Studies were included in the systematic review if they met the following criteria:

Only the prospective cohort study was eligible as it is the most appropriate design for examining the causal relationship between exposure and risk of disease. A pilot search found that a considerable number of prospective cohorts were published.

The participants at baseline were adults (≥18 years old) without a preexisting stroke.

The exposure group had one of the following phenotypes: MHOW, MHO, MUNW, MUOW, or MUO, and the reference group was MHNW. Overweight/obesity was defined by BMI in compliance with the standards from one of the following organizations: the International Obesity Task Force (IOTF), Working Group on Obesity in China (WGOC), and the World Health Organization (WHO). Currently, there is no consensus regarding the definition of metabolic health. We endorsed the methodology of the previous systematic reviews (16), and ”metabolic health was defined as individuals with less than two abnormal metabolic parameters.” This enabled a direct comparison between the previous evidence and our study.

The primary outcome was the first-ever stroke, which the participants could report—a clinical diagnosis confirmed by clinical signs/symptoms accompanied by brain images or a diagnosis in the medical records/national databases. The patients with ischemia, hemorrhage, or mixed stroke were eligible.

Studies were excluded if they were repeatedly published.

We searched the PubMed, Embase, and the Cochrane Library databases from their inception to 7 March 2021. The search strategy is presented in Supplementary Table 1. Two researchers (MMM and GYX) independently screened the literature by removing duplicates, reading the title/abstract, and reviewing the full text. We calculated the inter-rater agreement with the reference of kappa value, and the disagreements were resolved through discussion or consultation with a third researcher (NXJ).

Two researchers (MMM and GYX) extracted the data using the double-entry method and were checked by a third researcher (NXJ). The data items included author, publication year, country, sample size, mean or range of age, sex, period of baseline data collection, follow-up duration, overweight/obesity definition, definition of (un)metabolic health, the confounding factors that were adjusted for in the multivariable analysis, and adjusted or non-adjusted effect sizes [hazard ratio (HR) and relative risk (RR)] with the corresponding 95% confidence intervals (CIs).

Two researchers (KZR, LLL) independently evaluated the methodological quality of the included studies using the Newcastle–Ottawa Scale (NOS), a tool for assessing the quality of observational studies in meta-analyses (21). It was composed of three domains (eight items): selection of the study groups, comparability, and ascertainment of the outcomes for the cohort study. The assessment outcomes ranged from zero stars up to nine stars. Studies with seven or more stars were considered high quality. We calculated the inter-rater agreement concerning the intraclass correlation coefficient (ICC), and any disagreement was resolved by discussion or consultation with a third researcher (MMM).

Pooled HRs with 95% CIs were calculated using a random-effects model to assess the risk of stroke for MHOW, MHO, MUNW, MUOW, and MUO compared to MHNW. The heterogeneity was evaluated using Cochran’s Q test and I² statistics. If the heterogeneity was considered significant (p < 0.1 or I² > 50%), subgroup analyses were performed according to the duration of the follow-up, number of abnormal metabolic components, criteria for defining overweight/obesity, research sites, and proportion of females to explore the potential sources of heterogeneity. A sensitivity analysis was performed based on the methodological quality of the included studies and the leave-one-out method. The publication bias was evaluated by the visual inspection of asymmetry in the funnel plots and Egger’s test for at least 10 studies. The statistical significance was set at p < 0.05. All data analyses were conducted using Stata/MP version 16.0.

We initially identified 3,381 records from the database search and finally included 11 eligible cohort studies in the systematic review (13–15, 17–19, 22–26). The inter-examiner agreement between the two reviewers was good (kappa = 0.762, 95% CI: 0.5464–0.9776). The results from nine cohorts were quantitatively synthesized as two studies reported the relative ratio (RR) instead of HR (22, 25). A full flowchart of the study selection process is shown in Figure 1. A list of the excluded studies is shown in Supplementary Data Sheet 1.

The sample sizes of the included cohorts ranged from 5,171 to 3,495,777 participants. The included studies were conducted in different geographic locations: four studies were conducted in Europe (14, 15, 24, 25), two in North America (13, 26), and five in Asia (17–19, 22, 23). Female participants were enrolled in only three studies (13, 25, 26). The baseline information was collected between 1980 and 2018, and the duration of the follow-up ranged from 3 to 24 years. The BMI threshold for obesity varied, including BMI ≥ 25 kg/m² (International Obesity Task Force, IOTF) in two studies (17, 23), BMI ≥ 28 kg/m² (Working Group on Obesity in China, WGOC) in three studies (18, 19, 22), and BMI ≥ 30 kg/m² (World Health Organization, WHO) in six studies (13–15, 24–26). The definition of metabolic health was distinct, such as the Adult Treatment Panel III (ATP III) definition in three studies (14, 15, 26), the International Diabetes Federation (IDF) in one study (18), the Japanese diagnostic criteria in one study (17), modified harmonization and the Hammer’s study in three studies (19, 22, 23). The stroke event was reported by the participants in another study (22), diagnosed according to the clinical signs/symptoms and brain images in four studies (13, 14, 23, 24), and was collected from the medical records and registration information in the national databases. The full details of the basic characteristics of the included cohorts are presented in Table 1.

The inter-examiner agreement was very good (ICC = 0.771, 95% CI: 0.3503–0.933). Regarding the methodological quality of the included studies, one study was awarded six stars (17), three studies evaluated eight stars (13, 15, 22), and seven studies achieved nine stars (14, 18, 19, 23–26). The average quality score was eight, suggesting that the methodological quality was generally moderate to high. The results of the quality assessments are presented in Table 2.

We performed a sensitivity analysis by omitting only one study with the lowest methodological quality of the NOS (17). An advanced risk of stroke was detected in the MHO participants compared to the MHNW participants (HR = 1.10, 95% CI: 1.03–1.18, I² = 42.18%, n = 7 cohort studies, 4,531,197 participants) (Supplementary Figure 26), which differed from the primary analysis. An additional sensitivity analysis excluding one study at a time is shown in Supplementary Figure 27. The publication bias was not assessed because fewer than ten included studies in each phenotype.

This systematic review and meta-analysis including nine prospective cohorts showed that metabolically unhealthy concurrent with either overweight or obesity resulted in an increased risk of stroke, which was not affected by the duration of the follow-up, the number of abnormal metabolic parameters, the definition of obesity, geographic locations, or the proportion of females. Although the primary meta-analysis supported the fact that overweight/obesity with metabolic health did not increase the risk of stroke, some subgroup analyses that only included a specific type of participant might present a different result. This study further examined the causal relationship between overweight/obesity and stroke-onset from the perspective of metabolism, which would assist with the stratified management of overweight/obesity and precision prevention for stroke. Compared with a previous study (16), our study had several strengths in terms of study design and data analyses. First, we included more recent cohorts by an updated search, which improved the precision of the estimated effect size; second, we calculated the HR instead of RR, which considered the time-dependent effect; third, we performed comprehensive subgroup analyses to explore the sources of heterogeneity and to test the robustness of the analysis models.

We found that individuals with metabolically unhealthy phenotypes were at a higher risk of stroke, irrespective of the body weight, which was consistent with a previous systematic review (14, 16, 23). This suggests that multiple coexisting metabolic parameters play a leading role in the risk of incident stroke.

However, regarding the phenotypes of MHOW and MHO, our study suggested that the associated risk of stroke was not significantly increased, which differed from a previous systematic review (16). This was partially explained by the updates of the literature search and the different types of outcome data included in the meta-analysis. Furthermore, we observed additional findings when comprehensive subgroup analyses were performed. Concerning the participants with MHOW, a higher risk of stroke was identified by a longer duration of follow-up. This could be explained by the transition of the metabolic health status over long term. As studies have shown, MHOW is probably an unstable and intermediate condition, which could transfer to a metabolically unhealthy phenotype during the follow-up (28). Regarding the subgroup analysis of MHO, we found that MHO led to a higher risk of stroke only when the participants were all females and living in North America. This could be partially explained by the sex differences in the development of stroke. It has been well-recognized that the incidence of stroke is higher in women than in men after the menopausal period (29). The subgroup enrolled middle-aged female participants and continued the follow-up for incident stroke for more than 10 years immediately after their menopausal period. Another reason for the outstanding risk of stroke in the subgroup could be the high transition rate from metabolic health to unhealth in women. The Nurses’ Health Study reported that 84% of the women with MHO and 68% with MHNW were converted to metabolically unhealthy (13). The subgroup analysis also identified an associated risk of stroke for the MHO participants with obesity diagnosed based on the WHO criteria. This was not unexpected because the WHO defined obesity as a BMI of >30 kg/m² (30), which was higher than any other criteria, such as the IOTF and the WGOC standards.

This study has several potential implications in clinical practice. First, metabolically unhealthy and elevated body weight had combined effects on increasing the risk of stroke. This suggests that when a strategy for stroke prevention is developed, monitoring metabolic parameters and achieving optimal body weight should be highlighted. Second, the potential long-term risk of stroke in individuals with metabolic health overweight/obesity should not be neglected, as the metabolic status might change over time, particularly in women. This suggests that maintaining metabolic health is also a goal for individuals with elevated body weight in the primary prevention of stroke. Third, as the homogeneous subgroup analysis showed, women appeared to have a higher risk of stroke when they had MHO, MUOW, and MUO. This suggests that the stroke prevention strategies tailored to female individuals with elevated body weight should be further studied and developed.

This study has limitations. First, a few of the included cohorts performed multiple measurements of metabolic parameters during the follow-up, which might underestimate the impact of metabolic transitions on the risk of stroke. Second, our meta-analysis only calculated the risk of all stroke events, as most studies did not specify the type of stroke, and the confirmation method of the stroke events varied. This may not be sufficient to develop a precise prevention strategy for a specific type of stroke. Third, pooled HRs in this study were estimated by the random model to identify the average effect and the overall trend, therefore, although the statistical results potentially referred to a causal relationship of moderate-to-high strength, the HR value should not be considered as a typical risk effect (31, 32). Fourth, the definition of metabolic health varies, which may be changed alongside the ongoing studies. This suggests that an updated meta-analysis of the best available evidence is needed (33, 34). Finally, we only searched three representative English literature databases, which may have omitted some additional studies. However, methodological studies have suggested that restricting systematic reviews to English-language publications appears to have little impact on the effect estimates and conclusions of systematic reviews (35, 36).