Five databases were used to search for relevant articles, including Medline, Embase, PsycINFO, Cochrane Central Register of Controlled Trials (CENTRAL), and Allied and Complementary Medicine (AMED). The search strategy used was the combination of related keywords of “mindfulness” and “loneliness,” e.g., (loneliness OR lonel^* OR UCLA Loneliness Scale) and (mindfulness OR meditation or transcendental meditation). The complete search strategy was reported in Supplementary Table 1. Studies to be included must be (1) randomized controlled trial, (2) which recruited human participants (of any age, with or without any health condition) intervened with any intervention of mindfulness, (3) compared with a group without mindfulness component, and (4) with the assessment on loneliness.

Two authors (SLT and VL) working independently used a standardized data extraction sheet to extract the trials' characteristics and results. Any disagreement was resolved by discussion to reach a consensus. A third author's opinion (LHL) was sought after when needed. The authors extracted the information on study design, blinding status, participants' characteristics, interventions, comparators, clinical assessment, and the outcomes at baseline and postintervention. The clinical evaluation related to loneliness was the primary outcome. Besides, any adverse effect reported in the trials was considered the secondary outcome of interest.

The methodological quality of each trial was assessed by two independent reviewers (SLT and VL) using the Cochrane Risk of Bias (ROB) Tool 2.0 (Higgins et al., 2011; Sterne et al., 2019). The methodological evaluation domains included randomization, the effect of adhering to intervention, missing outcome data, outcome, and selection of the reported result (Higgins et al., 2011; Sterne et al., 2019). The funding of the trials was assessed within the domain of “other sources of bias.” Each trial was classified as having low risk (low ROB for all domains), high risk (high ROB for 1 or more domains), or some concerns (some concerns for one or more key domains, given no high ROB in any domain) (Higgins et al., 2011; Sterne et al., 2019).

The results were expressed as mean differences (MDs) with 95% confidence intervals (CIs) to determine the effect of mindfulness on loneliness and a continuous outcome. The change from baseline was compared between the mindfulness group and the comparator group. Data from trials measured using the same loneliness scales were pooled in a meta-analysis and expressed as MDs, using an inverse-variance method with a random-effects model (DerSimonian and Laird, 1986). Data from trials measured using similar loneliness scales were pooled in a meta-analysis using standardized mean difference (SMD). Using SMD, 0.20 indicated a small effect, 0.50 a moderate effect, and 0.8 a large effect (Cohen, 2013). The heterogeneity of the included trials was assessed using the chi-squared test and the I² test. For the chi-squared test, p ≤ 0.10 indicated statistically significant heterogeneity (Higgins et al., 2019). An I² value of more than 50% revealed substantial heterogeneity (Higgins et al., 2019). Subgroup analysis and explorative analysis were performed to add or remove any heterogeneity in participants, interventions, comparators, and outcome measurements.

For a meta-analysis with at least 10 trials included, publication bias was assessed using Egger's test (Egger et al., 1997) to calculate the significance level of funnel plot asymmetry, where p < 0.10 indicates significant funnel plot asymmetry (Sterne et al., 2011). The software used for data analysis was Stata version 14 (StataCorp; College Station, Texas, USA).

The overall quality of evidence was assessed independently by two authors (SLT and VL) based on the domains of study design, ROB of individual trials, heterogeneity, the directness of evidence, precision of effect estimates, and possibility of publication bias, using the Grading of Recommendations Assessment, Development, and Evaluation (GRADE) approach (Andrews et al., 2013). The overall quality of evidence ranged from high, moderate, low to very low. The high quality indicates a high degree of certainty that the estimated effect lies close to the true effect. In contrast, low quality means substantial uncertainty about the estimated impact (Guyatt et al., 2008).

Table 1 summarizes the characteristics of the eight included trials (Creswell et al., 2012; Jazaieri et al., 2012; Dodds et al., 2015; Mascaro et al., 2018; Zhang et al., 2018; Lee et al., 2019; Lindsay et al., 2019; Pandya, 2019). The trials were conducted in the USA (n = 6) (Creswell et al., 2012; Jazaieri et al., 2012; Dodds et al., 2015; Mascaro et al., 2018; Zhang et al., 2018; Lindsay et al., 2019), India (n = 1) (Pandya, 2019), and Korea (n = 1)(Lee et al., 2019).

The sample size of the trials was generally small, ranging from 33 participants (Dodds et al., 2015) to 153 (Lindsay et al., 2019) participants, except for one trial (Pandya, 2019) with a relatively bigger sample size of 378 participants. Two trials recruited younger populations who were students (Mascaro et al., 2018; Zhang et al., 2018), while the other six trials were adults (Creswell et al., 2012; Jazaieri et al., 2012; Dodds et al., 2015; Lee et al., 2019; Lindsay et al., 2019; Pandya, 2019). One trial recruited only women with a history of breast cancer (Dodds et al., 2015), while the other seven trials recruited both genders (Creswell et al., 2012; Jazaieri et al., 2012; Mascaro et al., 2018; Zhang et al., 2018; Lee et al., 2019; Lindsay et al., 2019; Pandya, 2019). In terms of mental health conditions, one trial recruited participants with a social anxiety disorder (Jazaieri et al., 2012), one trial with depression (Pandya, 2019), one trial with elevated stress level (Lindsay et al., 2019), and one trial with elevated loneliness level (Zhang et al., 2018). Three trials recruited participants with no obvious mental or cognitive conditions (Creswell et al., 2012; Dodds et al., 2015; Mascaro et al., 2018), and one trial did not report the characteristics of included participants (Lee et al., 2019).

The intervention of two trials was mindfulness-based stress reduction (MBSR) (Creswell et al., 2012; Jazaieri et al., 2012), two cognitively based compassion training (CBCT) (Dodds et al., 2015; Mascaro et al., 2018), one mindfulness-based cognitive therapy (MBCT) (Zhang et al., 2018), one brain education-based meditation (BEM) (Lee et al., 2019), one mindfulness meditation with an orientation of acceptation (Lindsay et al., 2019), and one yoga (Pandya, 2019). Almost all of the interventions (7/8) (Creswell et al., 2012; Jazaieri et al., 2012; Dodds et al., 2015; Mascaro et al., 2018; Zhang et al., 2018; Lee et al., 2019; Pandya, 2019) were conducted as a group session and one (Lindsay et al., 2019) smartphone-based intervention. All interventions consisted of guided meditations. However, for MBSR (Creswell et al., 2012; Jazaieri et al., 2012), MBCT (Zhang et al., 2018), and CBCT (Dodds et al., 2015; Mascaro et al., 2018), they consisted of the addition of yoga and stretching, group discussions, and a day-long retreat. MBSR (Creswell et al., 2012; Jazaieri et al., 2012) and MBCT (Zhang et al., 2018) also consisted of a day-long retreat. Notably, BEM (Lee et al., 2019) and Yoga (Pandya, 2019) had the additional exercise or stretching component in addition to meditation. More than half of the interventions (5/8) (Creswell et al., 2012; Jazaieri et al., 2012; Dodds et al., 2015; Mascaro et al., 2018; Zhang et al., 2018) carried out once-weekly sessions for around 2 h, except one with once weekly for 45 min (Pandya, 2019), one with twice weekly of 75 min (Lee et al., 2019), and one with once daily of 20 min (Lindsay et al., 2019). Almost all (7/8) interventions (Creswell et al., 2012; Jazaieri et al., 2012; Dodds et al., 2015; Mascaro et al., 2018; Zhang et al., 2018; Lee et al., 2019; Pandya, 2019) took at least 8 weeks with one with a longer duration of 2 years (Pandya, 2019), and only one of 2 weeks (Lindsay et al., 2019).

Three trials (Creswell et al., 2012; Dodds et al., 2015; Mascaro et al., 2018) used wait-list (i.e., control group's participants were placed on a wait-list while the trial was ongoing) as a comparator. Three other trials used active control group which included aerobic exercise (Jazaieri et al., 2012), health education class (Lee et al., 2019), and guidance in reflective thinking and problem solving without mindfulness content (Lindsay et al., 2019). One trial (Pandya, 2019) assigned no intervention to the comparator group, while the other trial (Zhang et al., 2018) did not mention the comparator type.

Based on the assessment using Cochrane's ROB tool version 2.0, almost all trials (7/8) (Creswell et al., 2012; Jazaieri et al., 2012; Dodds et al., 2015; Mascaro et al., 2018; Zhang et al., 2018; Lee et al., 2019; Pandya, 2019) had high ROB and one trial (Lindsay et al., 2019) with some concerns (Table 2). The trial had some concerns in randomization components with low ROB for all other components.

In the randomization component, generally appropriate sequence randomization was used. However, almost all trials (6/8) (Creswell et al., 2012; Jazaieri et al., 2012; Dodds et al., 2015; Mascaro et al., 2018; Zhang et al., 2018; Pandya, 2019) most likely did not conceal the allocation while one trial (Lindsay et al., 2019) did not specify and only one trial (Lee et al., 2019) mention about the appropriate randomization concealment.

For the effect of adhering to intervention, only one trial had low ROB (Lindsay et al., 2019), while the rest (7/8) had some concerns in nonadherence to the assigned intervention regimen that could have affected participants' outcomes. The trial with low ROB (Lindsay et al., 2019) provided raw data for baseline and postintervention for all patients, including those who have dropped out, to enable an appropriate analysis to estimate the effect of adhering to an intervention.

All trials had low ROB for missing outcome data, with outcome data available for all or nearly all participants randomized. All trials also had ROB for outcome domain, as all trials have used appropriate and only objective assessment (loneliness measurement scale). For the selection of the reported results, only one trial (Lindsay et al., 2019) had pre-specified and registered protocol while other trials (7/8) (Creswell et al., 2012; Jazaieri et al., 2012; Dodds et al., 2015; Mascaro et al., 2018; Zhang et al., 2018; Lee et al., 2019; Pandya, 2019) did not mention the availability of any pre-specified analysis plan. The details of the ROB assessment were available in Supplementary Table 2.

Two established loneliness scales were employed, with UCLA loneliness scale being the most common scale used to measure loneliness in five trials (Creswell et al., 2012; Jazaieri et al., 2012; Dodds et al., 2015; Mascaro et al., 2018; Lindsay et al., 2019). The other trials used De Jong Gierveld Loneliness Scale (Pandya, 2019), mental health test with a loneliness component (Lee et al., 2019), and a loneliness test in the author's indigenous language (Zhang et al., 2018). All except one trial used scales that indicated higher score with more loneliness (including UCLA loneliness scale) while the exceptional trial used scale showed lower scores with more loneliness (i.e., De Jong Gierveld Loneliness Scale).

Referring to the effects of individual trials as shown in Table 3, half of the trials (4/8) (Creswell et al., 2012; Mascaro et al., 2018; Zhang et al., 2018; Pandya, 2019) showed significant loneliness reduction after mindfulness intervention compared with the comparator group, and they shared the common feature of the intervention of at least 8 weeks. Another half of the trials (Jazaieri et al., 2012; Dodds et al., 2015; Lee et al., 2019; Lindsay et al., 2019) did not show significant changes.

Referring to Table 4, the main pooled analysis combining three trials in participants with no known mental health conditions, which employed slightly varied mindfulness interventions (i.e., two CBCT and one MBSR) and compared with wait-list, showed significant improvement in loneliness score reduction using UCLA-R scale with MD of −6.33 (95% CI: −9.39, −3.26), I² = 0.0%, p = 0.688; three trials; Grade low) (Supplementary Figure 1). There was no significant publication bias using Egger's test (p = 0.602). However, this value was only indicative owing to the small number of included studies in the meta-analysis. A subgroup analysis with only CBCT intervention, with the removal of one trial which employed MBSR, also showed similar results of MD of −6.05 (95% CI: −9.53, 2.58, I² = 0.0%, p = 0.425; two trials; Grade low) (Supplementary Figure 2). A subgroup analysis with only participants with mental health conditions found no significant improvement with small effect in loneliness score reduction using varied scales with SMD of −0.23 (95% CI: −0.80, 0.33), I² = 62.8%, p = 0.068; three trials; Grade very low) (Supplementary Figure 3). Another subgroup analysis comparing younger populations with adults and elderly showed significant improvement with large effect only in younger populations with SMD of −0.85 (95% CI: −1.36, −0.35), I² = 0.0%, p = 0.751; two trials; Grade low) (Supplementary Figure 4). No significant improvement with small effect in loneliness score reduction was found in adults and elderly with SMD = −0.12 (95% CI: −0.43, 0.19), I² = 18%, p = 0.300; five trials; Grade low) (Supplementary Figure 5) or SMD = −0.15 (95% CI: −0.46, 0.15), I² = 15.6%, p = 0.315; five trials; Grade low) (Supplementary Figure 6). Two pooled estimates were available as two estimates were reported from one trial with loneliness-related questions. Table 5 explains in detail on the GRADE.

Four explorative analyses were conducted by varying some components of the PICO (Table 4). Specifically, in explorative analysis 1, when 1 additional trial which used a slightly different comparator (i.e., aerobic exercise), a nonsignificant MD of −3.74 (95% CI: −8.45, 0.98, I² = 64.3%, p = 0.039; four trials) (Supplementary Figure 7). In explorative analysis 2, when one trial which utilized UCLA-8 scale (a simplified version of UCLA-R scale) was added to explorative analysis 1, also a nonsignificant SMD of −0.33 (95% CI: −0.76, 0.10, I² = 53.6%, p = 0.071; five trials) (Supplementary Figure 8) was obtained. Explorative analysis 3, involved two trials that employed MBSR interventions and wait-list control. All three analysis recruited different mental health status (one with no known mental health illness and one with SAD) and employed slightly different loneliness scale (one with UCLA-R and one with UCLA-8), with a pooled nonsignificant SMD of −0.48 (95% CI: −1.02, 0.06; I² = 13.9%, p = 0.281) (Supplementary Figure 9). In explorative analysis 4, seven trials which utilized the loneliness scales, which showed a similar trend of an increasing score with increase loneliness were pooled. Two pooled estimates were available as two estimates were reported from one trial with loneliness-related questions. The two different estimates produced a nonsignificant SMD of −0.34 (95% CI: −0.69, 0.01; I² = 48.8%, p = 0.068) (Supplementary Figure 10) and a significant SMD of SMD = −0.36 (95% CI: −0.70, −0.03; I² = 45.6%, p = 0.088) (Supplementary Figure 11), respectively.