This systematic review and meta-analysis study was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-analysis (PRISMA) guidelines (26), and the protocol was registered on Open Science Framework (https://doi.org/10.17605/OSF.IO/5BT6E).

The inclusion criteria were as follows: randomized controlled trials (RCTs) assessing the administration of MSCs for treating OA based on the American College of Rheumatology criteria (27) and a Kellgren-Lawrence grade of at least 1 (28); the intervention group received MSCs as monotherapy, with no restrictions based on dosage, route of administration or time of MSC application; the control group received a blank treatment or isopycnic placebo; the outcome indicators were the visual analogue scale (VAS) (29) or the total Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC) (30); safety was assessed as adverse events (AEs); and studies published in English.

The following studies were excluded: secondary analyses, including pooled analyses; reviews or conference abstracts; studies of MSCs combined with other surgeries, such as arthroscopic debridement and high tibial osteotomy (HTO); studies with a follow-up duration shorter than 6 months; and abstracts (insufficient data).

We systematically searched the Cochrane Central Register of Controlled Trials, PubMed, Embase, Scopus and Web of Science databases from inception to May 6, 2024.

We used a combination of relevant terms, including “mesenchymal stem cells”, “osteoarthrosis”, “placebo”, and “randomized controlled trial” ( Supplementary Table 1 ). In addition, the reference lists of both the included studies and relevant reviews were manually searched to identify additional eligible studies. We included only articles published in English.

The study selection process was performed independently by two review authors (Z.A. Qu and X.Y. Tian). We obtained the full texts of the studies to determine their eligibility. If there were multiple reports that described the same trial, only the most recent or complete study was included.

Relevant data from the selected studies were independently extracted in accordance with the inclusion criteria by two authors (X.Y. Tian and Z.A. Qu). A third author (B.C. Zhang) was consulted to resolve any disagreements regarding study selection and data extraction. If the means and standard deviations were not reported in the text of the articles, we extracted the values from the diagrams and tables as needed.

The following data were extracted from the included studies: author’s name, year of publication, patient information (including sex, mean age), MSC type, regimen in the intervention and control groups, and follow-up duration.

The primary outcome measures of interest were the mean change in the WOMAC score and the VAS score between baseline and the endpoint. The secondary outcome measures were the incidence rates of treatment-related AEs, such as arthralgia and swelling.

Two review authors (X.Y. Tian and Z.A. Qu) independently assessed the risk of bias for each study using the Cochrane risk of bias assessment tool (31). The tool assesses seven specific domains: sequence generation, allocation concealment, blinding of participants and personnel, blinding of outcome assessment, incomplete outcome data, selective outcome reporting and ‘other sources of bias’. Each domain was scored as having a low risk of bias, high risk of bias or unclear risk of bias. Disagreements were resolved by discussion and consensus.

A conventional meta-analysis was conducted to compare MSCs with placebo using Review Manager 5.3 (Cochrane Collaboration, Oxford, UK). Dichotomous data were examined using risk ratios (RRs) and 95% confidence intervals (CIs) based on the number of events in the control and intervention groups of each study. Continuous data were analyzed using the mean differences (MDs) and 95% CIs between the MSC and control groups. The WOMAC and VAS scores were converted to a common scale from 0 (no pain or disability) to 100 (worst possible pain or disability) before meta-analysis. The heterogeneity of the effect size across the studies was tested using the chi-square test (p < 0.1 was considered heterogeneous) and I² statistic (I² > 50% was considered heterogeneous). If significant heterogeneity existed between studies, a random effects model was used; otherwise, a fixed effects model was used. Subgroup analysis was performed based on the dose and type of MSCs, mean MSC counts per injection and autologous or allogenic MSCs. A sensitivity analysis was performed to examine the reliability of the results. When a control group served multiple experimental groups, the number of participants in the control group was divided by the number of experimental groups. The overall effect was tested using a Z score with the significance set at p < 0.05. We used funnel plots to assess publication bias if more than 10 included trials examined a particular outcome.

The PRISMA flow chart of study selection is shown in Figure 1 . We initially identified 1,612 studies from database searches. After removing duplicates, 846 studies remained. We then excluded 760 studies based on the titles and abstracts. We assessed the full texts of the remaining 86 papers and excluded 68 studies (details in Supplementary Table 2 ). Ultimately, we included 18 studies (14–16, 23–25, 32–43) in the systematic review and 16 studies (14–16, 23–25, 32–34, 36, 37, 39–43) in the meta-analysis.

The characteristics of the included studies are shown in Table 1 . This systematic review included 18 eligible studies with a total sample size of 1,174 participants. A total of 633 patients were included in the intervention group. Among the cell sources, 3 studies examined allogeneic AD-MSCs (14, 25, 39), 4 studies examined autologous AD-MSCs (24, 32, 34, 41), 3 studies examined allogeneic BM-MSCs (16, 23, 40), 5 studies examined autologous BM-MSCs (15, 33, 35–37), and 3 studies examined allogeneic UC-MSCs (38, 42, 43). In the included studies, the average number of MSCs per injection ranged from 1× 10⁶ to 150 × 10⁶. Fourteen studies compared the effects of a single dose of MSCs, with seven studies utilizing a dosage range of 50–100 × 10⁶ and an additional seven studies employing a dosage less than 50 × 10⁶. The remaining four studies examined varying doses of MSCs.

One study had a follow-up duration of 24 months (14), 13 studies had a follow-up duration of 12 months (15, 16, 23, 25, 32, 35–37, 39–43), and 4 studies had a follow-up duration of 6 months (24, 33, 34, 38). Seventeen studies were registered, and one study was not registered (35). Most of the included participants had Kellgren-Lawrence grades of 2 (28.8%) and 3 (55.6%).

The risk of bias assessment for all of the included studies is shown in Figure 2 . For 3 studies, all domains were judged as having a low risk of bias (23, 38, 41). Randomized sequence generation was not implemented adequately in 2 studies (34, 36); however, all of the included studies were RCTs. Allocation concealment was not implemented adequately in 7 studies (14, 15, 24, 25, 32, 39, 40). Twelve studies successfully reported blinding of participants (14, 23, 24, 33–35, 38–43), and the study personnel were at low risk of performance bias. Twelve studies reported the blinding of outcome assessors, thus leading to a low risk of detection bias (23, 24, 32–41). Five studies did not present loss to follow-up data and thus had an unclear risk of attrition bias (16, 33, 35, 37, 40). All studies exhibited a low risk of reporting bias.

Eleven included studies evaluating efficacy used a reduction in the WOMAC score as the primary or secondary outcome (14, 15, 23–25, 32, 33, 37, 40–42). First, we compared the mean change in the 6-month WOMAC between the MSC group and the placebo group. The reduction in the MSC group was significantly greater than that in the placebo group (n = 749; MD = -11.75; 95% CI = -16.26–7.25; p < 0.00001; I² = 61%) ( Figure 3A ). Similar results were obtained for the 12-month WOMAC (n = 454; MD = -15.94; 95% CI = -23.79–8.10; p < 0.00001; I² = 85%) ( Figure 3B ). Sensitivity analysis was conducted to examine the robustness of these results, and the efficacy was not substantially altered after removing any one study. The visual cues in the funnel plots indicated no conclusive evidence of publication bias ( Supplementary Figure 1 ).

Thirteen studies evaluated the analgesic effect of MSCs using the VAS score (14–16, 23–25, 33, 37, 39–43). The improvement in the VAS score significantly differed between the MSC group and the placebo group at 6 months (n = 891; MD = -11.94; 95% CI = -18.50–5.37; p = 0.0004; I² = 67%) ( Figure 4A ). Similar results were obtained for the 12-month VAS score (n = 742; MD = -14.25; 95% CI = -23.14–5.35; p = 0.002; I² = 83%) ( Figure 4B ). Sensitivity analysis was conducted to examine the robustness of these results, and the efficacy was not substantially altered after removing any one study. The visual cues in the funnel plots indicated no conclusive evidence of publication bias ( Supplementary Figure 2 ).

Previous studies have suggested that transient arthralgia and swelling can occur after MSC injection, so we evaluated the safety of MSCs by examining the incidence rates of treatment-related arthralgia and swelling. The risks of arthralgia (n = 650; RR = 1.22; 95% CI = 0.89–1.67; p = 0.21; I² = 0%) ( Figure 5A ) and joint swelling (n = 620; RR = 1.43; 95% CI = 0.69–2.94; p = 0.33; I² = 48%) ( Figure 5B ) were greater in the MSC group, but there was no significant difference compared with the placebo group. Sensitivity analysis was conducted to examine the robustness of these results, and the findings were not substantially altered after removing any one study. The visual cues in the funnel plots indicated no conclusive evidence of publication bias ( Supplementary Figure 3 ).

Based on the results described above, we analyzed the WOMAC index and the VAS score across different subgroups stratified according to the following variables: MSC type (BM, AD or UC), MSC source (autologous or allogeneic), and mean MSC count per injection (low: 0–49×10⁶ cells or high: greater than 50×10⁶ cells) ( Tables 2 , 3 ).

In the subgroup analysis of the WOMAC, 11 studies were synthesized at the 6-month follow-up, and 9 studies were synthesized at the 12-month follow-up. No significant difference in the 6-month WOMAC score was observed in the subgroup analyses when the data were stratified by MSC type and mean MSC count per injection. Overall, allogeneic MSCs may have a better curative effect than autologous MSCs (MD = -14.94, 95% CI = -21.17–8.71 vs. MD = -7.20, 95% CI = -11.39–3.00; p = 0.04, I² = 75.5%). Subgroup analyses did not reveal any significant differences in the 12-month WOMAC index score between subgroups. Additionally, among the BM-MSC group, the 12-month WOMAC index score was not significantly different between the MSC group and the control group (p = 0.07).

According to the subgroup analysis of the VAS score, 12 studies reported a 6-month follow-up, and 11 studies reported a 12-month follow-up. No significant difference in the 6-month VAS score was observed in the subgroup analyses when the data were stratified by MSC type, MSC source or mean MSC count per injection. Additionally, among the UC-MSC group, BM-MSC group and high cell count group, the 6-month VAS score did not significantly differ between the MSC group and the control group (UC-MSC group: p = 0.25; BM-MSC group: p = 0.14; high cell count group: p = 0.20). Similar results were obtained for the 12-month VAS score. No significant difference in the 12-month VAS score was observed in the subgroup analyses. The 12-month VAS score of the UC-MSC group, BM-MSC group and high cell count group did not differ from that of the placebo group (UC-MSC group: p = 0.25; BM-MSC group: p = 0.09; high cell count group: p = 0.44).