“Cure stroke with stem cells” — just send in new cells to replace the neurons that died. It sounds intuitive, but it is mostly wrong. The MASTERS trial threw out that naive image and aimed instead at the immune system. At 33 sites in the UK and USA, 137 patients with ischaemic stroke 24–48 hours from onset were randomly assigned; the primary efficacy analysis covered the 126 patients in the dose tier that used 1.2 billion cells (Groups 2 and 3: 65 cell, 61 placebo) — a double-blind, placebo-controlled trial, which is rare in cell therapy. The answer at 90 days: odds ratio 1.08, 95% CI 0.55–2.09, p=0.83. There was no difference.
Journal Information
- Paper title: Safety and efficacy of multipotent adult progenitor cells in acute ischaemic stroke (MASTERS): a randomised, double-blind, placebo-controlled, phase 2 trial
- Authors: David C Hess (first author and corresponding author) and colleagues, 14 in total (last author Robert W Mays / Athersys). Affiliations include Medical College of Georgia, Augusta University, and others. The trial itself was run at 33 sites in the UK and USA, which is separate from the authors’ affiliations.
- Journal / ID: The Lancet Neurology 2017, Volume 16, Issue 5, pages 360–368. DOI 10.1016/S1474-4422(17)30046-7 / PMID 28320635 / ClinicalTrials.gov NCT01436487
- Peer review / publication date / Impact Factor: Published online ahead of print on March 17, 2017 (in the May issue). The most recent IF at that time was 23.468 (JCR 2015). The often-quoted 26.284 is from JCR 2016, released on June 14, 2017 — after the paper was already out in the world. As of August 2026 it is 54.6 (JCR 2025), ranked 1st out of 296 journals in Clinical Neurology.
- Open access: No (
is_oa = false). There is no CC licence and no PMC full text, so no figures or tables are reproduced here. - Funding / conflicts of interest: The sponsor was Athersys, Inc. (the company developing MultiStem). The paper itself states that the funder was “involved in trial design and data interpretation” and that “one employee sat on the writing committee.” Of the 14 authors, only 4 declared no conflict of interest.
Corresponding author David C. Hess has been chair professor of neurology at the Medical College of Georgia (since 2001) and became the 27th Dean in April 2017. In the declaration of interests: Hess receives research funding from Athersys and, through the university, holds patents on MultiStem cells and receives licensing income. Last author Robert W Mays is a co-founder and employee of Athersys and a patent holder.
What Was Not Known Before This?
As of 2017, stroke treatment had a clear “time wall.” According to the authors, the thrombolytic tPA (alteplase) can be given only up to 4.5 hours from onset, and endovascular thrombectomy only up to 6.0 hours in the UK and USA. Fewer than 5% of ischaemic stroke patients can benefit at all, and even with thrombectomy, up to 50% are dead or left disabled at 90 days. Both are treatments that “open a blocked vessel,” not treatments that “repair dead brain.” What MASTERS aimed at was that empty zone — 24–48 hours from onset.
That said, the “6-hour wall” itself later moved. Following the 2018 DAWN and DEFUSE 3 trials, thrombectomy is recommended up to 24 hours from onset in cases where perfusion imaging and similar tools can show that “there is still salvageable brain,” and this is reflected in Japanese guidelines as well. Even if time has passed, call the emergency services first. Eligibility is judged not by time alone but by imaging of the vessels and the brain tissue, the neurological signs, and the person’s level of function before the stroke.
👦 Student: Once 6 hours have passed, isn’t it already too late? The cells die in the first few hours, right?
🧬 Dr. Exotaro: Two answers. First, the time wall has moved. The most dangerous thing of all is deciding for yourself that “it’s too late” and not calling an ambulance. On top of that — an ischaemic stroke is not a single fire but a two-stage disaster. The first stage is the few hours when blood flow stops. But there is a second stage. Contents leaking out of dead cells act as an alarm, and immune cells pour into the brain. The paper wrote that this may further worsen the ischaemic injury and impede repair and recovery. What MASTERS aimed at is this second fire.
In the first week after a stroke the immune system is activated, and splenocytes and similar cells can target the ischaemic brain and worsen the injury. After that comes a period of peripheral immunosuppression accompanied by shrinkage of the spleen. This biphasic pattern was the pillar of the intravenous-delivery strategy.
What was used is MultiStem, that is, multipotent adult progenitor cells (MAPC). They are allogeneic cells derived from healthy, unrelated donors. With autologous bone-marrow-derived cells, “there is not enough culture time to deliver a therapeutically meaningful number of cells within the first week,” and so what was needed was an allogeneic cell product that “does not require histocompatibility testing.”
And then the decisive gap: “no large, multicentre, randomised, placebo-controlled trial of allogeneic bone-marrow-derived cell therapy existed.” That is MASTERS’ reason for being.
👦 Student: The other trials were saying it “looks promising” — was there really a need to check all over again?
🧬 Dr. Exotaro: Because you have to question what “it worked” really is. Stroke patients recover over a few months even without treatment, and they do rehabilitation. A score measured on a bad day drifts back toward the average when you measure it next. A trial with no control group books all of that as “the effect of the treatment.”
What Did This Paper Find?
The design was phase 2, multicentre, randomised, double-blind, placebo-controlled, dose-escalation. 33 sites in the UK and USA, from the start of enrolment on October 24, 2011 to the final follow-up on December 7, 2015. Eligible patients were aged 18–83 (originally 18–79), with an NIHSS of 8–20 immediately before dosing (a 0–42 scale, higher meaning more severe), and a cortical infarct in the anterior circulation on diffusion-weighted MRI (DWI) of more than 5 mL and less than 100 mL. Brainstem and lacunar infarcts, serious comorbidities, and a history of splenectomy were excluded — that single exclusion criterion alone shows how heavily this trial had bet on “immune modulation via the spleen.”
Assignment came in three stages. After safety was confirmed in Group 1 (8 patients assigned 3:1, of whom 6 received 400 million cells), Group 2 was assigned 1200 million cells (1.2 billion) at 3:1, and Group 3 the same 1.2 billion at 1:1. The primary analysis covered Groups 2 and 3, the dose tier using 1.2 billion cells — in the paper’s words, “the largest single dose ever administered as an intravenous cell therapy.” The placebo contained the same PlasmaLyte-A, DMSO, and human serum albumin as the cell product at identical concentrations, giving the control arm even the DMSO-derived breath odour that could otherwise have broken the blind. However, whether the blind actually held was never verified, and fever and chills were skewed toward the cell arm at 4 vs 0 patients.
Partway through the trial, slow enrolment led to the treatment window being widened from 24–36 hours to 24–48 hours (amended on July 26, 2013). The reason was that “the operating hours of the cell-processing facility were limited” — an operational constraint on the product side moved the trial design.
Let us be precise about the numbers. Across the whole trial, 137 patients were randomised (8 in Group 1 plus 129 in Groups 2 and 3), and 134 were dosed (8 plus 126). The frequently quoted “129 / 126” are numbers for Groups 2 and 3 only, the 1.2-billion-cell dose tier, and the flow runs: 160 assessed for eligibility → 129 randomised (67 cell / 62 placebo) → after excluding 3 who withdrew consent, 126 were dosed (65 cell, 61 placebo). The paper calls these 126 the “ITT population,” but strictly it is a modified ITT, and the definition wavers.
Meanwhile, all three of the main prognostic factors that drive 90-day outcome slightly favour the cell arm. Infarct volume 43.7 mL vs 50.9 mL, time to dosing 37.2 vs 39.3 hours, reperfusion therapy 59% vs 53%. Yet the only adjustment made was for baseline NIHSS category — infarct volume, reperfusion therapy, and site differences were all left out.
The primary safety endpoint was met. It was defined as dose-limiting toxicity (DLT) through day 7 after dosing, comprising three items: (1) a study-product-related CTCAE grade 3/4 infusion-related allergic reaction within 24 hours of dosing; (2) a study-product-related grade 3/4 adverse event at day 7; and (3) a worsening of NIHSS by 4 points or more judged to be study-product-related. You cannot afford to skim past the fact that the “study-product-related” qualifier is attached to all three criteria. The result was zero qualifying events in either arm. Deaths were 5 (8%) vs 9 (15%), with no significant difference (p=0.21). However, study-drug-related adverse events were significantly more frequent, at 15 (23%) vs 5 (8%) (p=0.018).
And the primary efficacy endpoint was not met. The prespecified endpoint was “multivariate global stroke recovery at day 90” — combining (1) mRS (a 0–6 scale; 2 or less means broadly independent) of 2 or less, (2) an improvement of 75% or more in total NIHSS score, and (3) Barthel index (independence in activities of daily living, 0–100) of 95 or more, using generalised estimating equations (GEE). The result was “no difference … in global stroke recovery at day 90 (OR 1.08 [95% CI 0.55–2.09], p=0.83)”. At 1 year it was 1.48 (0.77–2.84), p=0.24, but a footnote in the paper states explicitly that all of the 1-year assessments are exploratory. The prespecified time point for judgement is day 90.
The prespecified secondary endpoints also showed no significant difference, every one of them. At day 90, mRS ≤2 was 37% vs 36% (p=0.93), NIHSS improvement of 75% or more was 40% vs 38% (p=0.79), and Barthel index ≥95 was 46% vs 44% (p=0.83) — differences of only 1–2 percentage points in each case. Only excellent outcome showed an 8-point gap, at 15% vs 7%, but at p=0.10 it falls short of significance. Every one of the numbers is slightly higher in the cell arm, but you must not read “the directions line up” as evidence of an effect. These are strongly correlated measures derived from the same underlying degree of recovery, not independent confirmation.
👦 Student: A p of 0.83 means “exactly the same,” doesn’t it…
🧬 Dr. Exotaro: That is exactly the trap. A large p-value does not mean “we proved they are the same.” It only means “we did not obtain evidence of a difference.” The 95% CI is 0.55–2.09. From “a harmful effect that lowers the odds by 45%” to “a beneficial effect that raises them more than 2-fold” — neither has been ruled out. The correct reading is not “it was proven not to work” but “the design could not detect even a clinically meaningful effect if one existed.”
This trial has another face. The exploratory biomarker analyses did show between-group differences suggesting that the immune response had shifted. In the mITT set (65 vs 58), 2 days after dosing the proportion of circulating CD3-positive T cells fell in the cell arm and rose in the placebo arm (p=0.001), and there was a difference in FoxP3-positive cells (regulatory T cells) as well (p=0.010). At day 7, TNF-α, interleukin 6, and interleukin 1 β were lower. The authors called this “some of the first data” showing that cell therapy modifies the immune response in humans with acute stroke.
But this is not established fact. A total of 8 cytokines were measured at 4 time points, yet the main text gives only p-values; effect sizes and confidence intervals are left to the appendix (the appendix was not checked for this article). On top of that, the 58 on the placebo side is the number left after excluding 3 patients who died before day 7 and yielded no post-dose data — the paper does not describe those 3 as “the most severe,” but inflammatory markers tend to run higher with greater severity, and the effect of the exclusion has not been examined. And above all, no causal link has been shown between the immune shift and the patients recovering.
👦 Student: The immune system really did shift, and yet the patients’ recovery didn’t change…?
🧬 Dr. Exotaro: Yes. That is the heaviest single line in the paper. The intended target appeared to have been engaged (target engagement). And still, nothing changed at 90 days. An arrow hitting the target and that one arrow deciding the match are two different stories.
In the limitations section the authors say that widening the time window from 36 hours to 48 hours “might have diluted” the effect, and they present a post hoc analysis restricted to patients under 36 hours. But global stroke recovery was OR 1.64 (0.75–3.60), p=0.21 at day 90, and 1.62 (0.75–3.49), p=0.22 at 1 year. Even the most favourably shaped analysis, excluding patients who also received reperfusion, gives p=0.06. The items that reached significance are all exploratory or post hoc, and not one of them is in the prespecified day 90 analysis.
How Will the Future Change? (The Road to the Clinic)
The paper closes by noting that “a clinical trial in an earlier time window (under 36 hours) is planned (NCT02961504).” We now know how that continued.
Act one: the TREASURE trial (Japan, phase 2/3). That NCT02961504 turned out to be TREASURE (Houkin K et al., JAMA Neurol 2024;81(2):154–162). 206 patients at 44 sites in Japan, given a single intravenous dose of 1.2 billion cells within 18–36 hours of onset. The result: excellent outcome at day 90 was 11.5% vs 9.8%, p=0.90. The MASTERS post hoc finding did not replicate once it was tested under prespecification.
And then — something very similar happened in TREASURE too. In a post hoc analysis at 1 year, global stroke recovery came out “significant” at 27.9% vs 15.7% (adjusted risk difference 11.0%, 95% CI 0.8–21.3, p=0.04). No difference at the prespecified 90 days, and a difference only in the 1-year analysis added afterwards. But they are not identical. MASTERS’ 1-year global stroke recovery was OR 1.48, p=0.24, which is not significant; what reached significance there were excellent outcome (p=0.0206) and mRS ≤1 (p=0.0410). The endpoints and the patient populations are both offset. Does it genuinely work with a delay, or did a similar analytical structure produce a similarly shaped false positive twice? There is not a single piece of prespecified evidence supporting the former.
Act two: MASTERS-2 (phase 3). NCT03545607, 300 patients, with an mRS shift analysis at day 365 as the primary endpoint. On October 10, 2023, Athersys released the conclusion of a prespecified interim analysis by the independent Data Safety Monitoring Board (DSMB): “the planned sample size of 300 provides insufficient power to meet the day 365 primary endpoint,” and no safety concerns were identified. Because the required sample size would be large, new enrolment was paused. This is not “a final efficacy analysis that concluded it does not work.” But as of August 2026 the data from roughly 200 patients has still not appeared in a peer-reviewed paper or in results registration. Data from the people who took part in a trial does not necessarily reach the world.
Act three: the company disappears, and where things stand now. Athersys, Inc. filed for Chapter 11 bankruptcy in the USA in January 2024, and on April 3 of that year Healios K.K. of Japan acquired substantially all of its assets, including MultiStem. On December 9, 2025, Healios stated that “for the time being we will prioritise developing HLCM051 as a treatment for ARDS,” and said that for ischaemic stroke it would “re-evaluate the development policy.” In the financial results materials of May 14, 2026 it still reads “Development policy review,” and the company’s website describes a global phase 3 as “under consideration”. This is not a formal discontinuation. But 9 years on from MASTERS, the next confirmatory trial has still not begun.
👦 Student: So does that mean cell therapy for stroke has failed across the board? I saw news that a drug called Akuugo went on sale in Japan.
🧬 Dr. Exotaro: That is a very important distinction. Akuugo Intracerebral Implant Injection was launched on May 21, 2026, and it is a cell medicine implanted directly into the brain. But here is the thing — its approved indication is “improvement of chronic motor paralysis associated with traumatic brain injury,” not ischaemic stroke and not intracerebral haemorrhage. ACTIsSIMA, which tested the same cells (SB623) in 163 patients with chronic ischaemic stroke, failed to meet its primary endpoint (p=0.6743; results registration for NCT02448641), while STEMTRA in traumatic brain injury did meet its endpoint and reached approval first.
If you skim past the target disease, you could easily come away thinking a drug effective for ischaemic stroke has been approved. To put it precisely: as of August 2026, there is no drug and no regenerative medical product established by high-quality controlled trials to reliably restore the after-effects of stroke themselves. That does not mean “zero approved indications,” however. The package insert for citicoline injection lists an indication of “promotion of upper-limb functional recovery in patients with stroke hemiplegia,” under conditions such as being within 1 year of onset and combined with rehabilitation. Having an approved indication and being backed by strong evidence of effect are two different things. Akuugo’s conditional and time-limited approval is likewise a scheme that “estimates” efficacy from a limited number of cases; HeartSheet, approved under the same scheme, had its approval withdrawn in 2024 — a scheme for “delivering quickly” is not a scheme for “staying permanently.”
How to Read This Study Critically (Limitations, and Ways to Raise the Quality Further)
First, in fairness. This trial clearly sits a level above the rest of stroke cell therapy in terms of evidence. Central randomisation, a placebo matched exactly in composition, prospective registration, 365-day follow-up. And above all — it stated plainly in the abstract that the primary endpoint was negative, and published it. The design paper (Int J Stroke 2014;9:381–86) was published online on May 22, 2013, about 19 months after enrolment began (the print version in April 2014). In other words it is a document recording the design from before the protocol amendment of July 26, 2013, and no paper describing the amended protocol can be found. Verifying what was prespecified requires cross-checking against the historical versions on ClinicalTrials.gov.
The disclosure of limitations is good. The 1-year analyses are labelled exploratory, the post hoc analyses “post-hoc,” and the footnotes to Tables 2 and 4 state clearly that no adjustment was made for multiplicity. It is not complete, though. Of the prespecified exploratory items, the change in MRI infarct volume appears neither in the main text nor in the tables. Conversely, “initial length of hospital stay,” which is not in the Outcomes, does appear in a table, and in a post hoc analysis it reaches significance at p=0.0164. What they said they would measure does not appear, and what they never said they would measure comes out significant. There are 6 points that deserve weight.
(1) The post hoc analysis breaks the randomisation. The comparison in Table 4 is “the cell arm dosed within 36 hours (n=31)” versus “placebo patients (n=61)” — only the cell arm is narrowed to under 36 hours, while the placebo arm keeps all patients from 24 to 48 hours as the comparator. “The effect of treatment” and “the effect of time to treatment” are confounded, and the significant difference at 1 year can equally be explained by the known fact that patients treated earlier have better outcomes. The fix is to make the time window a stratification factor in the randomisation, and to match the comparison as “cell <36h versus placebo <36h.”
(2) The power assumptions were unrealistic. The paper says “more than 90% power with about 125 patients,” but the assumption is a between-group difference in binary responder rates, not a difference in raw scale scores — 10 percentage points for the mRS responder rate, 20 points for NIHSS, and 20 points for Barthel (on a 0–6 mRS, a “10-point” difference is impossible on the scale). The observed differences were 1, 2, and 2 points respectively, and no supporting data are cited. The responder definition is too strict as well. For a patient with a median NIHSS of 13, “improvement of 75% or more” means roughly 10 points, more than twice the conventionally meaningful change (around 4 points). Improvement that is moderate but still meaningful is structurally missed. The fix is to align the responder threshold with a clinically meaningful change, and to make the primary endpoint an ordinal mRS shift analysis at day 90 alone.
(3) Around 45 p-values, with no multiplicity correction. 18 in Table 2, 18 in Table 4, and 9 or more in the main text. Run 45 tests and, even if every null hypothesis is true, an average of 2 will come out “significant.” And the findings labelled “significant” for clinical efficacy all fall between p=0.0081 and 0.0410, and not one of them passes the rough Bonferroni threshold (0.05/45 ≈ 0.0011) if they are treated as a single family of tests. The fix is to prespecify a hierarchical testing procedure, and to report Holm- or FDR-adjusted p-values alongside.
(4) The handling of missing data can mechanically manufacture the “good finding” at 1 year. Early dropout was 13 in the placebo arm and 7 in the cell arm, and those figures include the 9 and 5 deaths (dropouts and deaths are not counted separately). Missing data were handled by last observation carried forward (LOCF), and for binary outcomes the paper states that death was treated as “non-response.” LOCF disadvantages the arm with more dropout, which is consistent with the temporal pattern of no difference at day 90 and a difference only at 1 year. The fix is to make multiple imputation the primary analysis, and to handle death within a competing-risks framework or with a win ratio.
(5) The very thing they bet on was never once measured. This was a trial that bet on “immune modulation via the spleen” strongly enough to make prior splenectomy an exclusion criterion, and yet the authors write: “because we could not quantify spleen size, we could not determine the effect of MAPC treatment on spleen size.” It ended with the centre of its own hypothesis never observed. The fix is to measure spleen size serially with abdominal ultrasound. The very references this paper cites (Vahidy 2016, Sahota 2013) themselves measured spleen changes in human patients with acute stroke.
(6) Are these patients like your own family? The age range was 18–79 originally and 18–83 after amendment, so anyone 84 or older was excluded from the outset. Mean age was 61.8 / 62.6, and those 65 and over made up somewhere in the 40% range. Lacunar infarcts and brainstem or posterior-circulation infarcts were also out of scope. The most eloquent number of all: 33 sites over 4 years managed to randomise 137 patients, about 4 per site. Even if there were an effect, the range it could be applied to was narrow from the start. In fact, the mean age in TREASURE, which tested the same drug in Japan, was 76.5.
The cell product has weak points too. Although it was thawed and prepared individually at 33 sites, at least in the main text of the primary paper there are no data on post-thaw viability, elapsed time to administration, or lot-to-lot potency. “1.2 billion cells” is a nominal dose. Fischer UM, et al. (Stem Cells Dev 2009;18:683–92), which the paper cites, is a preclinical study in rats that showed most intravenously infused cells are trapped in the lungs (the pulmonary first-pass effect). Biodistribution was not actually measured in the MASTERS participants.
👦 Student: Even a trial this good has this many things to fix?
🧬 Dr. Exotaro: It is the other way round, really. The reason the criticism can be written this concretely is that this trial is transparent enough. With a single-arm trial that only says “patients got better after dosing,” you cannot even build a foothold for criticism. It is precisely because they labelled the post hoc analyses as post hoc, and noted in a footnote that there was no multiplicity correction, that we can debate it at all. Being verifiable is itself a scientific virtue.
Dr. Exotaro’s Perspective
I myself work with mesenchymal stem cells (MSC) and extracellular vesicles (EV), studying recovery in spinal cord injury (SCI) and neurological disease. What hit me hardest was this one point: the biomarkers moved, and the patients did not change. If I saw the same movement in an animal experiment, I would say “it’s working” without hesitation. And yet day 90 gave OR 1.08, p=0.83. Changes in inflammatory markers are not the goal in themselves, only a waypoint on the road to functional recovery.
The Discussion is austere about intravenous cell therapy, writing that when bone-marrow-derived cells are given intravenously, “direct entry into the brain and engraftment are limited, and neuronal replacement is unlikely to occur.” Intravenous MAPC is “a drug that calms the immune system”; the intracerebral implant type is “something that promotes plasticity inside the brain” — the same phrase “stem cell therapy” can point at completely different biology.
So what could EVs change? From here on, please read with my own position disclosed: I research EVs myself. What was never measured in MASTERS was where the cells went, and what I hope for from EVs is the possibility of shrinking that “could not be measured.” That said, you cannot conclude “it did not work because it did not get there,” and no standard has been established for how to measure EV potency either. And above all — as of August 2026, no results from a human controlled trial showing that EVs or “exosome infusions” work for stroke have been published (randomised trials themselves are ongoing). There are clinics offering exosome administration as self-pay treatment, but that sits short of the hurdle MASTERS could not clear. The caution that applies to stem cells applies just as much to exosomes.
To readers who are patients or family members. Even in the MASTERS placebo arm, 36% reached mRS ≤2 and 44% reached Barthel index ≥95 at day 90 — there were people who recovered this much without receiving any cell therapy. The Japanese Society for Regenerative Medicine has also warned that self-pay cell therapies “have not undergone review of their safety and therapeutic effect” (May 2022). A world-class team ran a 4-year controlled trial, and neither TREASURE nor MASTERS-2 that followed could produce an answer — more than 10 years have been spent, and we are still not at the stage of being able to say “it works.” Knowing that is, I think, the most practically effective preparation you can have.
Even so, I have hope for this field. MASTERS was not a failure; it was a trial that asked the question properly and received the answer properly. Up to the highest dose examined — a single intravenous infusion of 1.2 billion cells — no dose-limiting toxicity occurred — although a maximum tolerated dose was not identified, so the accurate statement is that the ceiling on safety is still not in sight. Only here can the next question finally be posed: is what is missing the amount, the timing, the way it is delivered, or the way patients are chosen? Not putting the data left behind by 137 patients and their families back on the shelf as “the story of something that did not work” — that, I believe, is the job of those of us who come after.