We followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines (48). We submitted our Protocol at the International Prospective Register of Ongoing Systematic Reviews (CRD42021250690). We searched up to February 2022 PubMed, ClinicalTrials.Gov, Embase, PsycINFO, WHO ICTRP database using a search string containing two sets of words referring to 1) schizophrenic patients and 2) metformin that were combined using the Boolean operator “AND.” There was no language, date, document type, or publication status limitations for inclusion of records. Additional articles were collected through the reference lists of reviews and eligible studies we found. We did not plan to contact authors for unpublished data. An example of a search string is fully described in the Supplementary Material S1.

We included only Randomised Controlled Trials (RCT) evaluating people diagnosed with SCZ and related disorders (such as schizoaffective disorder, schizophreniform disorder, and delusional disorder) who were treated with metformin as add-on therapy to antipsychotics. We did not use any criteria for age, nationality or sex of the participants, duration/stage of illness, treatment setting, current clinical state, or symptom clusters. We considered metformin compared to placebo or other types of pharmacological interventions for the treatment of weight gain. We considered behavioural interventions only when combined with a pharmacological intervention. Primary outcomes were the changes in psychiatric and cognitive scales: the psychometric properties of the measuring instrument should have been validated and the measuring instrument should have not been modified for that trial.

All titles and abstracts were assessed independently by two authors (GM, RL) to identify potentially relevant articles. Studies fulfilling the eligibility criteria were included and their full texts were retrieved and reviewed in duplicate (GM, ER). Discrepancies during the check of the two-step independent screening were resolved through the discussion with a third author (VB). Data were extracted by two researchers (RL, VB) and disagreements were resolved by consensus and consultation with the expert group (MP, CC, GC).

For every study the following data were extracted: First author; Year; Study duration; Study type (blinding/design); Number of subjects; diagnosis; number of males; age; antipsychotic(s) used and dose; control/comparator/placebo group; concomitant drugs; additional behavioural interventions; all outcomes of interest; Adverse Drug Reactions. Endpoint data were mainly chosen, mean change data if the former was not available.

This study was designed as an Intention-To-Treat (ITT) analysis. The risk of bias of included studies was assessed by three authors (GM, RL, ER) by using the Cochrane risk-of-bias tool for randomised trials (RoB 2) (49). Disagreements were resolved by consensus among them and a further consultation with the expert group (CC and VB). We planned to conduct a sensitivity analysis excluding studies rated with a high risk of bias if the number of remaining studies exceeds three.

A meta-analysis was performed by using the generic inverse variance method with a random effect model combining psychiatric scales reported by each study, which were Brief Psychiatric Rating Scale (BPRS) (50), Positive and Negative Syndrome Scale (PANSS) (51), Clinical Global Impression Scale (CGI) (52), Global Assessment of Functioning (GAF) (53), Scale for the Assessment of Negative Symptoms (SANS) and Scale for the Assessment of Positive Symptoms (SAPS) (54), Brief Assessment of Cognition in SCZ (BACS) (55) and Patient Health Questionnaire-9 (PHQ9) (56). Among scales, when more than one tool was available, a priority order was defined considering their impact on cognitive assessment. The procedure is clearly described in Supplementary Material S2. Thus, the final priority order adopted was as follows: BACS composite T score > BACS verbal memory T score > PANSS> BPRS > GAF > CGI > SANS and SAPS > PHQ9.

Assumptions were made regarding missing SDs using data of similar studies in terms of population, number of patients, and the point estimate. Forest plots were created for the main outcomes. Sensitivity analyses were performed excluding these studies to check their influence in the results.

In order to help the reader in the interpretation of results we also provided Mean Differences (MDs) of subgroup analyses concerning single scales. In these analyses, we included all studies reporting results of the scale of interest, we did not follow the priority order reported above.

RevMan 5 was the chosen tool to perform the meta-analysis (57).

It is known that the amount of adipose tissue of the patients impacts on the pharmacokinetic properties of antipsychotics and thus on their therapeutic effect (58–61).

We then explored the influence of the Body Mass Index (BMI) at baseline on the treatment effect (SMD) of our main meta-analysis to evaluate if a better response is more related to a better response to the psychiatric treatment than the efficacy of metformin in the regulation of psychiatric and cognitive symptoms. A random-effects meta-regression model with Knapp-Hartung method was performed. Data are provided by a regression bubble plot. The [meta] R package was used to perform meta-regression (62).

As specified in our protocol, we planned to check for eventual useful analyses that were not previously considered. We therefore decided to perform a sub-group analysis of studies reporting data at 12 and 24 weeks.

Since six studies reported an additional dietary and physical exercise control to patients, we also investigated the role of Lifestyle interventions in changes of scales with a sensitivity analysis.

We interpreted I² estimate greater than or equal to 50% together with a statistically significant Chi² statistic as evidence of substantial heterogeneity. We also visually inspected graphs to investigate the possibility of statistical heterogeneity and discussed it in the proper section.

We clearly state down below the deviations from the original protocol registered in PROSPERO:

The study selection and screening process is presented in the PRISMA 2020 flowchart (Figure 1). The electronic search identified 9,605 records from literature databases and 743 trials in study registers. After duplicates removal, 9,455 records were screened. Thirty-eight records were retrieved by manual search in the reference lists of relevant reviews and included studies for full-text analysis. Nine-teen studies eventually met our eligibility criteria and were included in the review: 12 were eligible to perform meta-analysis, 2 reported only qualitative data, and 5 were only present in trial registers without any result. These latter trials will be described in a separate section.

Full description of trials is available in Table 1 and Supplementary Material S1. All studies were designed as comparison between metformin and placebo, except two: Mondal et al. (64) that added one arm treated with topiramate, and Wu et al. (65) that explored the influence of lifestyle interventions. The duration of the included studies was between 12 weeks and 36 weeks. The prescribed dosage of metformin varied from 500 mg/die to 2000 mg/die. All studies included adult patients who were less than 80 years, diagnosed of SCZ (DSM-IV) and under stable treatment with antipsychotics. Four studies (66–69) included patients diagnosed with bipolar disorder and other psychiatric disorders (DSM-IV), however in a negligible percentage: as an example, Agarwal et al. 2 patients out of 30; Baptista et al. 4 patients out of 80. In 5 trials (66, 70–73) only overweight/obese patients were included, and 6 trials reported diabetes or prediabetes among the exclusion criteria (66, 69, 70, 72, 74, 75). All studies, except Mondal et al. (64), reported any other chronic disease (such as thyroid, liver or renal dysfunction, cardiovascular disease) and pregnancy among the exclusion criteria. The mean age at diagnosis was between 21 and 26 years. The disease duration varied from months (first episode) to more than two decades. Clozapine, olanzapine, aripiprazole, and risperidone were the most used antipsychotics. In 3 trials the concomitant use of mood stabilizers, benzodiazepines and antidepressants was permitted (65, 69, 75). Lifestyle intervention involving diet and physical exercise was provided in 6 trials (65, 66, 68, 72, 76, 77). All studies included validated psychometric scales for the clinical assessment; PANSS (65, 70, 72–74, 77) and BPRS (66–69, 71, 76) were the most used.

The following numerical results must be read as “metformin compared to placebo.”

In the analysis of 12 studies (65–74, 76, 77), metformin resulted in a favorable position against placebo (Figure 2), even if not statistically significant [SMD (95% CI) = −0.10 (−0.26; 0.06), OR (95% CI) = 0.8 (−1.4; 3.1)]. No significant differences were seen when studies with missing SDs or those with lifestyle intervention were excluded [SMD (95% CI) = −0.09(−0.27; 0.09), OR (95% CI) = 0.9 (−1.5;3.16) and SMD (95% CI) = −0.02(−0.22; 0.19), OR (95%CI) =1.0 (−1.4; 3.3), respectively].

Sub-group analysis examining studies with same duration of follow-up found non-significant results at 12 weeks [SMD (95%CI) = −0.01(−0.22; 0.20), OR (95% CI) = 1.0 (−1.4; 3.4)] (Figure 3); however, a subsequent improvement at 24 weeks [SMD (95% CI) = −0.40 (−0.82; 0.01), OR (95% CI) = 0.5 (−2.4; 3.4)] (Figure 4). No significant differences were seen when studies with missing SDs or those with lifestyle intervention were excluded.

Forest plots of single-scale analyses of PANSS, BPRS and GAF are available in Supplementary Material (Figure S1-S3 respectively). Better performances were detected by BACS-composite t-score [MD (95%CI) = 1.26 (−0.42; 2.94)], result from one study (66) and PANSS [MD (95% CI) = −2.26 (−5.90; 1.39)], result from 5 studies (65, 70, 72–74), compared to BPRS [MD (95% CI) = −0.57 (−2.56; 1.41)], result from 6 studies (66–69, 71, 76). One study (66) reported changes in BACS-verbal memory in favour of placebo [MD (95% CI) = −16.03 (−23.65; 8.42)]; on the other hand, another study (77) described non-significant results related to SANS [MD (95% CI) = −0.05 (−1.38; 1.28)] and SAPS [MD (95%C I) = 0.09 (−0.67; 0.85)] and similar results were described by Mondal et al. (64) Three studies (66, 69, 72) reported a not significant improvement in GAF [MD (95% CI) = 0.35 (−2.51; 3.21)] and PHQ9 [MD (95% CI) = −2.50 (−1.70; 2.07)], only one study (69).

Only one study (64) compared metformin to another drug used to control the increase of weight in schizophraenic patients. The Authors did not report any quantitative result, they only state that no differences were found in SAPS and SANS scales among groups after 24 weeks.

A non-significant influence of BMI at baseline in the treatment response [β (95%CI) = −0.0320 (−0.0982;0.0343), I² = 45.05%, R² = 4.60%, test of moderators: F = 1.1279; p = 0.3110] was found (Figure 5). Same results were reported by 5 single studies (66, 67, 75–77).

The general adverse events that were reported by authors were related to gastrointestinal discomfort, xerostomia, and extrapyramidal syndrome. Several psychiatric adverse events were also described, particularly some symptoms attributable to a relapse of SCZ (psychotic relapse/exacerbation, unstable/worsening of illness), others to mood alteration (depression, suicidality, irritated/bad mood), and finally some unspecific symptoms such as insomnia and agitation (Supplementary Table S2).

Five trials were registered on clinicaltrial.gov (78). NCT01654640 was terminated because they were not able to recruit enough patients; in NCT02140788, the Principal Investigator left the Institution, and the trial was interrupted. NCT03271866, reported as “unknown status,” focuses on the effect of metformin on cognitive impairment. NCT03708549 is a phase 4 trial that is still recruiting; the aim of the study is to compare berberine and metformin and the evaluation of the PANSS is among the secondary outcomes. NCT04865835 is a phase I trial that has been completed and it is likely under review; however, the aim of this study is to evaluate in pharmacokinetic of a novel substance compared to metformin, which is not specifically our outcome of interest.

Risk of Bias of the included studies is shown in Figure 6. In general, most of the studies reported high risk of bias (8/14). However, the randomisation process was favorably assessed in all studies and the “Deviations from intended interventions” domain was the one that highly influenced the general results because of the Per Protocol analysis used in 6 studies (64, 67, 68, 71, 72, 76). Three studies were considered with low-risk of bias in all domains (65, 70, 74), and only one study was at real high-risk of bias, since none of the fields were assessed without any concern (64).

Chronic treatment with SGAs is essential in the control of psychotic symptoms and the prevention of relapses in SCZ. For this reason, it is widely adopted in clinical settings, even though it can increase the risk of MetS and negatively impact cognitive performance thus worsening the therapeutic compliance, already impaired by the pathology itself. Therefore, the identification of a treatment that can contrast dysmetabolism and cognitive impairment in psychiatric patients would have a high impact in psychiatric clinical practice. Not only has metformin previously shown to be effective in reducing MetS, but it is also considered a low-cost drug, with a well-known safety profile. Our primary aim was to verify the hypothesis, previously emerged from several preclinical and clinical studies, that metformin may exert pro-cognitive effects also in psychotic patients, with or without DM2. This meta-analysis, in addition to its clinical relevance, represents an original perspective in the current literature background.

The first obstacle in investigating our primary objective was that only one study (66) used a specific assessment instrument for cognitive function, the BACS. This is the most widely adopted and validated scale that assesses cognition’s domains most impaired and correlated with outcome of SCZ (55, 92). Unfortunately, it is still underused in clinical practice, while the clinical course and functioning of SCZ are usually assessed by several validated psychometric scales, the main ones being BPRS, PANSS, CGI, GAF. Most of these latter scales contain specific items concerning the patient’s cognitive asset. Thus, since partial scores of these items were not available in the analyzed studies, we applied the priority order described above, that is an original method in order not to neglect valuable information for our primary aim. However, further studies with appropriate scales are warranted.

Another limit of our analysis is the relatively short period of investigation (only one study up to 36 weeks) while neuroprotective effects of metformin observed in previous RCT seem to emerge after long-term use (8 and 6 years) (42). Only 3 RCTs were assessed with low risk of bias, and we could not perform any sensitive analysis excluding those with high risk. However, considering that our aim was defining changes in measurements, the most important domains for our results were the quality of the randomisation process (domain 1) and the measurement of the outcome (domain 4), which were both considered at low risk of bias in the 64% of the included studies.