Electronic databases: including PubMed (ncbi.nlm.nih.gov/pubmed), EMBASE (embase.com) and Web of Science (webofknowledge.com), were used to comprehensively search for all relevant studies published in English from the earliest record to December 2014. The following keywords were used: ‘cartilage defect’, ‘cartilage repair’, ‘osteoarthritis’, ‘knee osteoarthritis’, ‘stem cells’, ‘mesenchymal stem cells’ (MSCs), ‘bone marrow concentrate’, ‘adipose-derived mesenchymal stem cells’ (ADMSCs), ‘synovial-derived mesenchymal stem cells’ and ‘peripheral blood-derived mesenchymal stem cells’, as medical subject headings or text words. In addition, Cochrane Systematic Reviews (cochrane.org/evidence) and ClinicalTrials.gov were manually searched for additional references. Articles were considered eligible if they met the following criteria: i) Patients were ≥18 years-old and had KOA symptom or diagnosed with KOA by clinical and imaging examination; ii) MSCs administered to at least one treatment group; iii) ≥3-month follow-up; iv) ≥1 valid outcome measurement before and after the administration of MSCs, such as the visual analogue scale (VAS), International Knee Documentation Committee (IKDC) Subjective Knee Form, Lysholm scale, and Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC); and v) outcomes were presented as continuous data [mean ± standard deviation (SD)]. Studies that lacked an intervention plan or pain and functional measurements were excluded.

Two independent reviewers searched the electronic databases and evaluated the eligibility of the searched articles and subsequently extracted data using a standardized form, including data on the study type, number of patients enrolled, patient characteristics, disease duration, dosage of MSCs, outcome measurements, follow-up time and adverse events. If additional data was necessary, the authors were contacted for further information. The Jadad scoring system was used to assess the methodological quality of the randomized controlled trials (RCTs) (21). The quality of the included RCTs ranged from 0–5 points, with a score of <3 indicating a low-quality study. The Newcastle-Ottawa Scale (NOS) was used to assess the quality of other studies according to selection, comparability, exposure, and outcome, including single-arm prospective and quasi-experimental studies (22). NOS was scored out of 9 points, with total scores <4 points defined as low quality. Discrepancies between the two independent evaluations of potential articles were resolved by discussion and consensus.

Data were extracted from four time points at or closest to the 3rd, 6th, 12th and 24th months after MSCs treatment. Effect size (ES) was calculated from knee joint pain and functional changes and the results were compared with the pretreatment baseline or between different treatment arms. VAS was extracted from the included articles. If >1 functional measurement was included in an article, only one functional scale in line with the order of IKDC, Lysholm and WOMAC was chosen. As multiple treatment groups wew included in some articles, each group was selected as a separate status set to analysis. Mean ± SD between the pretreatment baseline condition and functional scores after treatment was used to evaluate the effectiveness of MSCs therapy. Positive ES values demonstrated a pain or functional improvement, and vice versa. For studies in which the measurement score and SD was deficient, the value was calculated from the P-value of the corresponding hypothesis test. If the measurement scores and SD could not be extracted in some articles, a correlation of 0.5 was used to estimate the dispersion. A random effect model was used to pool the ESs with a 95% confidence interval (95%CI) on the basis of heterogeneity. A positive pooled ES with a 95%CI >0 indicated an advantage of MSCs compared with the pretreatment condition or reference treatments.

Statistical heterogeneity was assessed via the I-square and Cochran's Q tests. A P-value of <0.10 for χ² test or an I-square >50% was indicative of the existence of substantial heterogeneity (21). Subgroup analysis was performed according to variables of the study design, different dosages, arthroscopic debridement (AD), activation agent, as well as the severity of Kellgren-Lawrence (K-L) grades. Sensitivity analysis was performed by excluding some articles with extreme ES values to assess whether the movement resulted in serious changes in the total result. Funnel plots were used to assess the potential publication bias. All analyses were conducted using Review Manager Version 5.2 (The Cochrane Collaboration, Oxford, UK).

A total of 152 studies were initially searched, of which 117 were removed after title and abstract screening. Of the 35 citations, 18 clinical studies which met the inclusion criteria were identified for eligibility (Fig. 1); five case studies (17,22–26) were excluded and nine studies (24,27–34) were removed due to difficulties in extracting the outcome measurements. Four systematic reviews (35–38) were also excluded. An assessment of the remaining 18 studies revealed that 10 used a single-arm prospective design (18–20,39–45), four used quasi-experimental trials (46–49) and four used RCT (50–53) (Table I). A total of 565 participants (226 males and 339 females) were included from the 18 studies. The duration from the onset of knee pain to registration in each study was 3 months to ≥7 years. The follow-up period was 3–24 months. The majority of studies recruited patients with KOA with a severity grade of 1–4 on the K-L scale. K-L grade s 1–2, and grades 3–4 were defined as early OA and advanced OA, respectively (Table II).

Compared with the pretreatment condition, a pooled ES of 0.80 (95%CI, 0.42–1.17) was determined at 3 months, 1.72 (95%CI, 1.13–2.31) at 6 months, 2.03 (95%CI, 1.30–2.76) at 12 months (Fig.2), and 1.81 (95%CI, 1.62–2.00) at 24 months (Fig. 3), which all favored the status after MSCs treatment. Following the exclusion of an outlier with an extremely high ES, the beneficial effects from MSCs treatment remained, with an ES of 0.77 (95%CI, 0.41–1.13) at 3 months, 1.49 (95%CI, 0.93–2.04) at 6 months, 1.63 (95%CI, 0.99–2.27) at 12 months, and 1.74 (95%CI, 1.55–1.93) at 24 months. A significant superiority of MSCs intervention was demonstrated by a high summed ES at 12 and 24 months without an overlap of the 95%CI of ES at 3 months, which indicated that the treatment effect of MSCs on KOA patients improved significantly over time. However, after excluding the data from quai-experimental and single-arm prospective studies and only using the data from RCTs, the treatment of MSCs did not demonstrate superiority. Relative to the baseline, patients improved in the pain and functional scale scores at all time points.

Participants receiving MSC treatment were stratified according to the study design, administration dosage, AD, activation agents and K-L grades. Point estimates of the pooled ES in the single-arm prospective studies and quasi-experimental trials were higher than those in the RCTs, and an uncertainty in the treatment effectiveness emerged regarding participants in the RCTs at 6, 12 and 24 months, since the 95%CI of the summed ES crossed the value of 0. Stratified analysis failed to demonstrate a dose-responsiveness association in the MSC numbers. However, the treatment effectiveness in the MSC groups with AD or activation agents was superior to the MSC groups without AD and activation agents, particularly at 12 months in the activation agents group (ES, 3.13; 95%CI, 1.55–4.71) compared with the group without activation agents (ES, 0.67; 95% CI, 0.01–1.34). And the early OA group exhibited a higher ES point estimate at all time points than the advanced OA group (Table III).

Seven of the 18 trials reported adverse events after MSC treatment, in which the predominant symptoms were local swelling and transient regional pain. All of the adverse events reported by patients were self-limited or were remedied with therapeutic measures. None of the patients included in the present study were diagnosed with cancer that was associated with MSC therapy. Asymmetry was observed in the funnel plots based on the ESs of changes in the pain and functional scales from baseline (Fig. 4).