Two independent researchers systematically searched the PubMed, Embase, Scopus, and Web of Science databases. Manual searches were also performed on the Research Square, BioRxiv, and MedRxiv preprint servers to reduce the potential impact of publication bias. Retrieval strategies incorporating Boolean operators were utilized for the following keywords: (1) “insulin resistance” OR “homeostasis model assessment” OR “HOMA” and (2) “coronary artery calcification” OR “coronary calcification” OR “coronary artery calcium” OR “coronary calcium” OR “subclinical coronary atherosclerosis.” A supplementary manual search was conducted through references to relevant literature. The final literature search was updated to 16/09/2023. A detailed search strategy is provided in Supplementary File 2 . Endnote software was used in the literature screen process.

The criteria for inclusion in the literature were as follows: (1) adult participants; (2) reported on the relationship between IR and CAC; (3) assessed IR using the homeostasis model assessment of IR (HOMA-IR) method; (4) evaluated the prevalence or progression of CAC based on the Agatston score; (5) reported effect estimates while accounting for underlying confounding variables. Conference abstracts, case reports, reviews, letters, comments, and expert opinions were excluded. CAC was defined using the definition provided in the original study. The effect estimates were evaluated using risk, hazard, or odds ratio (OR) values.

Literature retrieval, data extraction, and quality evaluation were performed independently by two researchers (YH and XZ). In case of any discrepancies, a consensus was reached by consulting a third researcher (YS) to resolve the dispute. If vital data for analysis were found to be missing, we would approach the respective paper’s corresponding author to acquire the original data. The following information was extracted: author information, year, country, research type, study population characteristics, sample size, age, gender, disease status, HOMA-IR analysis method, reported outcomes, inclusion and exclusion criteria of subjects, variables adjusted, and outcome definitions. To evaluate the quality of the literature incorporated in this research, we applied the Newcastle–Ottawa Scale (NOS) (21). This scale, which employs a scoring system ranging from 0 to 9, was utilized for assessing the selection of research populations, comparability among cohorts, and outcome measurement.

We utilized pooled ORs to assess the association between HOMA-IR and CAC. One study separately reported ORs for the relationship between HOMA-IR and CAC based on glycated hemoglobin quartiles (22), whereas another reported ORs for different CAC risk levels (23). Therefore, we combined these groupings, calculated ORs, and analyzed them using a random-effects model. We selected the most appropriately adjusted model for confounding factors in multifactor analyses with multiple models and extracted the corresponding estimates.

When HOMA-IR was analyzed as a categorical variable, we extracted the estimates of its highest quartile to the first quartile. When HOMA-IR was presented as a continuous variable, we extracted the value of each unit increase in the variable. To ensure a normal distribution, a log transformation was applied to both the ORs and their corresponding 95% CIs. Heterogeneity was evaluated using the Cochrane Q and I² tests (24). A P-value below 0.1 for the Q-test indicated significant heterogeneity. Heterogeneity was considered low if I² was less than 40%, whereas 40%–60% indicated moderate heterogeneity. In cases where heterogeneity was evident, sources of heterogeneity were sought based on study characteristics and sensitivity analyses. If heterogeneity persisted, a random effects model was employed to combine the information and adjust for different sample sizes, which helped mitigate heterogeneity’s impact (25).

In order to guarantee the dependability of the combined findings, a sensitivity analysis was performed, wherein each study was sequentially excluded, and the corresponding effect on the pooled OR estimates was evaluated (26). Further subgroup analyses were performed to investigate potential differences between integrated ORs with different characteristics, such as demographics and comorbid conditions. In addition, the symmetry of funnel plots was assessed to determine the potential for publication bias. We utilized RevMan (Version 5.1; Cochrane Collaboration, Oxford, UK) and STATA software to conduct this systematic review.

Two researchers (LL and HZ) utilized the GRADE (Grading of Recommendations, Assessment, Development and Evaluation) approach to assess the quality of evidence. Following the GRADE principle, the initial level of evidence obtained from observational research was deemed as having a low quality. Five factors were taken into account to diminish the credibility of the evidence, which encompassed limitations in the study design, inconsistency, indirectness, imprecision, and the presence of publication bias (27). Additionally, three factors enhance the quality of evidence: large effect sizes, potential confounders, and dose–response relationships (28). If factors increased the evidence quality during the evaluation, it was upgraded to moderate quality. Conversely, if factors diminished the quality of evidence, it was downgraded to very low (29).

Following the exclusion of duplicate articles in the initial literature search, a total of 1,279 articles were obtained. However, upon reviewing the titles and abstracts, 1,247 articles were deemed irrelevant and subsequently excluded from further consideration. After a thorough review of 32 relevant studies, it was found that five of them used a non-HOMA-IR method for IR evaluation, and 11 did not report study-related results. One study was excluded to avoid duplication of data sets based on a survey of the same population (30), and we used a more recent study of this group. Ultimately, 15 articles were included in the current research (14, 15, 17, 18, 22, 23, 31–39). The entire process of literature selection is shown in Figure 1 .

Table 1 presents a comprehensive overview of the selected studies comprising 60,649 subjects from 15 studies. The studies were carried out across multiple countries, comprising the United States (17, 18, 23, 32, 35, 36, 38), South Korea (14, 22, 31, 37, 39), Iran (33), Mexico (34), and Japan (15). Five were cohort studies (14, 15, 18, 31, 38), whereas the remaining 10 were cross-sectional (17, 22, 23, 32–37, 39). The population of the studies comprised community residents or citizens (15, 17, 18, 33–36, 38), participants undergoing health screening (14, 22, 31, 32, 37, 39), and chronic kidney disease patients (23). There were 12 studies that excluded individuals with cardiovascular disease (CVD) (14, 17, 18, 22, 31–37, 39). One study partially excluded individuals with CVD by only excluding those with coronary heart disease (15), whereas two other studies included individuals with CVD in their study population (23, 38).

The participants involved in the studies exhibited a considerable range in size, with sample sizes ranging from 208 to 18,504. Additionally, the studies displayed diversity in the mean age of the participants, ranging from 38.8 to 63.8 years. In terms of the analysis of HOMA-IR, eight studies used categorical variables (14, 17, 18, 22, 31, 34, 38, 39), whereas five used continuous variables (15, 23, 33, 35, 36) and two used both (32, 37). There were 12 studies that reported CAC prevalence (17, 18, 22, 23, 32–39), two reported CAC progression (14, 31), and one reported CAC prevalence and progression in both (15).

The studies used different criteria to determine the prevalence and progression of CAC, with all studies utilizing the Agatston score methodology for assessment. For the prevalence of CAC, 10 studies used an Agatston score above 0 as the criterion for “judging” (15, 17, 18, 22, 32, 34, 35, 37–39), one had a CAC score more than 10 (33), and two studies classified it into multiple risk classes based on the Agatston score (23, 36). Regarding the progression of CAC, three studies had their criteria for judging the progression of CAC (14, 15, 31). The NOS score for the selected studies ranged from 7 to 9. Details are presented in Table 1 .

When analyzing HOMA-IR as a categorical variable, a fixed-effects model was utilized to pool data from eight studies (14, 17, 18, 22, 32, 34, 37, 39). The outcomes of the analysis showed that individuals who belonged to the highest HOMA-IR group demonstrated an augmented occurrence of CAC among subjects who fell within the lowest classification [OR: 1.13, 95% CI 1.06–1.20, I² = 29%, P <0.001; Figure 2A ]. A similar outcome was obtained when HOMA-IR was treated as a continuous variable (15, 23, 32, 33, 35–37). The application of a random-effects model to analyze the data revealed that higher HOMA-IR values were also correlated with an elevated prevalence of CAC [OR: 1.27, 95% CI: 1.14–1.41, I² = 54%, P < 0.001; Figure 2B ].

Upon conducting sensitivity analyses, we observed that removing any of the studies did not significantly alter heterogeneity or pooled OR values. Specifically, the ORs for categorical variables ranged from 1.11 to 1.18. Likewise, the ORs for continuous variables ranged from 1.24 to 1.30. All of the above analyses demonstrate P values less than 0.001. See Tables 2 and 3 .

To account for the potential impact of different clinical characteristics of study subjects on the results, we conducted a subgroup analysis (refer to Table 4 ). Our analysis revealed that the source of study participants did not significantly affect the pooled ORs in either community or non-community residents, physical examination, or non-physical examination population (all P > 0.05).

We conducted a risk factor analysis for CVD. Our analysis included examining whether participants with CVD were excluded from the study population. When HOMA-IR was used as a continuous variable, there was no statistical difference in subgroups between complete exclusion and partial exclusion/no exclusion of participants with CVD (P > 0.05). When HOMA-IR was used as a categorical variable, one study did not exclude patients with CVD from the population (38). However, our sensitivity analysis in Table 2 showed that excluding this study did not significantly change the results. Considering that IR might differ in diabetes, we performed subgroup analyses according to whether participants had diabetes or not. The results showed that the association between IR and CAC was statistically significant only among the non-diabetic subgroup (OR: 1.29, 95% CI: 1.17–1.42) but not in the diabetes subgroup (OR: 1.22, 95% CI: 0.72–2.05).

Due to data limitations, subgroup analyses for people with or without hypertension and different lipid profiles could not be performed. However, out of all the studies in the multifactorial analysis, only one did not adjust for hypertension/blood pressure (36) and the other did not adjust for lipid status (23). Notably, excluding these two studies from the sensitivity analysis did not affect the overall combined-OR value.

In addition, differences in CAC prevalence evaluation criteria among the included studies were considered. Subgroup analysis showed that an Agatston score >0 did not show a difference from the subgroup with a non-Agatston score > 0.

A meta-analysis of two longitudinal cohort research (14, 31) showed that HOMA-IR was statistically related to CAC progression when used as a categorical variable (OR: 1.44, 95% CI 1.04–2.01, I² = 21%, P < 0.05). In addition, a cohort study (15) of community residents investigating HOMA-IR as a continuous variable identified its independent relationship with CAC progression (OR: 1.25, 95% CI: 1.01–1.55, P < 0.05).

In order to evaluate the probability of publication bias, we employed funnel plots. The scatter on the funnel plot displayed a visually balanced distribution, thereby suggesting a reduced likelihood of publication bias ( Figure 3 ). Additionally, the outcomes from the Egger regression analysis remained consistent when examining HOMA-IR as both a categorical and continuous variable (P =0.149 and 0.297, respectively).

Of the downgrading factors regarding the CAC prevalence, in terms of study limitations and indirectness, the included studies have a broad representation of the population and use standard measurement methods for both exposure and outcome indicators, which reduces potential concerns. At the same time, the point estimates of the included studies show high similarity, with overlapping confidence intervals in most studies. Additionally, there was low to moderate heterogeneity among the included studies, all of these indicating consistencies. The study’s large sample size also mitigates concerns about imprecision. Furthermore, there is no evidence of publication bias, eliminating any concern in this regard. However, the quality of evidence cannot be upgraded as the pooled OR values are all less than 2. Confounding factors pose a challenge as not all studies excluded patients with cardiovascular disease and only some considered insulin/hyperglycemic drugs. These factors prevent us from eliminating concerns about plausible residual biases. Furthermore, the inability to evaluate the dose–response relationship also hinders its upgrade. Regarding CAC progression, we have concerns about potential bias in outcome measures because of the use of different Agatston method-based evaluation methods in each of the three included studies. Therefore, the quality of evidence rating will directly be very low.