Two reviewers independently (HSX and XYG) searched PubMed, Embase, the Cochrane Library, Chinese National Knowledge Infrastructure (CNKI), Wanfang Database, Chinese Biomedical Literature Database (CBM), and China Science and Technology Journal Database (VIP) from their inceptions to 5 October 2022. Search terms used depression, attention bias modification, and randomized controlled trial. The full search strategies for all databases are shown in Supplementary Appendix 2. We manually searched the reference lists of all identified articles, gray literature, and relevant registration websites² for possible eligible studies. In addition, we consulted the relevant experts for potential studies.

Studies were included if they fulfilled all the inclusion criteria: (1) Patients diagnosed with depression based on the Diagnostic and Statistical Manual of Mental Disorders (DSM) (29), International Classification of Diseases (ICD) (30), Chinese Classification and Diagnosis of Mental Diseases (CCMD), or validated scales (24, 31). There were no restrictions on race, gender, or age. (2) Intervention included ABM alone, or ABM plus conventional treatment (CT). CT contained medication and psychological intervention. There were no limitations on task types, stimulus types, and training directions of ABM. (3) Participants in the control group received attention control training (ACT) alone, ACT plus CT, or CT alone. (4) The primary outcome was depressive symptoms evaluated with widely accepted and validated scales. Secondary outcomes included rumination and attentional control. (5) RCTs that investigated the effect of ABM on patients with depression were included.

Studies were excluded if they met any of the following criteria: (1) studies using interpretation bias modification; (2) cross-over RCTs, review articles, and conference abstracts; (3) overlapping publications; (4) the full texts were not available through various approaches.

Endnote X9 was used to manage the retrieved records. After removing duplicates, two independent reviewers (H-sX and X-bL) screened the titles and abstracts to identify the potential studies. Then, the rest records were scrutinized in full text. Any inconsistency was resolved through consultation with the third reviewer (JL). If multiple publications reported data from the same trial, we included the article with the most complete or latest data.

Two independent reviewers (H-sX and X-yG) extracted data from included studies with a standard extraction form. The following data were extracted: (1) study information: first author, publication year, and country; (2) participant characteristics: diagnostic criteria, sample size, and age; (3) details of interventions: types, paradigms, stimulus types, stimulus directions, sessions, and total trials of per session; (4) comparators: types of intervention, frequency, and duration; (5) primary outcome and secondary outcomes; (6) information related to the risk of bias. With regards to missing data, corresponding authors were contacted via email for missing or incomplete data. For multi-arm RCTs, we extracted the eligible comparisons or the comparison with an inferior effect size. If the data was displayed in the graph, the GetData Graph Digitizer 2.26 was used to extract the data. After cross-checking, disagreements were settled through consultation with an experienced reviewer (Y-xL).

Two researchers (JF and D-lZ) separately evaluated the risk of bias using version 2 of the Cochrane risk-of-bias tool for randomized trials (RoB 2.0). There are five domains in RoB 2.0: randomization process, deviations from intended interventions, missing outcome data, measurement of the outcome, and selection of the reported results. Each domain is rated as “low risk of bias,” “some concerns,” or “high risk of bias.” In case of disagreements, a third investigator (JL) was involved.

The Grading of Recommendations Assessment, Development, and Evaluation (GRADE) system was used to assess the certainty of the evidence of each outcome (32). Each outcome was evaluated from the following five aspects: limitations, inconsistency, indirectness, imprecision, and publication bias. The certainty of the evidence was categorized as “high,” “moderate,” “low,” or “very low.” (33).

Data synthesis was conducted using RevMan (version 5.4) and Stata (version 12.0). Among the included studies, different measurement tools were used to evaluate the symptoms of depression, and the standardized mean difference (SMD) was calculated (34–36). The ability of attention control among the included studies was evaluated using an attention control scale (ACS), and the ruminative symptoms were assessed with a ruminative response scale (RRS); thus, weighted mean difference (WMD) was used to synthesize these results. The uncertainty was expressed with 95% confidence intervals (CIs). The statistical heterogeneity across the included studies was assessed by the Chi-squared test and I² statistic. The fixed-effect model was performed when p > 0.1 or I² values < 50%. Otherwise, the random-effect model was used. Forest plots were used to display the pooled estimates, and a value of p < 0.05 was regarded as statistically significant. If the pooled data could not be synthesized, we conducted the descriptive analysis.

Subgroup analyses were conducted according to age, task types, training target stimuli, training directions, and training settings, whereas meta-regressions were performed based on BDI scores at baseline, publication year, gender, number of training sessions, and number of training trials per session.

The sensitivity analysis was conducted by eliminating studies with a high risk of bias to verify the robustness of the results.

The funnel plot was used to assess possible publication bias when ≥10 studies were included in the analysis.

A total of 2,560 articles were identified. After removing 513 duplicates, 2,004 articles did not conform with the eligible criteria and were excluded. Among the remaining 43 records, 24 studies were excluded after reviewing the full text. Then, we included 20 datasets from 19 reports. The diagram of the screening process is shown in Figure 1. The list of excluded records with reasons is provided in Supplementary Appendix 3.

The characteristics of the included trials are shown in Table 1. A total of 19 trials involving 20 datasets with 1,262 patients with depression were included, wherein Baert et al. (37) observed the effect of ABM for patients with depression diagnosed by DSM-IV criteria and BDI-II scales separately; therefore, we extracted these two datasets, respectively. Among included studies, nine studies were performed in China (17, 18, 39, 43, 44, 48, 49, 51, 52), two were in Belgium (37), two were in the United States (41, 50), two were in the United Kingdom (45, 46), and one in Norway (38), Netherlands (40), Poland (42), Israel (47), and Canada (26). The sample size of these studies varied from 30 to 301. The age of the included patients ranged from 14 to 45. Sixteen studies involved adults (17, 26, 37–39, 41, 42, 44–50, 52), and four studies included adolescents (18, 40, 43, 51). Among included studies, four studies were three-arm trials (17, 40, 41, 48), and the rest studies were two-arm trials. As for comparison, 17 studies compared ABM with ACT (17, 18, 26, 37–42, 45–48, 50–52), two studies compared ABM plus CT versus ACT plus CT (43, 49), and only one study compared ABM plus CT versus CT (44). The number of training sessions ranged from 1 to 28, and the duration of training was between 1 and 4 weeks.

The plot of the risk of bias (RoB 2.0) for each included study is presented in Figure 2, and the proportions of individual studies are presented in Figure 3.

In the randomization process, all included studies showed no statistically significant difference between groups at baseline. Four studies (38, 40, 41, 45) were judged as low risk, while the rest 15 studies were assessed as having some concerns due to no details of randomization or allocation concealment.

Considering the deviation from intended interventions, 14 trials (17, 18, 38–43, 45–47, 50–52) performed appropriate analysis on all randomly assigned participants, which were judged as low risk of bias. Two studies (48, 49) did not report blinding, which was considered as some concerns. The remaining three trials (26, 37, 44) were regarded as high risk due to no blinding in outcome assessors and inappropriate analysis.

As for the missing outcome, 13 studies (17, 18, 26, 37–42, 44, 45, 47, 51) reported the number of drop-outs or lost to follow-up. Among these studies, three studies (38, 40, 41) did not report the details of drop-outs, which were rated as some concerns.

With regard to the measurement of outcomes, six studies (26, 37, 39, 44, 48, 49) were assessed as some concerns due to the lack of a blinding method of outcome assessors. The remaining 13 studies were a low risk of bias.

For the selection of the reported results, six trials (26, 40, 41, 45–47) provided protocol information and reported most of the expected outcomes comprehensively, which were considered as low risk. The rest trials did not provide protocol information, which was assessed as some concerns.

In summary, the overall risk of bias in one trial was considered as low risk, three trials were considered as high risk, and the remaining were considered as some concerns.

The results of the GRADE are shown in Supplementary Appendix 4. The certainty of the evidence of depression (ABM plus CT versus ACT plus CT) was graded as “moderate,” and the rest outcomes were considered as low or very low. The reasons for downgrading were mainly attributed to the risk of bias of included studies and imprecision and publication bias generated by small sample sizes.

Due to limited RCTs, previous meta-analyses demonstrated that ABM had no effect on relieving depressive symptoms (23, 24, 53). In the present meta-analysis, with 20 RCTs included, the synthesized data indicated that ABM might be an effective treatment for depressive disorder. Neurophysiologic studies have confirmed that ABM could modify functional brain connectivity within neural networks related to attentional control (54, 55). Specifically, Beaver’s study (56) found that ABM could reduce negative attention bias and enhance connectivity between the middle frontal gyrus and the dorsal anterior cingulate cortex (ACC). The dorsal ACC involves in attention control through connections with other frontal regions and plays an important role in the cognitive regulation of emotional information. Another study highlighted that greater activation in the lateral prefrontal cortex (PFC) and rostral ACC was observed when the direction of patients’ attention was opposite to ABM training (54). Hakamata et al. discovered that ABM enhanced the pulvinar control over the ventral frontoparietal network (vFPN) to maintain endogenous attention to behavioral targets and diminished the information exchanges between the postcentral gyrus and vFPN to inhibit the capture of exogenous attention by potential threats (57). Moreover, ABM could increase the levels of cortisol awakening responses, which were related to the development and progression of depression (58). Nevertheless, the mechanism of ABM for depression needs further exploration.

Significant reduction in rumination after ABM treatment was noted in our study. Depressive rumination is defined as a maladaptive emotion regulation strategy, which focuses one’s attention on sad mood and negative thoughts (59). The current study revealed that rumination was associated with negative attention bias and attentional control deficits with depression (60). Nolen-Hoeksema et al. (61) conducted a 3-year follow-up visit of 82 patients with depression and found individuals who engaged in rumination were more likely to develop depressive disorders. In addition, several studies concluded that ABM could decrease maladaptive ruminative processing by reducing negative attention bias, thus producing antidepressant effects (17, 62, 63). These findings suggested that ABM was able to promote resilience to the normal pattern of emotional regulation in depression by reducing rumination.

Attentional control is a type of cognitive control schema and defined as the effortful allocation of attention toward goal-relevant information in the face of conflicting prepotent attentional demands (64). Evidence showed that patients with depression manifested hypoactivation in cortical structures of attentional control, which might be related to the impairment of cognitive performance (65). Attentional control appears to have an impact on depressive symptoms through rumination, and poor mood states can be regulated by improving attention control performance (65). Previous studies discovered that ABM might enhance attentional performance through the repetitive activation of neural circuitry with information processing and attentional control (62). However, based on limited studies, the results of our data showed that ABM was not effective to improve attentional control. Wang et al. (48) interpreted that long material presentation times may cause patients with depression to induce attentional avoidance toward negative stimuli in the later stages of attention processing. Therefore, the effect of ABM on attention control requires further investigation.

According to subgroup analysis, ABM training with the dot-probe task was more effective than ACT in reducing depression scores. It is reported that the dot probe has increasingly become an optimal type for attentional modulation (66, 67). However, Robert et al. (68) argued that the dot-probe task was not reliable in measuring reaction time, thus limiting its application in clinical practice. Future studies should identify the reliability of the dot-probe task for depression and compare the effect of different ABM tasks.

Our studies showed that ABM using training direction presented by left–right had a larger effect, while those using top-bottom training did not. Heeren et al. (69) explained that it was more ecologically relevant than processing faces presented horizontally rather than vertically. In contrast, Hakamata et al. (19) and Beard et al. (70) found that top-bottom training had a better effect than those with left–right. Different ABM protocols such as stimuli types or stimuli presentation time may be the reasons for the inconsistent findings.

Regarding training target stimuli, ABM using face stimulus was superior to those using word stimulus in our study. Similarly, Browning et al. (58) found that ABM training with face stimulus reduced the risk of depression, while ABM with word stimulus did not produce such beneficial effects. Jones et al. (71) revealed that ABM studies benefited more from using word stimulus combined with top-down training than those using face stimulus combined with left–right training.

ABM training delivered in the laboratory tended to yield a larger effect than those conducted at home, which was consistent with previous studies (23, 24, 69). Heeren et al. (69) inferred that patients who received ABM training in the standardized laboratory were less susceptible to outside interference.

Of note, ABM appeared to be beneficial for depressive adults and had no effect on adolescents. However, Hang et al. (10) concluded that younger participants could benefit more from ABM as they have a greater potential for attention control. In addition, it has been reported that the cognitive abilities of normal adults may decline in their 20s and 30s (72). Since few studies pay attention to depressive adolescents, more studies are needed to verify this finding.

In addition, the results of meta-regression also showed that BDI scores at baseline were the influencing factor of ABM, and lower BDI scores at baseline yielded a larger effect. Li et al. (73) found that severely depressive individuals exhibited deficits in executive function and attention compared to those with mild depression. A neuroimaging study showed that severe patients with depression manifested dorsolateral prefrontal cortex hypoactivity during attention control (74). This may elucidate why depressive patients with lower BDI benefit more from ABM.

As mentioned earlier, task types, target stimuli, training directions, training settings, age, and BDI scores at baseline were closely associated with the effect of ABM on depression. However, due to limited studies, the optimal protocols and potential influencing factors of ABM for patients with depression were undetermined, and more rigorously designed RCTs are needed to address these issues.

(1) Comparators mentioned in the protocol were sham ABM alone or sham ABM plus conventional rehabilitation or any other active intervention. In this review, participants in the control group received ACT alone, ACT plus CT, or CT alone. Sham ABM and ACT mean the same thing, while the ACT was widely used in the literature (10, 75); thus, we used ACT. (2) The age of the included patients was different. In the protocol, all patients over 18 years of age were included. Previous studies showed that ABM was also widely used for depression in adolescents (18, 40). To enlarge the applicability of this study, we did not impose age restrictions in this review. In addition, we conducted a subgroup analysis by different age groups to identify the advantages of ABM for specific age groups. (3) Different from the protocol, rumination and attentional control were added as secondary outcomes in the review. Rumination and attentional control are closely related to the occurrence and development of depression. The addition of secondary outcomes would facilitate a more broad and more comprehensive exploration of the effect of ABM on depression.

There were several potential limitations in our study. First, high heterogeneity was detected among the included studies; although the subgroup and meta-regression analyses were carried out, we still failed to find out the source of heterogeneity. Second, the risk of bias in most included studies was some concerns, and most of the evidence was low or very low certainty evidence, the findings should be taken with caution. Third, we included studies published in both Chinese and English, and publication bias might exist.