Results

Study characteristics

Figure 1 provides a summary of the article selection procedure. Initially, 1,330 records were obtained through database searches. After removing duplicates (n = 238) and records deemed ineligible by automated screening (n = 50), 1,042 unique records were reviewed at the title and abstract level. Following the exclusion of 935 records, 107 articles underwent full-text review. A final cohort of 38 preclinical in vivo studies was included in the quantitative synthesis. The key characteristics of the included studies are summarized in Table 1. The 38 studies encompassed both human-derived (28 studies) [12-39] and animal-derived (10 studies) [40-49] MSC-EVs evaluated across diverse preclinical OA models. Of these, 26 human-derived studies and 8 animal-derived studies provided extractable quantitative data for the primary OARSI score meta-analysis. The remaining two animal-derived studies [48,49] met the inclusion criteria and are described in Table 1, but were excluded from the quantitative synthesis because OARSI histological score values (mean ± SD) were not reported in numerical form in the text, tables, or supplementary data of the original publications, precluding mean difference calculation. Rodent models (C57BL/6 mice and Sprague-Dawley rats) predominated, with OA primarily induced via surgical techniques (e.g. destabilization of the medial meniscus (DMM), anterior cruciate ligament transection (ACLT)) or chemical induction (e.g. monosodium iodoacetate [MIA]) [12-49]. EV cellular origins were highly heterogeneous, including bone marrow, adipose tissue, synovial membrane, umbilical cord, and induced pluripotent stem cells.

Figure 1.: PRISMA flow diagram of the literature search and selection of studies for meta-analysis

Study ID | EV source | Tissue origin | EV source species | Experimental subject | Age, week | Sex/n | Weight, g | OA model | Isolation method | Characterizationmethod | Ref.
SD1 | iMSC-Exo | Induced pluripotent stem cells | Human | C57B/L10 mice | 6 | Female/35 | NR | CIOA | Ultracentrifugation | RPSA, TEM, WB | [12]
SMMSCs-Exo | Synovial | Human | C57B/L10 mice | 6 | Female/36 | NR | CIOA | Ultracentrifugation | RPSA, TEM, WB
SD2 | SMMSCs-Exo | Synovial | Human | SD rats | ~ 12 | Male/30 | 300 to 350 | Complete transection of the medial collateral ligament, ACLT and DMM | NR | NR | [13]
SD3 | ESCMSCs-Exo | Embryonic | Human | C57BL/6 J | 8 | NR/32 | NR | DMM | Ultracentrifugation | NR | [14]
SD4 | IPFPMSC-Exo | infrapatellar fat pad | Human | C57BL/6 mice | 9 | Male/39 | NR | DMM | ExoQuick™ reagent kit and ultrafiltration | TEM, NTA, WB | [15]
SD5 | ADSC-EVs | Adipose | Human | SD rats | 7 | Male/50 | 200 to 250 | MIA | TFF | TEM, NTA, WB, FC, DLS | [16]
Adipose | Human | C57BL/6 mice | 9 | Male/NR | 20 to 25 | DMM | TFF | TEM, NTA, WB, FC, DLS
SD6 | BMSCs-Exo | Bone Marrow | Human | C57BL/6 J mice | 6 | Male/23 | NR | CIOA | Total exosome isolation kit |  | [17]
SD7 | SMMSCs-Exo | Synovial | Human | SD rats | NR | Male/40 | 200 to 220 | DMM | Ultracentrifugation | TEM, NTA, WB | [18]
SD8 | SMMSCs-Exo | Synovial | Human | C57 mice | 8 | Male/20 | 25 to 30 | Complete transection of the medial collateral ligament, ACLT and DMM | Ultracentrifugation | TEM, NTA | [19]
SD9 | UCMSCs-Exo | Umbilical cord | Human | SD rats | 8 | Male/24 | NR | ACLT | Commercial kit | TEM,FC | [20]
SD10 | ADSCs-Exo | Adipose | Human | BALB/c mice | 3 | Female/35 | NR | ciprofloxacin induced OA | Ultracentrifugation | TEM, DLS, FC | [21]
BMSCs-Exo | Bone Marrow | Human | BALB/c mice | 3 | Female/35 | NR | ciprofloxacin induced OA | Exocib exosome extraction kit | TEM, DLS, FC
SD11 | BMSCs-Exo | Bone Marrow | Human | SD rats | NR | Male/20 | 358 ± 5 | ACL+MM | Gradient centrifugation | TEM, FC, WB | [22]
SD12 | ADSC-EVs | Adipose | Human | ICR (CD-1) mice | 8 | Female/30 | NR | Bilaterally OVX | TFF | TEM, FC | [23]
SD13 | SMMSCs-Exo | Synovial | Human | C57BL/6 J mice | 10 | Male/NR | NR | DMM | Ultracentrifugation | NTA, TEM, WB | [24]
SD14 | UCMSCs-Exo | Umbilical cord | Human | SD rats | NR | NR/18 | NR | Surgically induced cartilage defect model | NR | TEM, NTA, FC | [25]
SD15 | UCMSCs-Exo | Umbilical cord | Human | SD rats | 12 | Female/40 | 250 ± 20 | ACLT and DMM | Ultracentrifugation |  | [26]
SD16 | ADSCs-Exo | Adipose | Human | SD rats | NR | NR/15 | 230 to 280 | MIA-OA | Ultracentrifugation |  | [27]
SD17 | iPSCs-Exo | PBMS | Human | NZ Rabbits | NR | Female/9 | 400 | ACLT | Sequential ultracentrifugation | TEM, WB, DLS | [28]
SD18 | UCMSCs-Exo | Umbilical cord | Human | C57BL/6 mice | 8 | Male/NR | NR | DMM | Differential centrifugation | TEM, NTA, WB | [29]
SD19 | WJMSC-Exo | Umbilical cord | Human | SD rats | 8 | Male/24 | NR | ACLT | Sequential ultracentrifugation | TEM, AFM, SEM, FC | [30]
SD20 | ADSC-Exo | Subcutaneous fat (SC) | Human | SD žrats | 6 | Male/NR | 180 to 200 | DMM+ACLT | Ultracentrifugation | TEM, NTA | [31]
ADSC-Exo | Subcutaneous fat (SC) | Human | C57BL/6 J mice | 7 | Male/NR | 18 to 22 | DMM | Ultracentrifugation | TEM, NTA
SD21 | ADSC-Exo | Subcutaneous adipose | Human | SD rats | 8 | Male/30 | NR | ACLT | Ultracentrifugation | TEM, NTA | [32]
SD22 | UCMSCs-Exo | Umbilical cord | Human | SD rats | NR | Male | 350g | ACL rupture-induced OA | Gradient centrifugation | TEM, NTA | [33]
SD23 | DPSC-Exo | Dental pulp stem cell | Human | C57BL/6 mice | 7 | Male/15 | 20-25g | MIA | Sequential ultracentrifugation | TEM, NTA, WB | [34]
SD24 | pExo | Placenta | Human | SD rats | 10 to 12 | Male/NR | NR | MCLT+MMT | Sequential ultracentrifugation | NTA, WB | [35]
SD25 | UCMSCs-Exo | Umbilical cord | Human | SD rats | 6 to 8 | Male/24 | 200 ± 20 | MIA | NR |  | [36]
SD26 | UCMSCs-Exo | Umbilical cord | Human | SD rats | 8 | Male/NR | 300 to 350 | CIOA | TFF | TEM, NTA,FC | [37]
SD27 | UCMSCs-Exo | Umbilical cord | Human | C57BL/6 J mice | 8 to 12 | Male, Female/ NR | NR | CIOA | Ultracentrifugation | NTA, FC | [38]
SD28 | UCMSCs-Exo | Umbilical cord | Human | SD rats | 12 | Male/NR | 300 to 350 | ACLT + pMMx | Ultracentrifugation | TEM, ZS, WB | [39]
SD1 | BMSCs-Exo | Bone marrow | C57BL/6 mice | C57BL/6 mice | 3 days | NR/45 | NR | CIOA | Ultracentrifugation | NTA, TEM, DLS, FC | [40]
SD2 | BMSCs-Exo | Bone marrow | C57BL/6 mice | C57BL/6 mice | NR | NR/60 | 25 to 30 | Surgically induce instability in the lumbar spine | Ultracentrifugation | TEM, BCA, WB | [41]
SD3 | BMSCs-Exo | Bone marrow | C57BL/6 mice | SD rats | Adult | Male/36 | 200 to 250 | ACLT+DMM | Ultracentrifugation | TEM, NTA, WB | [42]
SD4 | BMSCs-Exo | Bone marrow | SD Rats | SD rats | 6 | Male/80 | 190 ± 10 | ACL+MCL | Ultracentrifugation | WB | [43]
SD5 | BMSCs-Exo | Bone marrow | SD Rats | SD rats | 10 | Male/24 | NR | MIA-OA | Ultracentrifugation | TEM, NTA, WB | [44]
SD6 | BMSCs-Exo | Bone marrow | SD Rats | SD rats | NR | Male/24 | 200 to 220 | MIA-OA | ExoQuick-TC™ | TEM, NTA | [45]
SD7 | BMSCs-Exo | Bone marrow | Rabbit | NZ rabbits | NR | Male/20 | 2500 ± 500 | ACL+MM | NR | TEM, NTA, FC | [46]
SD8 | BMSCs-Exo | Bone marrow | C57BL/6 mice | SD rats | NR | NR/40 | NR | MCL+ACL+PCL | ExoQuick Extraction Kit | TEM, NTA, WB | [47]
SD9 | ADSCs-Exo | Adipose | C57BL/6 mice | C57BL/6 mice | 8 | Male/102 | 25 to 30 | LFJ OA | NR | TEM, NTA, WB | [48]
SD10 | BMSCs-Exo | Bone marrow | SD Rats | SD rats | 10 | Male/24 | 220 | ACL+MM | Ultracentrifugation | TEM, NTA, WB | [49]

Quality of included studies

To contextualize the heterogeneity observed across the included studies and to identify systematic gaps in methodological transparency, we performed a structured reporting-quality assessment of all 38 studies against the MISEV 2023 guidelines. Each study was scored across seven domains derived directly from the comprehensive methodological matrix (Supplementary material, Table S2): isolation method, characterization method (MISEV triad of morphology, particle metrics, and protein markers), size distribution, surface marker analysis, functional testing, procedural reporting clarity (centrifugation force, time, temperature, and equipment), and statistical methods. No study achieved complete adherence to MISEV 2023 reporting standards. Substantial adherence was observed in 7 of 38 studies (18.4 %), Partial adherence in 20 of 38 studies (52.6 %), and limited adherence in 11 of 38 studies (28.9 %). Domain-level analysis revealed that the strongest reporting was in surface marker characterization (32/38, 84.2 % adequate) and the MISEV characterization triad (31/38, 81.6 % adequate), reflecting widespread adoption of the basic transmission electron microscopy/nanoparticle tracking analysis/Western blot characterization framework. In contrast, the most pronounced reporting deficits were observed in numeric size distribution reporting, with only 4 of 38 studies (10.5 %) providing explicit modal diameter or size-range values in the methods or results text rather than in figure panels alone, and in procedural reporting clarity, where only 5 of 38 studies (13.2 %) reported all three of relative centrifugal force/time, isolation temperature, and ultracentrifuge or rotor model. Statistical methods were fully specified (test, software, and significance threshold) in only 10 of 38 studies (26.3 %), whereas functional validation, combining both in vivo OA model outcomes and in vitro mechanistic assays, was reported in 14 of 38 studies (36.8 %).

Critical gaps were observed in the reporting of pharmacological dosing parameters. Methodological quality was further evaluated using the SYRCLE risk-of-bias tool (Figure 2). Most studies demonstrated a low risk of bias regarding random sequence generation and selective reporting.

Figure 2.: Evaluation of methodological quality of the included studies using the SYRCLE risk of bias tool.

However, blinding of participants/personnel, allocation concealment, and blinded outcome assessments were frequently rated as high or unclear risk, highlighting persistent methodological vulnerabilities in the preclinical EV landscape that may inflate perceived efficacies.

Effect size (primary outcome)

Quantitative synthesis of the OARSI histological score demonstrated robust therapeutic efficacy of MSC-EV administration. For human-derived MSC-EVs, the pooled mean difference (MD) was -3.27 (95% CI: -4.66 to -1.88; p <0.0001), indicating significant structural protection of articular cartilage compared with untreated controls. Heterogeneity remained substantial (_I_2 = 94.64%, _τ_2 = 11.92), reflecting the wide variance in EV sourcing and dosing (Figure 3). Crucially, animal-derived MSC-EVs exerted an even more pronounced protective effect, yielding a pooled MD of -5.58 (95% CI: -7.13 to -4.03; p <0.0001). While statistical heterogeneity was also present in the animal-derived subgroup (_I_2 = 84.19%, _τ_2 = 4.07), the highly significant p-value confirms the potent cross-species biological activity of MSC-EVs in mitigating OA progression (Figure 4).

Figure 3.: Forest plot of preclinical studies [12-22,24-30,32-34,36,38] evaluating human MSC-derived exosomes in OA models using OARSI histological scoring. The random-effects model (REML) yielded a pooled estimate of -3.27 (95% CI: -4.66 to -1.88; p <0.0001), with substantial heterogeneity (I2 = 94.64%, τ2 = 11.92, Q = 265.59, df = 25)

Figure 4.: Forest plot of preclinical studies [38-44,46] evaluating animal-derived MSC-derived exosomes in OA models using OARSI histological scoring. The random-effects model (REML) yielded a pooled estimate of -5.58 (95% CI: -7.13 to -4.03; p <0.0001), with substantial heterogeneity (I2 = 84.19%, τ2 = 4.07, Q = 53.37, df = 7)

Dosage characteristics

A focused evaluation of EV dosing revealed profound heterogeneity, underscoring a major barrier to ADMET standardization (Supplementary material, Table S3). Dose per injection (particles) was established as the primary metric, while protein concentration served as a secondary proxy. As summarized in Tables 2 and 3, particle-based dosing was reported in 14 of 38 studies (37 %), spanning 8×107 to 1010 particles per injection (median ~3×108; IQR: 108 to 109). Protein-based dosing was reported in 13 of 38 studies (34 %), ranging from 0.25 μg to 500 μg per injection. Critical contextual parameters for pharmacokinetics, such as precise injection volumes, dosing intervals, and intra-articular retention times, were omitted in the majority of reports, thereby preventing precise dose-exposure modelling and highlighting an urgent need for standardization of pharmacological reporting in EV therapeutics.

Metric | Median (IQR) | Range | Number of reporting studies
Dose per injection (particles), μg | ~3×108 (108 to 109) | 8×107 to 1010 | 14
Dose per injection (protein concentration), μg/mL | 40 (10 to 100) | 0.25 to 500 | 13

Parameter | Number of reported dose (n / %) | Number of not reported dose (n / %)
Particles per injection | 14 (37) | 24 (63)
Protein per injection | 13 (34) | 25 (66)

Subgroup analysis

To delineate drivers of inter-study variance, a priori subgroup analyses were conducted for both human-derived and animal-derived MSC-EVs, stratified by EV cellular source, OA induction model and experimental subject.

Human-derived MSC-EVs

No significant differences in therapeutic efficacy were observed across tissue origins of the EVs, as indicated by the test for subgroup differences (Q_b(8) = 12.79, p = 0.119). Adipose-derived EVs (ADSC-Exo) demonstrated a statistically significant reduction in OARSI scores (MD = -3.01, 95% CI: -4.29 to -1.73). Similarly, bone marrow-derived EVs (BMSCs-Exo) also yielded a significant pooled effect (MD = -2.79, 95% CI: -4.58 to -1.00). Furthermore, induced MSC-derived EVs (iMSC-Exo) showed a strong reduction in scores (MD = -6.23, 95% CI: -8.35 to -4.11). In contrast, Wharton's jelly-derived EVs (WJMSC-Exo) were associated with a positive mean difference (MD = 2.20, 95% CI: 0.91 to 3.48) (Figure 5). The overall random-effects model across all EV types indicated a significant overall reduction in OARSI scores (MD = -3.27, 95% CI: -4.66 to -1.88) (Figure 5).

Figure 5.: Meta-analysis of EV therapeutic efficacy on OARSI scores subgrouped by tissue origin [12-39]. Forest plot detailing the effect of human-derived extracellular vesicles (EVs) on OARSI scores, stratified by the tissue source of the EVs (e.g. ADSC-Exo, SMMSCs-Exo, UCMSCs-Exo, BMSCs-Exo, WJMSC-Exo, iMSC-Exo). The plot displays MD and 95% CI for individual studies and the pooled random-effects (RE) models for each subgroup. The test for subgroup differences indicates no statistically significant variation in therapeutic efficacy based on EV tissue origin (Q_b(8) = 12.79, p = 0.119). The overall random-effects model across all studies demonstrates a significant reduction in OARSI scores (MD = -3.27, 95% CI: -4.66 to -1.88)

A subgroup analysis on experimental subjects (e.g. C57 mice, SD rats, NZ Rabbits) indicated that the host species did not significantly moderate the therapeutic efficacy of human-derived EVs (Q_b(6) = 9.08, p = 0.169). This suggests that the structural benefits of human MSC-EVs are conserved across various preclinical xenograft models (Figure 6).

Figure 6.: Meta-analysis of EV therapeutic efficacy on OARSI scores subgrouped by experimental subjects [10-37]. Forest plot of a mixed-effects meta-regression assessing the impact of host species and animal models on the therapeutic rescue of EVs, measured via OARSI scores. Subgroups represent the various preclinical xenograft models utilized across studies, including SD rats, multiple strains of mice (e.g. C57BL/6 J, C57BL/6, C57B/L10, BALB/c) and NZ rabbits. The analysis indicates that the host species does not significantly moderate the therapeutic efficacy of the EVs (Q_b(6) = 9.08, p = 0.169)

Subgrouping by OA induction methodology revealed no statistically significant differences in therapeutic efficacy across the various models, as indicated by the test for subgroup differences (Q_b(10) = 6.62, p = 0.761). The MIA model yielded an effect size of MD = -4.33 (95% CI: -6.10 to -2.56). Similarly, Collagenase-Induced OA (CIOA) produced an MD of -5.87 (95% CI: -11.44 to -0.29). Surgical DMM models exhibited high internal variance and a non-significant pooled effect, as the confidence interval crossed zero (MD = -1.29, 95% CI: -2.82 to 0.24) (Figure 7).

Figure 7.: Meta-analysis of EV therapeutic efficacy on OARSI scores subgrouped by OA induction methodology [10-37]. Forest plot evaluating the therapeutic impact of EVs on OARSI scores, stratified by the methodology used to OA in the experimental models. Subgroups include chemically induced models (e.g. MIA, CIOA, ciprofloxacin) and surgical models (e.g. DMM, ACLT, ACL rupture). The test for subgroup differences reveals no statistically significant variance in therapeutic efficacy across the different induction methodologies (Q_b(10) = 6.62, p = 0.761)

Animal-derived MSC-EVs

Consistent with trends in the human subgroup, both adipose- and bone marrow-derived animal MSC-EVs exerted protective effects against cartilage degradation. While bone marrow remains the most frequently investigated source, yielding a highly reproducible significant benefit (MD: -5.21, 95% CI: -6.75 to -3.67), preliminary evidence indicates that adipose-derived EVs may yield an even more potent structural rescue (MD: -8.70, 95% CI: -11.54 to -5.87) (Figure 8).

Figure 8.: Meta-analysis of animal MSC-EV therapeutic efficacy on OARSI scores subgrouped by EV tissue source [38-44,46]. Forest plot detailing the effect of animal-derived extracellular vesicles (EVs) on OARSI scores, stratified by the tissue origin of the EVs: bone marrow-derived (BMSCs-Exo) and adipose-derived (ADSCs-Exo). The plot displays MD) and 95% CI for individual studies and the pooled random-effects (RE) models. The overall random-effects model across all studies demonstrates a significant reduction in OARSI scores (MD: = -5.58, 95% CI: -7.13, -4.03). The test for subgroup differences indicates no statistically significant variation in therapeutic efficacy between the two tissue sources (Q_b(1) = 1.95, p = 0.162)

Therapeutic efficacy varies across induction methods. The chemically induced MIA-OA model yielded a highly significant effect size (MD: -4.49, 95% CI: -5.71 to -3.26), while specific surgical instability models, such as ACL+MM and LFJ OA, demonstrated even more profound improvements (MD: -6.46 and -8.70, respectively) (Figure 9).

Figure 9.: Meta-analysis of animal MSC-EV therapeutic efficacy on OARSI scores subgrouped by OA induction methodology [38-44,46]. Forest plot evaluating the therapeutic impact of animal-derived EVs on OARSI scores across OA induction models. Subgroups include chemically induced models (e.g. MIA-OA, CIOA) and multiple surgical instability models (e.g. ACL+MCL, ACL+MM, ACLT+DMM, Lumbar spine instability, LFJ OA). While the test for subgroup differences yields a statistically significant p-value (Q_b(6) = 53.33, p = 0.000), a notation clarifies that because 6 out of the 7 subgroups consist of only a single study (k = 1), the between-model comparison is descriptive rather than inferential. The overall model demonstrates a pooled MD of -5.58 (95% CI: -7.13, -4.03). Note: With k = 1 in 6 of 7 subgroups, between-model comparison is descriptive rather than inferential

Animal-derived EVs successfully attenuated cartilage degeneration across diverse hosts, including C57BL/6 mice (MD: -7.50, 95% CI: -9.63 to -5.37), SD rats (MD: -3.98, 95% CI: -5.50 to -2.45), and NZ rabbits (MD: -6.46, 95% CI: -8.65 to -4.27). Efficacy was preserved across both small-rodent and larger-animal hosts, reinforcing the robust translational and cross-species potential of MSC-EV therapies (Figure 10).

Figure 10.: Meta-analysis of animal MSC-EV therapeutic efficacy on OARSI scores subgrouped by experimental host species [38-44,46]. Forest plot assessing the impact of different preclinical animal hosts on the therapeutic rescue of EVs, measured via OARSI scores. Subgroups represent the host species utilized across the evaluated studies, which include SD rats, C57BL/6 mice and NZ rabbits. The test for subgroup differences indicates that the choice of experimental host significantly moderates the therapeutic efficacy of the EVs (Q_b(2) = 7.55, p = 0.023). The overall random-effects model reflects a significant therapeutic benefit across all subjects (MD = -5.58, 95% CI: -7.13, -4.03).

Evaluation of publication bias and sensitivity analysis

Potential publication bias and small-study effects were initially evaluated using Egger’s regression test for funnel plot asymmetry. Highly significant asymmetry was detected across both datasets, with human-derived MSC-EVs (z = -7.50, p <0.0001) and animal-derived MSC-EVs (z = -4.56, p <0.0001) demonstrating a departure from symmetry (Figure 11).

Figure 11.: Funnel plots assessing publication bias and small-study effects. (A) Funnel plot of 26 preclinical studies evaluating human-derived MSC-EVs on OARSI scores. Visual asymmetry indicates the presence of small-study effects, formally corroborated by Egger's regression test (z = -7.50, p <0.0001). (B) Funnel plot of 8 studies evaluating animal-derived MSC-EVs. Similar asymmetry is observed, confirming significant small-study effects (z = -4.56, p <0.0001)

To rigorously assess whether this potential publication bias invalidated our primary findings, we performed a nonparametric Duval and Tweedie Trim-and-Fill analysis. For both the animal-derived dataset (k = 8) and the human-derived dataset (k = 26), the model estimated 0 missing studies on the right side. Consequently, the imputed models did not alter the original effect sizes, confirming the robustness of the significant therapeutic benefits observed for both animal-derived (MD = -5.58, 95% CI: -7.13 to -4.03) and human-derived EVs (MD = -3.27, 95% CI: -4.66 to -1.88).

To further test the stability of these findings against unpublished null results, Fail-safe N calculations were conducted. The Rosenthal approach indicated that 2,058 and 813 missing studies with an effect size of zero would be required to nullify the statistical significance of the human and animal datasets, respectively. Similarly, the Rosenberg Fail-safe N corroborated this high tolerance, requiring 906 and 485 missing studies to bring the significance level above. Finally, the Orwin approach demonstrated that 26 and 8 missing studies would be needed to reduce the respective average effect sizes exactly in half. Together, these sensitivity analyses indicate that although funnel plot asymmetry is present, the observed protective effects of MSC-EVs on cartilage integrity remain statistically robust and are highly unlikely to be overturned by unpublished negative data.