This study protocol was prospectively registered with the PROSPERO (CRD42019121522). Search strategy, study selection, data extraction, risk of bias assessment, and data analysis were performed in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines.

The prepiloted search strategy was designed with an experienced medical librarian with an aim of high sensitivity (Supplementary Material). We searched the Ovid MEDLINE, EMBASE, Web of Science Core Collection, and the Cochrane Central Register of Controlled Trials from inception to August 2020. Reference lists of previous systematic reviews and included studies were reviewed, but gray literature sources were not included. Covidence systematic review software (Veritas Health Innovation, Melbourne, Australia) was utilized during eligibility assessment. Microsoft Excel and Endnote were utilized for data collection, data checking, and reference management.

Studies that included participants with ≤10 mm UIAs in both the anterior and posterior circulation that did not have neurosurgical or endovascular repair and underwent follow-up were evaluated for inclusion. Randomized and non-randomized clinical trials, prospective and retrospective cohort studies, and case series with >10 patients published in English were evaluated. We included studies that reported aneurysm rupture as an outcome. When aneurysm rupture was reported as a clinical outcome in a mixed cohort of unruptured intradural and cavernous internal carotid aneurysms, we only included data if > 85% of the cohort harbored intradural aneurysms. For studies that included data published more than once, the study report with the most informative dataset was included. Two investigators independently evaluated studies for eligibility against the pre-specified inclusion criteria. Initially, titles and abstracts were screened. Full-text reports were obtained for all the potentially relevant reports and when there was ambiguity on whether or not the articles met inclusion criteria, disagreements were resolved by consensus.

Two investigators performed independent data extraction using a pre-piloted standardized electronic datasheet that included: study author, publication year, study start and end years, country of source population, study design, number of patients, mean or median age, number of aneurysms, number of patients with multiple aneurysms, number of patients with prior SAH, mean or median follow-up, and number of aneurysms that ruptured during follow-up. When aneurysm data were stratified by size, data were extracted under the following size thresholds commonly reported in the literature: ≤10, ≤7, ≤5, and ≤3 mm. Aneurysm shape and anatomical location data were also extracted where available. Irregular aneurysms were defined as UIAs that were reported as irregular, multilobular, or with a daughter sac or aneurysm bleb. For anatomical location analysis, the total number of anterior and posterior circulation UIAs of ≤10 mm was extracted. If specific anatomical location was reported, the total number of anterior communicating artery and anterior cerebral artery aneurysms for UIAs of ≤10 mm was extracted. Discrepancies were resolved by consensus. To improve accuracy of analyses for size, we included details from additional publications of the same study cohort and emailed the corresponding authors of original publications. A separate investigator independently checked the data for consistency and completeness following double data extraction and receipt of unpublished data. An experienced investigator performed all the data analyses, which were checked independently by a separate experienced investigator.

To assess the quality of the included non-randomized studies, the pre-piloted modified Newcastle–Ottawa Scale (NOS) (9) was used (Supplementary Material). Two investigators independently rated included studies on selection, comparability, and outcome and recorded information on each study to justify the judgment made. Any disagreements were resolved by consensus. The NOS ratings were categorized as good quality, fair quality, and poor quality based on the Agency for Healthcare Research and Quality (AHRQ) standards (10).

The pre-specified outcome was rupture of the index UIA included at study entry. The proportion of UIAs that ruptured was extracted from each study. Aneurysmal SAH from de novo, dissecting, or fusiform aneurysms was excluded. A post-hoc decision was made to perform a per aneurysm analysis, since one aneurysm in one patient ruptures at one time, while a patient may harbor multiple aneurysms. Aneurysm multiplicity was taken into account by extracting the total number of aneurysms for UIAs of ≤10 mm. If this was not reported, the total number of aneurysms in the ≤10 mm cohort was derived by applying the proportion of multiple aneurysms in the total observation cohort to the number of patients. If this was not reported, then patients with multiple aneurysms were assumed to harbor two UIAs. For aneurysm shape analysis, if the total number of irregular aneurysms for UIAs of ≤10 mm was not reported, the total number of aneurysms in the ≤10 mm cohort was derived by applying the same proportion of irregular aneurysms in the total observation cohort. For anatomical location analysis, if anatomical location data were not reported for UIAs of ≤10 mm, the number of aneurysms in the ≤10 mm cohort was derived by applying the same proportion of the anatomical location information in the total observation cohort. Median follow-up values were used as best estimates when mean values were missing (11). If mean follow-up was not specifically reported for the ≤10 mm cohort, the mean follow-up of the total observation cohort for all the aneurysm sizes was utilized.

The I² statistic was used to identify heterogeneity, which was categorized as low (<25%), moderate (25–75%), or significant (>75%) (12). Heterogeneity was explored using outlier and influence analyses. If a reasonable rationale due to clinical and/or methodological diversity was evident, the outlier study was excluded from data synthesis to reduce bias in the pooled proportion estimate and additional sensitivity analyses were performed with and without outlier studies (13).

Random-effects data synthesis was carried out using a random intercept logistic regression model for meta-analysis of proportions. The Wilson procedure was utilized for the 95% CIs. Random-effects meta-regression was used to explore categorical or continuous covariates for residual between-study heterogeneity. Additional information is given in the Supplementary Materials. Prespecified covariates included aneurysm size and exposure to prior SAH. Additional covariates included aneurysm shape and anatomical location. Residual heterogeneity was identified using the I² statistic. p-values were two-sided with values < 0.05 that are considered as statistically significant. Statistical analyses and graphical output were performed in R (version 3.6.3) using the following packages: dmetar (14), meta (15), and metafor (16).

Prespecified sensitivity analysis was performed using size thresholds of ≤7, ≤5, and ≤3 mm. Additional post-hoc sensitivity analyses were performed including outlier studies, for high-quality studies, and using the leave-one-out method to confirm that the pooled rupture proportion was not driven by a single study. Reporting bias and small-study effects were examined by visual assessment of a funnel plot, calculation of Egger's regression intercept, and examination of the random-effects and fixed-effects estimates (17).

Baseline study characteristics are given in Table 1. In the 32 studies, source populations were Japan (7 studies) (23–29) United States (7 studies) (20, 22, 30–34), international (5 studies) (3, 35–38), Korea (4 studies) (39–42), Australia (2 studies) (43, 44), and one each from China (45), Netherlands (46), United Kingdom (47), Switzerland (18), Germany (48), Poland (21), and Finland (19). Majority (23 studies) were retrospective and of the prospective studies, two of the three largest studies were from Japan (26, 27). There were no randomized controlled trials. Participants were included in studies from 1956 (19) to 2019 (44). Study publication occurred between 1995 and 2020, with majority (21 studies) published in the last decade.

Across the 32 studies, a total of 12,214 participants with 16,615 UIAs were managed without repair across all the size strata. Of these, 13,966 were ≤10 mm UIAs with mean follow-up of 4.3 years (range 1.4–21 years). Overall, 70.8% of participants were female; mean age was 58.3 ± 7.1 years. The mean proportion of included patients included with prior SAH was 14.5% (range 0–92.3%).

Twelve studies were judged of the good AHRQ standard with lower risk of bias (Table 2). Only one study scored three stars in the outcome domains having complete follow-up (19).

Preliminary data synthesis included 32 studies with 229 rupture events in 13,966 ≤10 mm UIAs over a mean of 4.3 years with significant heterogeneity (I² = 85%) (Supplementary Figure 1). Outlier analysis identified an outlier (19), with a study rupture proportion of 18.7% (95% CI 13.5–25.3) compared to the pooled proportion of 1.1% (95% CI 0.7–1.7). With leave-one-out analysis, this was the only study that contributed greatly to heterogeneity, with I² reduction from 85 to 53% (Supplementary Figure 2). Further outlier and influence diagnostics also confirmed these findings (Supplementary Figures 3, 4).

Examination of the outlier study (19) revealed underlying clinical and methodological diversity responsible for between-study heterogeneity. All the patients in this study were managed without aneurysm repair due to lack of treatment availability. In the remaining 31 studies, treatment was considered and a clinical decision made to manage the patient without aneurysm repair. In addition, the outlier study had the longest follow-up (21 vs. mean 3.7 years), the highest proportion of exposure to prior SAH (92.3 vs. mean 11.9%) and an earlier recruitment period (1956–1978 vs. median midpoint 2006). A post-hoc decision was made to exclude this outlier study (19) from the main analysis because of lack of availability of aneurysm repair at the time of the study and to appropriately address heterogeneity (13).

The meta-analytic proportion of ruptures without aneurysm repair was synthesized across the remaining 31 studies and included 198 rupture events in 13,800 ≤10 mm UIAs (Figure 2). Aneurysm rupture occurred in 1.1% of ≤10 mm UIAs (95% CI 0.8–1.5; I² = 53%) over mean 3.7 years of follow-up.

Meta-regression included the following study-level covariates: proportion of patients with exposure to prior SAH, proportion with aneurysm size ≤5 mm, proportion with multiple aneurysms, proportion with irregular aneurysms, proportion with posterior circulation aneurysms, proportion with anterior circulation aneurysms, proportion with anterior communicating artery or anterior cerebral aneurysms, mean age, source population categorized as Japanese or non-Japanese, study type categorized as prospective or retrospective, and length of follow-up.

In subgroup analysis according to whether or not the source population was Japanese, a larger proportion of ruptures was identified in Japanese compared to non-Japanese populations during follow-up [test for subgroup differences, p < 0.001, residual I² = 0% (95% CI 0.0–19.4%)] (Supplementary Figure 5). However, the mean age in Japanese samples (64.5 years) was greater than in non-Japanese samples (57.4 years, p = 0.001).

Associations between other covariates in the remaining meta-regression analyses including proportion with prior SAH, proportion with ≤5 mm aneurysms, proportion with multiple aneurysms, proportion with irregular aneurysms, proportion with posterior circulation aneurysms, proportion with anterior circulation aneurysms, proportion with anterior communicating artery or anterior cerebral aneurysms, study type, study-level age as a continuous variable, and follow-up time as a continuous variable were not detected (Supplementary Figures 6–15).

In leave-one-out sensitivity analysis, the meta-analytic proportion of ruptures remained consistent with no clinically relevant effect (Supplementary Figure 16). Heterogeneity could be further reduced by exclusion of Murayama et al. (27) (residual I² = 42%) or Jeon et al. (41) (residual I² = 44%), but these levels of reduction in heterogeneity were considered unimportant.

Findings were consistent in prespecified sensitivity analyses at all the size thresholds (Supplementary Figures 17–19). For ≤7 mm UIAs managed without repair, data were synthesized across 28 studies, 134 rupture events, and 11,371 UIAs: rupture occurred in 0.9% UIAs (95% CI 0.7–1.3; I² = 49.1%) over 3.6 years of follow-up. For ≤5 mm UIAs managed without repair, data synthesis occurred across 24 studies, 74 rupture events, and 7,280 UIAs: rupture occurred in 1.0% UIAs (95% CI 0.8–1.3; I² = 0%) over 3.5 years of follow-up. For ≤3 mm UIAs managed without repair, data synthesis occurred across 18 studies, 10 rupture events, and 1,228 UIAs: rupture occurred in 0.8% UIAs (95% CI 0.4–1.5; I² = 0%) over 3.8 years of follow-up.

There were also consistent results in post-hoc sensitivity analyses that included the outlier study (Supplementary Figure 20). Notably, in sensitivity analyses limited to high-quality studies, the pooled proportion of rupture was greater, 1.8% (95% CI 1.5–2.0; I² = 0%) over 3.9 years (Supplementary Figure 21).

Funnel plot distribution was asymmetrical on visual assessment and confirmed on Egger's test of the intercept (Supplementary Figure 22). Multiple possible sources of asymmetry other than random chance were identifiable. These include non-reporting bias, selective reporting bias, small-study effects, and residual heterogeneity. The presence of small-study effects was investigated with additional post-hoc sensitivity analysis considering a fixed-effects model estimate with no clinically relevant effect on the meta-analytic proportion of ruptures (Supplementary Figure 23).