Scope of This Review
This page is a narrative review of the published systematic reviews and meta-analyses of intra-articular mesenchymal stromal cell (MSC) therapy for knee osteoarthritis. It is not itself a systematic review, and no pooled analysis was performed for it. Every effect estimate below is attributed to the published synthesis or trial that produced it, quoted at that paper's own strength, with its confidence interval and its authors' own caveats.
That distinction matters more here than usual, because the central feature of this literature is that the published syntheses disagree with each other — and the disagreement follows a pattern worth understanding before reading any single number.
1. Why This Question Is Hard
Knee osteoarthritis is among the leading causes of disability worldwide [1], and in the United States lifetime risk of symptomatic disease approaches one in two by age 85 [2].
Every rung of the existing treatment ladder has real limits. Long-term NSAID use carries vascular and upper gastrointestinal risk [3]. Intra-articular corticosteroid gives short-term relief, but in a two-year randomised trial against saline, triamcinolone produced greater cartilage volume loss with no significant pain benefit at 24 months [4]. Hyaluronic acid shows modest effects in network meta-analysis, in a field where intra-articular placebo performs strongly in its own right [5]. Knee replacement is effective for end-stage disease but is major surgery [6].
That gap is why cell therapy attracts the attention it does. It is also why this literature needs reading carefully rather than enthusiastically, for four reasons.
Comparator choice changes the answer. A trial against hyaluronic acid asks a different question from a trial against saline, and the two produce systematically different results [15].
Contextual effects are unusually large. Needle, aspiration, attention and expectation all contribute. A 2025 meta-analysis dedicated to this question found contextual effects accounted for roughly 63% of the observed treatment response at six months and about 50% at twelve [27].
"MSC" is not one product. Bone marrow, adipose, umbilical cord; autologous and allogeneic; culture-expanded and minimally manipulated; doses spanning two orders of magnitude. Pooling these assumes a common effect that may not exist.
Early risk of bias was high — and that finding is itself informative. The first systematic review of the field found that all five available RCTs reported superior patient-reported outcomes, and that all five were at high risk of bias, yielding level-3 evidence overall. On that basis its authors declined to recommend the treatment [14]. Consistently positive results from consistently weak studies is the signature that better-controlled trials tend to deflate — and, as Section 2 shows, partly did.
2. The Central Finding: The Syntheses Disagree
2.1 Against placebo, the pooled effect is not significant
The most direct test is a meta-analysis of 13 Level-I randomised controlled trials in Arthroscopy. Against placebo it found:
The proportion of patients crossing the minimal clinically important difference [28]also did not differ: RR 0.93 (0.55–1.57) for VAS pain and RR 0.40 (0.13–1.21) for WOMAC total. The authors state the conclusion in their title — intra-articular MSC injections were no different from placebo [15].
A network meta-analysis of 43 trials and 5,554 patients reached a compatible conclusion by another route, ranking corticosteroid and hyaluronic acid most likely to be effective and adipose MSC among the least likely — while noting that all differences were small and possibly not clinically meaningful [17].
2.2 Against hyaluronic acid, benefit is reported consistently
The same 13-RCT meta-analysis that found nothing against placebo found MSC injection superior to hyaluronic acid: VAS pain MD 2.00 (0.94–3.07), WOMAC pain MD 4.58 (0.49–8.67), WOMAC total MD 14.86 (10.59–19.13). Its own caveat: the smallest treatment effect did not exceed the MCID, and the proportion crossing the MCID threshold was not significantly different (RR 0.57, 0.21–1.55) [15].
A 2025 meta-analysis of 10 RCTs and 818 patients with Kellgren-Lawrence grade I–III found MSC superior to hyaluronic acid at 12 months, WOMAC total mean difference −10.22 (95% CI −14.86 to −5.59), reaching the MCID for function at both 6 and 12 months, with comparable safety [21].
2.3 Across the whole literature, the balance is positive
An umbrella review published in 2024 pooled 28 meta-analyses covering 32,763 participants. It opens by acknowledging that prior meta-analyses have been contradictory — and concludes that MSC and platelet-rich plasma therapies were significantly associated with improvement in knee osteoarthritis scores, supporting their beneficial effects [16].
A 2026 meta-analysis of 28 randomised controlled trials found significant improvement in pain — ΔVAS mean difference −1.67 (p = 0.007), post-treatment VAS −3.55 (p = 0.01), KOOS pain +15.37 (p = 0.03) — and in function, KOOS activities of daily living +12.84 (p = 0.04) and KOOS sport +11.76 [20].
A 2025 network meta-analysis of 16 RCTs and 622 patients found MSC therapy significantly improved pain and knee function at 3, 6 and 12 months [23].
A 2025 meta-analysis restricted to patients who had not had surgery — 8 RCTs, 502 patients — found significant WOMAC improvement at 6 months (MD 7.44, 95% CI 1.45–13.42) and 12 months (MD 10.31, 95% CI 0.96–19.67), with no significant excess of adverse events [22].
Underneath these syntheses sit the individual randomised trials that repeatedly report benefit. A triple-blind, placebo-controlled trial of 43 patients found the MSC group improved significantly more than placebo in WOMAC total, WOMAC pain, physical function and painless walking distance [31]. Allogeneic bone marrow MSCs were compared with hyaluronic acid in 30 patients [29], and two doses of autologous bone marrow MSCs against hyaluronic acid in a multicentre phase I/II trial [30]. A randomised trial of autologous adipose MSCs reported clinically significant pain and functional improvement sustained at 12 months [32]. None of these is large, and several were open or single-blind — but the direction is consistent, and that consistency is what the positive syntheses are built on.
2.4 The two largest placebo-controlled trials split
Positive. A Phase III, multicentre, double-blind trial randomised 261 patients with Kellgren-Lawrence grade 3 to a single injection of autologous culture-expanded adipose-derived MSCs or placebo. At six months the treated group improved significantly more on a 100-mm VAS (25.2 vs 15.5, P = 0.004) and total WOMAC (21.7 vs 14.3, P = 0.002), with significantly more patients above the MCID and no serious treatment-related adverse events [18].
Null. ADIPOA2, a phase 2b double-blind trial in Annals of the Rheumatic Diseases, randomised 135 patients with mild-to-moderate disease to autologous expanded adipose MSCs at 2 × 10⁶ cells, 10 × 10⁶ cells, or saline. At six months, strict OARSI/OMERACT responders were 47.3% with cells versus 54.8% with placebo — RR 0.86 (95% CI 0.58–1.28), P = 0.46. No secondary outcome, including MRI cartilage thickness, showed a significant group difference [19].
Both are well-conducted double-blind trials of autologous expanded adipose MSCs against saline. They disagree, and no amount of pooling dissolves that.
2.5 What the pattern most likely means
The most economical reading: a real but modest symptomatic effect exists, of a size that placebo-controlled trials struggle to resolve. Against a weak active comparator such as hyaluronic acid it becomes visible; against a strong contextual response it often does not [27].
That is a different claim from "it works", and a different claim from "it does nothing". It is what the data support.
3. How Large Is the Effect, and Does It Matter to a Patient?
The MCID is the smallest change a patient reliably notices [28]. Three observations:
- Where syntheses report benefit against hyaluronic acid, the point estimates reach or approach the MCID — WOMAC total MD −10.22 [21]and 14.86 [15]on scales where roughly 12 points is the conventional threshold
- The lower bound of those intervals often does not clear the MCID, which is exactly what the 13-RCT synthesis flagged [15]
- Where responder analysis is reported rather than mean difference, the picture is weaker still: RR 0.93 and RR 0.40 for MCID crossing against placebo [15], and RR 0.86 for strict OARSI/OMERACT response in ADIPOA2 [19]
The honest summary is that the average effect is real in some comparisons and modest in all of them — closer to the boundary of perceptibility than the field's enthusiasm often implies, and well short of transformative.
4. What the Evidence Does Not Support
Structural modification. This is the clearest negative in the literature, and it matters because "regeneration" is the field's central promise.
- The Phase III trial found no significant difference in MRI cartilage defect change versus placebo at six months [18]
- ADIPOA2 found no significant difference in MRI cartilage thickness [19]
- A 2025 randomised trial of umbilical cord MSCs found no significant MRI structural change except in the synovium [26]
- The 28-RCT meta-analysis found structural scores were not significantly improved, and concluded the therapies are "primarily symptom-modifying rather than structure-modifying" [20]
Earlier reviews of a heterogeneous, largely low-quality body of work reported more encouraging imaging findings [14]. The better-controlled the trial, the weaker the structural signal has been. That trajectory is the finding.
Superiority to corticosteroid. A 2025 meta-analysis of 9 studies and 671 patients found stromal vascular fraction inferior to corticosteroid at three months and adipose MSCs not superior to hyaluronic acid at any timepoint, with complications arising from the liposuction harvest [25]. The 43-trial network meta-analysis ranked steroids first [17].
Durability beyond two years. Most randomised evidence stops at 6 or 12 months. Beyond that the data conflict: a double-blind placebo-controlled trial with 60-month follow-up found scores drifting back toward baseline by year five [38], while a separate five-year follow-up reported maintained benefit — uncontrolled, n = 11 [39]. A ten-year follow-up of 72 patients found no tumours and no deep infection [40]— reassuring on safety, but not an efficacy result.
Avoiding surgery. No trial has compared cell therapy against joint replacement; a 2026 umbrella analysis built to make that comparison searched seven databases across a decade and found zero direct comparisons [47]. The intervention with randomised evidence for deferring surgery is structured non-surgical care: among patients eligible for knee replacement who completed a 12-week supervised exercise, education and weight-management programme, two in three had still not had surgery two years later [48][49]. A cost-effectiveness model of a one-year delay strategy using platelet-rich plasma found replacement from the outset dominated it — more quality-adjusted life years at lower cost [50].
5. Dose: Four Independent Lines, One Conclusion
This runs against intuition, and against a great deal of marketing.
A dose-ranging RCT. Across four doses of allogeneic bone marrow MSCs, the lowest tested dose (25 million) was the most effective, the result did not reach significance against placebo, and adverse events were predominant in the higher-dose groups [34].
A survey of the field. A review of 914 registered MSC trials concluded that both lower and higher doses were less effective than an intermediate range [35].
A dose-escalation trial. Umbilical cord MSCs at 2, 20 or 80 million cells: all doses safe, but 100% of the high-dose group experienced injection-related knee swelling, and improvement was greater in the low- and medium-dose groups [36].
A 2026 network meta-analysis. Eleven RCTs and 602 participants, stratified by dose and tissue source: moderate-dose adipose MSCs produced the greatest improvement in pain and function at 6 and 12 months. High-dose was superior only at 3 months and induced more adverse events; low-dose carried significantly lower adverse-event risk [37].
ADIPOA2 adds a fifth data point from the null direction — neither 2 × 10⁶ nor 10 × 10⁶ separated from placebo [19].
Repeat dosing is a separate question, and it fares better. A 2025 network meta-analysis found repeated injections superior to single injections at 6 and 12 months — while also being associated with a higher incidence of adverse events [23]. In a randomised Phase I/II trial, repeated 20 × 10⁶ umbilical cord MSC dosing outperformed hyaluronic acid at 12 months on WOMAC-A (1.1 ± 1.3 vs 4.3 ± 3.5, P = 0.04) and VAS (2.4 ± 2.1 vs 22.1 ± 9.8, P = 0.03), in nine patients per arm [33].
More total cells delivered as a bigger single dose, and more cells delivered as a second injection months later, are not the same intervention. The evidence separates them.
6. Cell Source
No synthesis has established the superiority of one tissue source outright. The 2026 network meta-analysis found adipose-derived preparations performed best across most dose strata [37]; the 43-trial network meta-analysis ranked adipose MSC among the least effective interventions studied [17]. Bone marrow, adipose and umbilical cord have each produced positive trials and each produced null ones — ADIPOA2 and the Phase III trial used the same source and reached opposite conclusions [18][19].
What differs most is the quantity of evidence. Umbilical cord evidence rests on a smaller base: a dedicated meta-analysis found improved function and reduced pain (WOMAC MD −25.85, 95% CI −41.50 to −10.20) from only three randomised trials, which its authors flag as too few for confidence [24]. A 2025 randomised trial of 55 patients found overall improvements favouring umbilical cord MSCs that did not reach statistical significance, with benefit appearing in the subgroup that also had synovitis [26].
Allogeneic cells are given without donor matching, but MSCs are best described as immune evasive rather than immune privileged — they are cleared by the recipient, and their evasiveness is conditional [13].
7. Disease Severity
The European ESSKA-ORBIT consensus positions cell-based therapy as a second-line option in Kellgren-Lawrence grades 1–3, with inferior results in grade 4, and records that 22 of its 27 statements rested on grade C or D evidence [43]. Its companion consensus on blood-derived products issued three grade-A recommendations for platelet-rich plasma in KL ≤3 [44]— a useful measure of how much further ahead that evidence currently is.
The trial populations reflect this. The Phase III trial recruited only KL grade 3 [18]; the meta-analysis against hyaluronic acid covered KL I–III [21]; ADIPOA2 enrolled mild-to-moderate disease [19]. Evidence in advanced, bone-on-bone disease is sparse and less favourable.
8. Safety
In the joint. The 13-RCT meta-analysis found MSCs did not increase adverse events relative to either hyaluronic acid or placebo [15]; the 2025 meta-analysis of 8 RCTs found no significant excess [22]; the Phase III trial recorded no serious treatment-related events [18]. Against that, the 28-RCT synthesis found a higher frequency of local reactions [20], and the dose-escalation trial found universal injection-related swelling in its high-dose arm [36]. Post-injection joint pain and swelling are the common, expected events.
Systemically. The largest safety meta-analysis of MSC therapy across all indications found no association with acute infusional toxicity, organ system complications, infection, death or malignancy — and a significant association with transient fever (OR 16.82, 95% CI 5.33–53.10) [41]. Both halves of that finding are the finding.
Product quality. Unregulated preparations have caused real harm: a 2021 investigation traced 20 infections to a single umbilical cord product, with 65% of tested vials contaminated [42]. Manufacturing standard is not a detail.
Long term. Ten-year follow-up of 72 patients found no tumours and no deep infection [40]— the longest published follow-up, reassuring, but too small to exclude a rare event.
9. Why the Trials Disagree
Several explanations are plausible and not mutually exclusive.
The contextual effect is doing much of the work. If roughly half to two-thirds of the observed response is contextual [27], a genuine but modest biological effect will appear or vanish depending on whether the control arm captures that response. This alone explains the placebo-versus-hyaluronic-acid pattern without requiring anyone's data to be wrong.
Product heterogeneity. Different sources, doses, expansion protocols and delivery methods are pooled as though interchangeable [37].
Risk of bias in the early literature. The consistently positive early trials were consistently at high risk of bias [14], and better-blinded trials have tended to report smaller effects.
Outcome selection. A mean difference and a responder analysis can point different ways in the same dataset — as they do in the 13-RCT synthesis, where the MSC-versus-hyaluronic-acid mean differences were significant while the proportion crossing the MCID was not [15].
Publication and timing. Several of the largest null results have appeared only recently [19], so older syntheses could not include them.
10. Where the Guidelines Stand
Cell therapy is not a recommended intervention in the OARSI guidelines for non-surgical management of knee, hip and polyarticular osteoarthritis — whose core recommendations are arthritis education and structured land-based exercise with or without dietary weight management [45]— nor in the AAOS clinical practice guideline for non-arthroplasty management of knee osteoarthritis [46].
The ESSKA-ORBIT consensus does address it directly, positioning cell-based therapy as second-line in KL 1–3 on predominantly grade C–D evidence [43].
That deserves weighing in both directions. Guideline bodies are deliberately conservative and can lag genuine advances. But several independent groups reviewing the same literature and declining to recommend is information, not merely caution.
11. How the Mechanism Is Currently Understood
MSCs were originally expected to work by differentiating into cartilage [7]. Current understanding is that they act largely through paracrine signalling — secreting anti-inflammatory cytokines and trophic factors that modulate the joint environment rather than replacing tissue [8][9][10][12]. Preclinical work, including a goat model of osteoarthritis, showed chondroprotective and anti-inflammatory effects that motivated the clinical programme [11].
The gap between that mechanism and the clinical results is instructive: a treatment that modulates inflammation would be expected to relieve symptoms without necessarily rebuilding structure — which is precisely what the better-controlled trials show [20].
12. What Would Settle It
- Adequately powered placebo-controlled trials with pre-registered primary endpoints and responder analyses, in defined KL strata
- Head-to-head trials against corticosteroid and hyaluronic acid — the treatments patients actually face as alternatives
- Trials against joint replacement, which have never been run [47]
- Standardised product characterisation — source, dose, potency, viability — so that pooling means something [37]
- Follow-up beyond two years, where controlled data currently stop
- Prospectively powered structural endpoints, rather than imaging reported post hoc
Until several of those exist, the honest position is the one this page takes: a modest symptomatic effect, real against weak comparators, unproven against placebo, and not shown to change the joint.
Frequently Asked Questions
Does stem cell therapy work for knee osteoarthritis?
For symptoms, in mild-to-moderate disease, probably to a modest degree. An umbrella review of 28 meta-analyses concludes net benefit [16]and a 28-RCT meta-analysis finds significant pain and function improvement [20]. But the best test against placebo found no significant difference [15], and the most recent large placebo-controlled trial was null [19]. Anyone who tells you the question is settled, in either direction, is ahead of the data.
Why do different studies give such different numbers?
Mostly because they use different comparators, products and doses in a condition where placebo injection is itself strongly effective [27]. Section 9 covers this.
Will it regrow my cartilage?
There is no good evidence that it does. Neither large placebo-controlled trial found an MRI difference [18][19], and the 28-RCT synthesis describes the effect as symptom-modifying rather than structure-modifying [20].
How long does the benefit last?
Controlled evidence mostly ends at 12 months. Five-year data conflict — a placebo-controlled trial with 60-month follow-up found scores near baseline [38]; an uncontrolled 11-patient series reported maintained benefit [39].
Are more cells better?
No. A dose-ranging RCT found its lowest dose most effective with adverse events concentrated at higher doses [34], a review of 914 trials found an intermediate optimum [35], a dose-escalation trial found universal swelling and smaller improvement at high dose [36], and a 2026 network meta-analysis found moderate dose best at 6 and 12 months [37].
Is it safe?
Joint-level safety is reasonable — no increase in adverse events versus placebo or hyaluronic acid in pooled randomised data [15][22]— with expected post-injection pain and swelling [20]. Systemically, the largest safety review found no association with organ toxicity, infection, malignancy or death, but a significant association with transient fever [41]. Manufacturing standard matters a great deal [42].
Should I have it instead of a knee replacement?
No study has compared the two [47]. If the goal is to postpone surgery, the intervention with randomised evidence behind it is a supervised exercise, education and weight-management programme [48][49].
Why doesn't this page give one overall effect size?
Because pooling syntheses that disagree would manufacture a precision the literature does not have. The disagreement is the result.