Portable MRI at the stroke bedside what the evidence shows.

What did the 2020 press release actually report?

The American Heart Association release, dated 12 February 2020, described a Yale team led by Kevin Sheth taking a portable, low-field MRI into hospital rooms and imaging 85 stroke patients within seven days of symptom onset. The group spanned the main stroke types: 46% ischemic, 34% intracerebral haemorrhage, and 20% subarachnoid haemorrhage, in patients aged 18 to 96.

The practical details were the point. The scanner ran near other hospital equipment without interference, staff did not have to clear metal from the room, and the average exam took about 30 minutes. No significant adverse events were reported. Five patients could not fit through the 30-centimetre opening and six experienced claustrophobia.

The detail that matters six years on

That 85-patient cohort was an oral presentation at a conference, and it was never published as a peer-reviewed paper. The group's published feasibility study, in JAMA Neurology later that year, reports a different and smaller cohort: 50 intensive care patients. Anyone citing the release should not describe the 85-patient work as peer-reviewed, and should not treat the two as the same study.

Why does bedside imaging matter at all?

It is worth stating the case for this technology properly before picking at it, because the case is strong and rests on two independent problems: time and access.

The time problem

The standard reference is Jeffrey Saver's 2006 quantification in Stroke, which modelled what an untreated large-vessel ischemic stroke costs per unit of time. His estimate: about 1.9 million neurons a minute, roughly 120 million an hour, such that the ischemic brain ages about 3.6 years for every untreated hour. These are modelled figures from typical infarct volumes rather than per-patient measurements, but they set the scale.

Imaging sits in that critical path, because you cannot treat a stroke until you know whether it is a clot or a bleed. The 2026 American Heart Association and American Stroke Association guideline recommends that hospitals establish protocols so emergent brain imaging happens as rapidly as possible, giving 25 minutes as the example target, at its strongest recommendation class.

That target is frequently missed. In the Florida Stroke Registry, covering 63,265 patients, door-to-imaging within 25 minutes was achieved in 56% of cases, improving from 36% in 2010 to 72% in 2018. A separate multicentre analysis of 23,364 patients in north Texas found only 19.1% met it.

The access problem

The second problem is starker. A World Stroke Organization and World Health Organization survey of stroke services across 84 countries found that cranial CT was available in just 57.2% of the centres surveyed. Stroke units existed in 91% of hospitals surveyed in high-income countries against 18% in low-income ones. That survey measured CT only, so it says nothing about MRI availability.

The broader picture matches. The 2021 Lancet Commission on diagnostics concluded that 47% of the world's population has little to no access to diagnostics at all, and that fewer than 10% live within two hours of a facility offering imaging of any kind.

Underneath those service figures is raw equipment scarcity. Drawing on the International Atomic Energy Agency's IMAGINE database, a 2021 Lancet Oncology Commission reported the number of people served by a single CT scanner by country income group:

That is roughly a 68-fold gap, and the Commission notes the disparity is wider still for MRI. As a country-level illustration, a 2021 survey found Ethiopia had 38 CT scanners and 11 MRI machines for a population of about 112 million. These scanner-density figures come from a commission framed around cancer care, so they describe imaging capacity in general rather than stroke services specifically.

What has the peer-reviewed research found since?

Four studies carry the weight, and they answer progressively harder questions: is it feasible, does it find blood, does it find infarcts, and how accurately.

The Stroke 2023 result is the most instructive of the four, because it isolates a variable. The same scanner, reading the same kind of patient, went from 77.8% to 96.6% sensitivity for haemorrhage depending on whether deep-learning image reconstruction was applied. The FDA cleared that reconstruction software in November 2021. The improvement came from the algorithm, not from a better magnet.

Two caveats belong with these numbers. The Science Advances 90% is a detection rate, not a sensitivity: every patient in that cohort already had a confirmed stroke, so the study could not measure false positives. And in the Stroke 2023 study, raters saw the clinical information alongside the images, which is realistic but flatters accuracy relative to a blinded read. The authors say so themselves.

How accurate is portable MRI really?

The best answer comes from the study with the least at stake. In 2026 a Mayo Clinic-led team, independent of the Yale group and the manufacturer, pooled 9 studies and 474 patients with acute ischemic stroke.

A sensitivity of 73.3% means roughly one in four confirmed infarcts was missed, and the review notes the missed lesions were generally small or sub-centimetric. That is a materially lower figure than the 90% detection rate from the manufacturer-affiliated case-only cohort, and it is the number a careful reader should carry.

The authors' conclusion is the practical one: portable MRI should not be relied upon as a stand-alone rule-out test for acute ischemic stroke. A negative scan does not clear a patient.

Substantial heterogeneity across the pooled studies (I² = 74.2%) means these estimates cover a range of scanner generations, software versions, and reading conditions rather than one settled performance level. Newer hardware and sequences continue to improve, which is a reason to expect the numbers to move rather than to treat any of them as final.

What can a portable MRI not do?

The physics sets the ceiling. These scanners run at 0.064 tesla against the 1.5 to 3 tesla of a hospital MRI, roughly 23 to 47 times weaker, with gradient hardware around a tenth as strong. Less signal means lower resolution and more noise, and diffusion-weighted imaging, the sequence that finds early ischemic stroke, suffers most.

• Small and posterior lesions get missed. In the haemorrhage study, the false negatives clustered in the posterior fossa: cerebellar, pontine, and thalamic bleeds that all or most raters missed.
• It cannot image blood vessels. These systems do not perform angiography or perfusion imaging, so they cannot identify a large-vessel occlusion or triage a patient for thrombectomy.
• Fine structure is out of reach. Compared against 3 T scanning, gross brain volumes corresponded reasonably well, but mean cortical thickness showed essentially no correlation (r = −0.05, not significant), and AI super-resolution only lifted it to 0.14.
• Body habitus excludes some patients. The opening is 30 centimetres. Five patients in the original release and nine in the haemorrhage study could not be scanned for this reason.

A 2023 review in Radiology, from researchers working on the technology, puts the constraint plainly: portable low-field MRI has the potential to transform neuroimaging but is limited by low spatial resolution and low signal-to-noise ratio.

What regulators actually cleared

This is the most commonly misstated fact about these devices. Hyperfine has received 16 FDA 510(k) clearances between February 2020 and December 2025, and across all of them the indication has stayed narrow and essentially unchanged. The system is cleared for producing images of the head where full diagnostic examination is not clinically practical, to be interpreted by a trained physician.

It is not cleared as a replacement for conventional MRI or CT, and it carries no stroke-specific diagnostic indication. The cleared role is filling a gap where the alternative is no imaging at all, which is a genuinely valuable thing to do and a different claim from matching a hospital scanner.

Who funded this research?

Largely the manufacturer, and the papers disclose it. Hyperfine funded the foundational studies. The company's founder and chairman holds significant stock, a co-author is a company co-founder and equity holder, and several co-authors are Hyperfine employees. The senior investigators report Hyperfine grants.

Disclosed industry funding does not invalidate a finding, and this work was published in serious journals with blinded raters and adjudication. It does argue for weighting independent replication more heavily, which is why the Mayo-led meta-analysis carries the most interpretive weight on this page, and why its lower pooled sensitivity is the number in our summary rather than the manufacturer-affiliated 90%.

What does this have to do with rehabilitation and chronic pain?

Directly, very little. Portable MRI is an acute diagnostic tool for people in hospital beds, not a treatment, and it has no bearing on chronic pain care. Karuna is not a stroke or imaging provider, and nothing here should be read as clinical guidance about stroke.

What carries over is the design principle. The interesting move is bringing the capability to the person instead of moving the person to the capability, and it is the same reasoning behind tele-monitored VR rehabilitation at home after stroke and behind virtual-first chronic pain care generally. Access is a clinical variable, not just a convenience one.

The second thing that carries over is how to read an announcement. A press release about a conference abstract, an impressive single-centre number, a manufacturer's funding line, and an independent meta-analysis that lands lower: that sequence recurs across medical technology, VR therapy included. It is the same discipline we apply to the Cochrane review on VR for stroke and to our own outcome data.