This scoping review has been conducted according to the Manual for Evidence Synthesis (Chapter 10 - Scoping reviews) from the Joanna Briggs Institute [17] and reported according to the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) guidelines (Checklist 1) [18]. A detailed study protocol for this scoping review has previously been published, along with a description of any deviations from the original protocol [19].

In this review, the term “CGM-derived metrics” covers all metrics derived from CGM device data. Blood glucose metrics like HbA_1c or measurements from anything other than CGM data will not be included (eg, metrics based on self-monitored blood glucose measurements taken with finger-prick or measurements from blood samples such as HbA_1c). CVD outcomes were grouped as either clinical or subclinical. The following outcomes were considered clinical CVD: cardiovascular mortality, major adverse cardiovascular events, coronary artery disease, heart failure, stroke, and peripheral artery disease. Synonymous terms (eg, ischemic heart disease) and clinical events (eg, undergoing coronary artery bypass surgery) were also included.

Subclinical outcome measures were grouped into 6 subcategories: arterial stiffness, flow resistance, arterial wall thickness, arterial wall composition, cardiac and pulse-related measures, and arterial lumen. CVD outcomes did not include broader risk factors nonspecific to cardiovascular risk (eg, age and sex).

The eligibility criteria are reported in Textbox 1, with further details provided in the review protocol [19].

The MEDLINE and Embase databases were searched from inception to March 11, 2025, using a search strategy tested against 13 key articles within the field [12,14,20-30] by an information specialist (ON) and reviewed by another (THA; Multimedia Appendix 1) [31].

Following the search, all identified citations were collated and uploaded into EPPI Reviewer 6, and duplicates were removed [32]. A meeting was held after about 5% of all titles and abstracts had been screened to create consistency among the reviewers (HBT, BL-J, AH, and AAI). In the screening phase, 2 independent reviewers screened the titles and abstracts to assess eligibility. When all titles and abstracts had been screened, full-text versions of relevant articles were retrieved and assessed in detail against the eligibility criteria by 2 independent reviewers. The reasons for exclusion during full-text screening were recorded and reported. Any disagreements that arose between the reviewers at any stage of the selection process were resolved through discussion. Disagreements unresolved through discussion were settled by the senior researcher (AH).

After the first screening phase, the software tool citationchaser was used for backward and forward citation searching [31]. The tool was applied to all included studies, and the screening process was repeated until no additional studies were found through backward citation and forward citation searching [19].

Research questions were predefined and published in the scoping review protocol (Textbox 2) [19], and a corresponding data extraction table was developed based on the PRISMA-ScR checklist [18]. Data were extracted by the first author (HBT) and verified by the last author (AAI).

Study characteristics were aggregated using descriptive statistics, and a narrative summary accompanied the tabulated results.

We considered the most adjusted models to be more clinically relevant and therefore extracted results only from these in the main population of each study, if studies reported numerous estimates due to multiple adjustment levels, subgroup stratifications, and varied combinations of CGM metrics and CVD outcomes. Results are reported for all studies, both clinical and subclinical, but only the specific details of association studies investigating clinical outcomes have been prioritized and presented in the main text, and evidence from the subclinical studies has been presented in Multimedia Appendix 2. A full list of all the CGM metrics and CVD outcomes found in the literature is presented in Multimedia Appendix 3.

The search identified 5253 records, of which 369 were duplicates and therefore removed (Figure 1). After title and abstract screening, 4802 records were excluded, leaving 82 records for full-text screening.

For full-text screening, 1 study could not be retrieved and 30 studies were excluded for the following reasons: not an original article (n=3), CGM-derived metrics were not based on CGM data (n=5), the patient group did not have CVD as an outcome (n=13), the study was in the postsurgery stage (n=6), the study language was not understood by the authors of this review (n=2), and the study was not yet published in a journal (n=1). Two additional studies were identified through backward and forward citation searching, resulting in a total of 53 included studies on clinical and subclinical outcomes (Multimedia Appendices 4 and 5).

The most common patient group was people with type 2 diabetes (34 out of 53 studies), followed by people with type 1 diabetes (20 out of 53 studies) and those with prediabetes (4 out of 53 studies). The patient group mostly included adults, with 7 studies focusing on children younger than 18 years [33-39]. Almost all studies included both male and female participants, with the exception of 2 studies [40,41], which included only male participants. The geographic distribution of the studies was uneven, as most studies were from Asia (27 out of 53 studies) and Europe (22 out of 53 studies). Three studies included data collected in North America, and 1 study was from Australia. Only 1 study included data from an African country; however, the study population was still predominantly White [42]. No studies were from South America (Multimedia Appendix 6).

One study was a randomized controlled trial [34], and the remaining 52 were observational studies. The majority of studies (49 out of 53 studies) performed cross-sectional analyses, and only 5 studies conducted longitudinal analyses. The size of the study population varied greatly, ranging from 17 to 6225, with a median of 152 (IQR 75‐469). Eleven studies included routinely collected CGM data from the participants’ own devices, and 42 studies used CGM data that had been actively collected with a device provided as part of the study. Studies analyzed data with Spearman or Pearson correlation (n=11), Cox proportional hazards regression (n=3), linear regression (n=23), or logistic regression (n=22; Multimedia Appendices 7 and 8). All studies in this review were association studies, with only 3 studies reporting prediction measures from secondary analyses [28,43,44], and none of these 3 studies used machine learning methods. None of the studies shared data or code.

Medtronic devices were most frequently used (28 out of 53 studies), followed by Abbott (9 out of 53 studies) and Dexcom (5 out of 53 studies). Other companies were Menarini, Meiqi Company, and SIBIONICS (Multimedia Appendix 7). Sampling frequencies varied from 3 to 15 minutes but were only reported in a minority of studies (21 out of 53 studies).

Among the 53 studies, the most common CGM-derived metrics were time in range (TIR; n=23), mean amplitude of glycemic excursions (MAGE; n=22), mean blood glucose (n=21), SD (n=19), coefficient of variation (CV; n=19), time below range (TBR; n=16), and time above range (TAR; n=15; Multimedia Appendix 9). Most studies did not find a statistically significant association between CGM-derived metrics and CVD. For example, only 7 out of 23 (30%) studies involving TIR found an association between TIR and CVD. However, among studies involving MAGE, a high proportion (14/22, 64%) reported an association between MAGE and CVD. The use of CGM-derived metrics differed among studies involving different diabetes populations. For example, 10 out of 23 (43%) studies involving TIR and 5 out of 22 (23%) studies involving MAGE had populations with type 1 diabetes. Among studies on people with type 1 diabetes, limited studies detected a statistically significant association between CGM-derived metrics and CVD (eg, TIR: 0/10, 0%; MAGE: 2/5, 40%). In contrast, studies on people with type 2 diabetes more often detected a statistically significant association (eg, TIR: 5/11, 45%; MAGE: 10/14, 71%). Studies on people with type 2 diabetes had a larger median sample size (TIR: 510, IQR 405-600; MAGE: 251, IQR 89-411) than studies on people with type 1 diabetes (TIR: 214, IQR 119-547; MAGE: 57, IQR 30-215; Multimedia Appendix 9).

Three studies included predictive analyses investigating MAGE as a predictor of CVD using logistic regression models. The reported area under the receiver operating characteristic curve (AUC) was 0.61 in one study [43] and 0.62 in another study [28]. Both studies reported MAGE to be a superior predictor when compared to HbA_1c, which had AUC values of 0.55 and 0.58, respectively. In the third study, a receiver operator characteristic curve analysis was conducted to ascertain the optimal threshold for dichotomizing MAGE as part of the variable selection process, but the AUC was not reported [44]. The authors found that MAGE ≥3.4 mmol/L was a risk factor for stenosis and/or occlusion, with a sensitivity of 0.60 and a specificity of 0.61.

Of the 53 included studies, 13 (25%) focused solely on clinical outcomes, 37 (70%) focused solely on subclinical outcomes, and 3 (6%) investigated both outcomes. The design/demographics and main findings of the studies on clinical cardiovascular outcomes are summarized in Tables 1 and 2, respectively, and the results of the studies on subclinical cardiovascular outcomes are summarized in Multimedia Appendix 7.

Two longitudinal studies found an association between a CGM-derived metric (TIR and hypoglycemia) and cardiovascular mortality [21,46].

Two studies assessed major adverse cardiovascular events [45,46]. Both studies included nonfatal myocardial infarction, nonfatal stroke, and cardiovascular death. One study also included unstable angina leading to hospitalization [46]. Hypoglycemic events and low blood glucose index values were associated with major adverse cardiovascular events; however, no associations were found for other CGM-derived metrics, including glucose management indicator, high blood glucose index, average of daily risk range, m-value, and blood glucose risk index.

Five studies explored the association between CGM-derived metrics and nonfatal cardiovascular events regardless of anatomical location as a composite CVD outcome, with some variation between the studies in terms of the complications included [27,47-50]. All 5 studies included cerebrovascular accident, 4 included peripheral artery disease [47-50], 3 included coronary artery disease [48-50], 2 included ischemic heart disease [27,47], and 1 each included stenosis [48], heart failure [50], and ankle-brachial index <0.9 or abnormal carotid intima-media thickness [27]. One study [27] found an association between TBR and cardiovascular complications, while another study [47] did not find an association. A study by Deng et al [49] found an association between CV and diabetic macroangiopathy in people who were Hp1 carriers but not in people with the Hp2‐2 genotype. Furthermore, no studies found evidence for associations between macrovascular complications and the following CGM-derived metrics: CV, SD, TIR, TAR, glucose management indicator, time in tight range (TITR), lowest blood glucose value, and area under the TBR curve [27,47,48,50].

Six studies investigated coronary artery disease [28,40,43,46,48,53]. Studies reported an association between MAGE ≥3.4 mmol/L and coronary artery disease [28,43]. Furthermore, a difference was observed between the control group and the high blood glucose fluctuation group based on dichotomizing the following CGM metrics: SD <1.40 mmol/L, MAGE <3.90 mmol/L, largest amplitude of glycemic excursions <4.40 mmol/L, and mean of daily differences <0.83 mmol/L [28,43,53]. TIR <20% was found to have an association, but none of the other TIR intervals [53] or TITR [48] had an association. Lastly, hypoglycemic events were found to have an association with myocardial infarction but not with unstable angina pectoris [46]. Four studies investigated the severity of coronary artery disease [55] by using the Gensini score [28,40,43,54]. Three studies found an association with MAGE [28,43,54]. Further associations were found when comparing the control group with the high blood glucose fluctuation group [28,40]. Two studies investigated the complexity of coronary artery disease [56] using the SYNTAX score [30,54]. MAGE was found to have an association with the SYNTAX score [30,54] together with blood glucose excursions during the night from 00:00 to 03:00, in the mornings from 06:00 to 08:00, and at midday from 11:00 to 13:00. No association was detected for all the other times during the day [30].

Three studies investigated stroke or cerebrovascular accidents. TIR, but not hypoglycemic events, was associated with stroke [46,51]. TIR and TITR were both associated with cerebrovascular accidents [48].

A study by De Meulemeester et al [48] included peripheral artery disease, while a study by Li et al [52] included lower extremity artery disease. Overall, they investigated CV, SD, TIR, and TITR, and only an association with TIR was found in some analyses [52].

This scoping review identified 53 studies focusing on the relationship between CGM-derived metrics and CVD risk in individuals with diabetes. The literature included inconsistent findings across association studies, which also had highly diverse clinical and subclinical CVD outcomes. CGM-derived metrics are widely studied, but their predictive value for CVD outcomes remains unclear since MAGE was the only metric whose predictive value was tested.

We observed patterns regarding study population size, diabetes type, and reporting of evidence for associations between CGM-derived metrics and CVD outcomes, as studies focusing on type 1 diabetes were often conducted in smaller study populations and rarely found evidence for associations. In contrast, studies focusing on type 2 diabetes were conducted in larger study populations and more consistently found evidence for associations. These patterns may explain some of the inconsistent findings for each CGM-derived metric and could indicate a lack of statistical power in some studies, suggesting that future studies, particularly those focusing on type 1 diabetes, should emphasize having sufficiently sized study populations. However, differences between type 1 and type 2 diabetes populations extend beyond sample size and include distinct pathophysiology, treatment regimens, and cumulative exposure to cardiovascular risk factors, all of which may influence the relationship between CGM-derived metrics and cardiovascular outcomes.

The most frequently investigated CGM-derived metric in our review was TIR, and we found inconsistent results across the included studies. Similar results were reported in the review by Yapanis et al [14]. The authors argued that low TIR is a risk factor for macrovascular disease and mentioned that the large sample size of a supporting study [21] provides more reliable evidence than the inconsistent results reported from smaller sample sizes. The same study [21] was the largest in our review, and its size makes its conclusion compelling. The inconsistency across other TIR studies is likely due to limited power and cross-sectional designs. MAGE was the CGM-derived metric most consistently associated with cardiovascular outcomes across studies and the only CGM-derived metric used for prediction; however, it exhibited poor discriminative ability. The more consistent associations observed for MAGE across studies, despite generally smaller sample sizes compared to TIR, may suggest a stronger link between MAGE and CVD risk than between TIR and CVD risk. Even though MAGE was studied across a diverse range of subclinical outcomes (Multimedia Appendix 7), studies on clinical outcomes were limited to coronary artery disease and severity scores, and studies on other clinical outcomes are needed to confirm this pattern, which would suggest that within-day glycemic variability may be an important cardiovascular risk factor. However, since MAGE is biased toward detecting hyperglycemic excursions [57], it may underestimate the impact of hypoglycemia on CVD. Although some studies have reported associations between hyperglycemia-focused CGM-derived metrics and cardiovascular outcomes [26,43], other studies have reported associations between hypoglycemia and cardiovascular outcomes [27,39,46,58-60], with some concluding that hypoglycemia is associated with macrovascular complications and hyperglycemia is associated with microvascular complications [27,59].

The heterogeneity in the reported associations between CGM-derived metrics and CVD outcomes may also reflect differences in how the metrics were defined and analyzed. Thresholds for TIR, TAR, and TBR, as well as observation periods, varied across studies. Statistical adjustment strategies also differed. Some studies adjusted for HbA_1c or other covariates, while others made no adjustments. However, we found no substantial differences between the unadjusted and the most adjusted estimates in the studies, suggesting that attenuation due to adjustment for other covariates played only a minor role. Study comparisons would have been easier if the analysis code were available; however, none of the authors provided this information. Together, these discrepancies highlight how diverse definitions and analytic approaches can contribute to conflicting findings and complicate the interpretation of the evidence in this field. The heterogeneity in this field makes it impossible to present findings in a quantitative meta-analysis, and there is a need for more standardized study designs if studies cannot generate definitive evidence by themselves.

Several recurring methodological issues also emerged. First, most studies were geographically concentrated in Asia and Europe, limiting generalizability to other health care settings and populations, particularly those in South America and Africa. Second, many studies assessed multiple combinations of CGM metrics and CVD outcomes in separate models without prespecified hypotheses or correction for multiple testing. Third, and most importantly, only 3 studies performed prediction modeling analyses, all of which were carried out as secondary analyses. None of the studies reported any external validation or performance metrics beyond discrimination (AUC, sensitivity, and specificity), indicating modest performance [61]. No study applied machine learning methods or used raw CGM time-series data, which may further constrain predictive ability. Thus, the predictive utility of CGM-derived metrics for CVD outcomes remains essentially untested. There is a clear need for sufficiently powered, longitudinal prediction studies using clinical CVD outcomes in ethnically diverse populations [62].

The vast majority of identified studies were cross-sectional, limiting their clinical relevance due to potential reverse causality, as established CVD can alter lifestyle behaviors and glucose patterns. This bias can skew the results in 2 directions. First, it may produce false-positive associations if distinct CGM patterns only emerge after a CVD event. Second, it can yield false-negative findings if incident CVD, or the resulting intensive medical treatment, masks or attenuates a pre-existing glucose pattern. However, there were too few longitudinal studies to assess if reverse causality systematically skewed the estimates provided by cross-sectional studies and thereby led to divergent results between the 2 study designs.

Collectively, the methodological challenges identified in this review indicate a need for clearer methodological alignment in future studies if systematic reviews are to be feasible. Specifically, researchers should adhere to consensus guidelines, such as the ATTD (Advanced Technologies and Treatments for Diabetes) consensus recommendations [63], together with prespecified covariate adjustment strategies, standardized classification of cardiovascular outcomes, and transparent reporting of analytic decisions. Addressing these areas would improve comparability across studies and strengthen the interpretability of future evidence.

Most studies used CGM data collected specifically for research, with only a few studies drawing on routinely collected real-world data despite the growing prevalence of CGM use. This represents a missed opportunity, as routine data are typically larger, more cost-effective, and more representative of CGM users. Underuse may reflect challenges in accessing data stored on proprietary manufacturer platforms or linking these data to individual health records. Open, publicly available datasets have driven advances in many fields (eg, medical image analysis) [64-66], but no comparable dataset exists for studying CGM data and complications. In the absence of such resources, aligning existing databases with FAIR (findability, accessibility, interoperability, and reusability) principles could help accelerate research in this area [67].

A key strength of this review was the differentiation between association and prediction studies, highlighting the lack of knowledge on how well CGM-derived metrics perform in CVD prediction models. We performed a more comprehensive literature search, yielding an additional 40 studies compared to a previous review [14]. The distinction between clinical and subclinical CVD allowed a more detailed synthesis of how CGM-derived metrics relate to both CVD manifestation and early vascular changes. Furthermore, we provided a detailed methodological overview and revealed common methodological weaknesses, including variations in the calculation of CGM-derived metrics and the definitions of cardiovascular outcomes.

This review also has limitations. First, the feasibility of synthesizing effect sizes consistently across studies was limited by heterogeneity in study designs, CGM metrics, and CVD outcome definitions. Therefore, this review summarized studies based on P values, which is suboptimal, as P values are influenced by both the effect size and the sample size [68]. P values do not accurately reflect the effect size, clinical relevance, or estimate precision. This greatly limits our ability to compare the strength of associations across studies. Furthermore, underpowered studies are more likely not to find evidence for associations, thereby adding noise to the literature. Second, we reported only the most adjusted models from each study. While this approach was deemed necessary, it may have excluded potentially informative results from alternative model specifications. Third, identifying all relevant studies in this field proved challenging. We decided to limit the search to MEDLINE and Embase only, as these are core databases for biomedical literature searching. Given the resources available to the review team, we were not able to extend the database search further. However, the search retrieved a high number of records, both relevant and irrelevant, owing to inconsistent terminology and overlapping search categories (eg, “blood glucose monitoring” and “glycemic control” both encompass finger-prick measurements). We therefore designed a broad search strategy to ensure that we did not miss any relevant studies in the 2 databases that we chose to search. Acknowledging that searching only 2 databases may have resulted in missing relevant studies, we systematically screened all references and citing articles (backward and forward citation searching) of the included studies. This process resulted in the identification of 2 additional articles [44,69], of which 1 article (Koroleva et al [44]) was not indexed in the medical databases we searched. Nevertheless, to the best of our knowledge, this comprehensive search strategy enabled us to identify more relevant studies than any previous review on this topic.

This scoping review mapped a broad landscape of association studies examining associations between CGM-derived metrics and CVD outcomes, with a smaller number also addressing prediction. The included studies were methodologically heterogeneous, making it difficult to synthesize evidence and draw firm conclusions about clinical cardiovascular risk.

Within these constraints and using statistical significance as a pragmatic indicator of consistency across heterogeneous studies covering different CVD outcomes, TIR was associated with CVD in the largest single study, and MAGE was the CGM-derived metric most consistently associated with CVD outcomes across multiple studies covering subclinical outcomes, coronary artery disease, and severity scores. Notably, MAGE was the only CGM-derived metric to have its predictive value assessed, and it exhibited only modest discriminatory performance. None of the studies used any machine learning–based methods, suggesting that the predictive value of CGM-derived metrics for CVD outcomes and the possibilities of using machine learning–based methods are underexplored. There is a fragmented evidence base in which metric definitions, study designs, and analytical strategies vary widely. In the future, more standardized analytical strategies could enable meta-analyses across individual studies to synthesize more substantial evidence.