h later (nonsignificant indirect effect of β = 0.06 [95% CI: −0.07, 0.18]). Rather, time spent in nonsedentary activity in the 3 h between cognitive tests mediated about 10% of the effect of very high glucose on perceptual speed 3 h later (indirect effect of β = −0.08 [95% CI: −0.11, −0.05]). Specifically, more time in very high glucose predicted less activity in the following 3 h (β = −0.002 [95% CI: −0.003, −0.001]), which then predicted slower perceptual speed (β = 38.2 [95% CI: 30.9, 45.8]).

Use of CGM as part of usual care moderated 3 out of 42 potential associations between cognition and glycemia (3 associations per model for seven glucose metrics and two cognitive tests). Among people not using a personal CGM, slower sustained attention was associated with higher average glucose (β = −0.01 [95% CI: −0.02, −0.01; Wald χ² = 6.42, P = 0.01) and more time in very high glucose (β = −0.006 [95% CI: −0.01, −0.002; Wald χ² = 4.82, P = 0.03) over the next 3 h, with no association for participants using a personal CGM. At the same time, among those using a personal CGM, faster perceptual speed was associated with more time in high glucose over the next 3 h (β = 0.01 [95% CI: 0.003, 0.02]), with no association for people without a CGM (Wald χ² = 8.01, P < 0.001).

Gender and history of severe hypoglycemia moderated 2 out of 42 potential associations. For females, higher mean glucose predicted slower perceptual speed 3 h later (β = −0.343 [95% CI: −0.536, −0.186]), with no association for males (β = −0.04 [95% CI: −0.25, 0.19]) (Wald χ² = 5.34, P = 0.02). Gender also moderated the association between perceptual speed and time in high glucose over the next 3 h, with faster perceptual speed predicting more time in high glucose over the next 3 h for females (β = 0.005 [95% CI: 0.001, 0.01]), but not for males (Wald χ² = 5.07, P = 0.02). For individuals without a history of severe hypoglycemia, higher mean (r = 0.13 [95% CI: 0.09, 0.17]) and immediate (r = 0.13 [95% CI: 0.09, 0.17]) glucose were associated with faster perceptual speed at the end of the 3-h time frame, while there were no associations for people reporting a severe hypoglycemic event (r = 0.00 [95% CI: −0.10, 0.11], Wald χ² = 5.30, P = 0.02; r = 0.01 [95% CI: −0.06, 0.09], Wald χ² = 5.30, P = 0.02, respectively).

Presence of diabetes complications and fatigue moderated 1 out of 42 associations. Presence of complications moderated only the relationship between perceptual speed and immediate glucose 3 h later (Wald χ² = 4.49, P = 0.03). For people reporting a complication, the relationship with immediate glucose was trending negative (β = −0.009 [95% CI: −0.021, 0.004), while it trended positive for those reporting no complications (β = 0.008 [95% CI: 0.000, 0.016). Fatigue moderated the relationship between perceptual speed and CV over the next 3 h (Wald χ² = 4.56, P = 0.03). For individuals with fatigue 1 SD below the mean, the relationship trended negative (β = −0.006 [95% CI: −0.014, 0.005]), while it trended positive for individuals with fatigue 1 SD above the mean (β = 0.008 [95% CI: −0.001, 0.017]). Age did not moderate associations among glycemia and cognitive performance.

This study adds to our understanding of how glycemia and cognition are dynamically related in the daily lives of diverse adults with T1D. Examining glucose regulation over 3-h periods using CGM metrics commonly used in clinical practice, we found a lagged association between more time in hypoglycemia and slower perceptual speed at the end of those 3 h, with no association found 3 h later. Higher mean glucose and more time in high (>250) over this period were each associated with faster perceptual speed initially, followed by slower perceptual speed 3 h later. While higher immediate glucose was associated with faster perceptual speed within a 15-min time frame, it did not predict perceptual speed 3 h later. Results complement findings from the GluCog study, which found within-person associations between dysglycemia and reduced processing speed 5 min later, but no associations between glucose levels and subsequent sustained attention (10). However, the current study found robust between-person links between reduced sustained attention and higher average and immediate glucose levels, less TIR, and more time with high and very high glucose levels, as well as evidence that better sustained attention predicted improved glycemic quality at the within person level.

Findings from mediation analyses support decreased physical activity as a partial explanatory pathway for short-term effects of hyperglycemia on subsequent lower perceptual speed. Unmeasured physiological mechanisms could also be implicated in mediating these effects and deserve further investigation, including effects from food or insulin, changes in cortisol, greater production of glutamate, and other compensatory processes that may have delayed effects on cognition (33,34). Finally, our exploration of potential moderators found minimal evidence, considering the number of effects tested, and should be interpreted with caution. We found that a history of severe hypoglycemia weakened the association between hyperglycemia and faster perceptual speed at the end of the 3-h time frame. Additionally, females experienced a stronger negative effect of elevated glucose on slower perceptual speed 3 h later. While it appears that females may be more vulnerable to the effects of elevated glucose, this is less clear for those with severe hypoglycemia, and we did not find other indicators of vulnerability and poor health to strengthen the effects of glucose on processing speed. In contrast, the GluCog study found evidence for increased vulnerability with older age and various indicators of more severely dysregulated glycemia and poor health (10). In our study, moderators more consistently influenced the relationships between cognitive performance and subsequent glycemic quality, including personal CGM use, gender, presence of complications, and fatigue. Of interest, those with complications showed a negative trend between faster perceptual speed and less time in high, which is contrary to the positive relationship observed overall. Future research is needed to identify patient characteristics that could make certain groups of patients more vulnerable to the effects reported here.

Our findings linking reduced sustained attention to subsequent deterioration in glycemic quality suggest that sustained attention may be an important cognitive process involved in diabetes self-management behaviors. Moderation analyses indicated that personal CGM use weakened this association, suggesting that CGM use could attenuate the negative effects of reduced attention on self-management. At the same time, faster perceptual speed predicted worse glycemia over the following 3 h, and CGM use strengthened this association. Given the bidirectional link between perceptual speed and glycemia, slower speed may act as an indicator of poor glycemic quality, and further work is needed to identify how this may relate to the enactment of self-management behaviors and personal CGM use. The finding that worse sustained attention predicted worse glycemic quality aligns with cross-sectional data (14–16), although the lack of between-person associations for faster processing speed and better glycemic quality, and the opposing relationship within persons, does not. If observed associations are causal, interventions that improve sustained attention in particular could have positive effects on glycemic management in diabetes. These novel findings need replication.

Results support prior findings indicating that experiencing hypoglycemia, hyperglycemia, and glycemic variability can have short-term effects on cognitive performance (10,35–37). However, the dynamics over time, including degree of exposure to dysglycemia and residual deficits over short time frames, are not well understood. The literature shows a variety of thresholds across different samples for hypoglycemia impacting cognitive performance, with considerable inconsistency even for the same test (38). Additionally, the recovery time for different cognitive tasks following hypoglycemia varies considerably, ranging between 10 and 90 min (39). Our results show that effects of hypoglycemia on perceptual speed are observed within 3 h, with a larger percentage of time in hypoglycemia associated with worse deficits. We also show residual effects of hyperglycemia on perceptual speed that can be observed up to 6 h later, with more time in hyperglycemia associated with initial improved performance followed by worse deficits. There was a generally consistent pattern with the immediate glucose metric as with our metrics averaged over 3 h, particularly mean glucose. However, lagged effects with perceptual speed were markedly weaker. This suggests that observed lagged associations with our 3-h glucose metrics are more reflective of exposure to, and effects of, glucose that accumulate over time and are less immediate, at least over the relatively brief 3-h intervals examined.

Limitations of our study design include that causal conclusions cannot be drawn from our findings. Another limitation is that the occurrence of very low glucose was infrequent. For this reason, we collapsed low and very low glucose into one category, limiting our ability to differentiate between different levels of hypoglycemia and cognitive function. Given the low occurrence, we lose some precision in understanding the temporal impact of hypoglycemia on perceptual speed, with more work needed in this area. Another limitation is that the within-person reliability of the sustained attention measure is lower than has been observed in other EMA studies (10,21). We also acknowledge the possibility of data being missing not at random, although missing data were limited. Additionally, the number of analyses required to evaluate these associations over time and across glucose metrics and cognitive domains was inherently large and increases the chances of spurious findings; further replication of our results is needed. Further, the current study did not include momentary questions assessing additional health behaviors that were fine-grained enough to include in current analyses, such as adherence to insulin and self-management generally (although constructs were measured once daily [18]). Examining these behaviors may yield a better understanding of the relationship between sustained attention and subsequent changes in glycemic quality. A strength includes that the study was conducted among a racially/ethnically and socioeconomically diverse sample, strengthening generalizability of findings. At the same time, participants were recruited from clinics where they were engaged in care, which could somewhat limit generalizability of findings.

This study builds on previous work examining the interplay of glycemia and ambulatory cognitive functioning in persons with T1D and provides an important step toward further needed work in this area. Current findings highlight the strengths of EMA methodology for better understanding real-world dynamics among glycemia and cognitive performance and emphasize the need to consider cumulative effects and different day-level time frames. While the clinical relevance of the modest but statistically significant associations observed is unclear, the consequences may be cumulative, which can magnify their impact over a lifetime. Further, findings suggest that several factors may influence the effects of glycemia on cognition, and vice versa, supporting individualization of treatment depending on specific needs.

Funding. This study was supported by grant R01 DK121298 from the National Institutes of Health and a grant from Abbott Laboratories (A.P. received donated devices from Abbott Diabetes Care). This study was also partially supported by the Einstein-Mount Sinai Diabetes Research Center (P30 DK020541) and the New York Regional Center for Diabetes Translation Research (P30 DK111022).

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. C.J.H. contributed to the discussion and wrote, reviewed, and edited the manuscript. R.H. and S.S. ran the statistical models, contributed to the discussion, and wrote, reviewed, and edited the manuscript. A.P., M.H., E.A.P., and J.S.G. contributed to the discussion and reviewed and edited the manuscript. All authors approved the final version of the manuscript. J.S.G. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Prior Presentation. This work was presented in abstract form at the 83rd Scientific Sessions of the American Diabetes Association, San Diego, CA, 23–26 June 2023.

Handling Editors. The journal editors responsible for overseeing the review of the manuscript were Steven E. Kahn and Stephanie L. Fitzpatrick.