By Sharena Rice, Contributing Editor
September 17, 2026 | Machine Medicine’s FDA breakthrough designation for their video assessment platform, hiring at Metamorphic and Epia Neuro, and a leaked Apple AirPods demo all point to the same patterns: computer vision can give better situational understanding and standardization in a world where human raters find challenges with consistency. Because cameras, LiDAR, and other types of spatial sensor systems are becoming inexpensive and ubiquitous, computer vision systems are increasingly easier to implement both in research settings and in everyday life.
Metamorphic, which builds foundation models from large scale neural recordings, recently posted job listings for a computer vision research scientist and a computer vision research engineer. Epia Neuro, which emerged from stealth in April with a read/write brain computer interface for stroke, describes a system that fuses neural signals with contextual data from external sensors to infer user intent. Neither firm leads with cameras in their public stories. Both are building in directions where computer vision carries weight, and the same discipline that lets a phone unlock on a face has been doing load bearing work inside neuro devices for years.
Nowadays, frameless neuronavigation for transcranial magnetic stimulation depends on optical tracking: infrared cameras follow markers on the patient’s head and the coil, and software registers that geometry against the patient’s MRI so the operator aims at a cortical target rather than a scalp landmark. Focused ultrasound faces greater challenges. Steering an acoustic focus through skull and tissue to a target is crucial: subtle shifts in the ultrasound transducer location relative to the skull can make an order of magnitude difference for how much energy makes it to the target region of interest. Drift between real anatomy and the plan degrades the dose delivered and the reliability of dosing, and this is exacerbated as distances between the scalp and the brain region of interest increase. Eye tracking platforms, used as both assessment tools and control channels, are vision systems end to end. The sensor systems are not an afterthought: the computer vision stack is the instrument that keeps the therapy reliable and precise while making it easier to validate whether the setup was done properly.
DeepLabCut, released in 2018, let neuroscientists track unmarked animals from ordinary video with the precision that had previously required physical markers, turning behavior into a dense, cheap signal that can be aligned to neural recordings frame by frame. The lesson for the field of neurotech is that behavior is an information-rich readout that can improve understanding of neural correlates, and the tooling to extract it has matured to the point where a small team can deploy it.
On sensing, vision lets a device read the body without adding channels to the brain. A stroke interface that infers movement intent can cross check that inference against what the limb is visibly doing, tightening a closed loop without a denser array, the profile of an intent fusion architecture like Epia’s. On stimulation, vision makes targeting repeatable across operators and sites, since most variability in TMS and focused ultrasound response isn’t biological. It comes from how consistently the device was aimed and held, which today depends on operator skill. Registration by camera and correction for patient motion as it happens turn that manual step into something an automated system can check, which is the point for a device company: a system whose targeting no longer hinges on operator expertise is easier to validate and to run consistently regardless of the treatment setting that determines whether a therapy scales past its pivotal trial.
The same argument applies to measuring the patient rather than aiming energy at the nervous system. Machine Medicine Technologies built KELVIN, a platform that records a patient performing standard motor tasks and scores them with pose estimation, tracking movement to the fingertip for symptoms such as tremor and bradykinesia. It attacks rater variance: scoring on the MDS-UPDRS motor exam is subjective and drifts within a rater and across sites, and that variance lands in the endpoint of every trial that uses it. KELVIN is in use at DBS centers internationally, and the company reports FDA granted its technology breakthrough device designation as software for automated clinical assessment in Parkinson’s disease.
The access case is larger. Machine Medicine also runs an abbreviated version at home, on a phone, with an avatar demonstrating each task so the patient performs it correctly without a clinician present. Movement disorder expertise is concentrated in a small number of academic centers, patients far from one get assessed less often, and a guided video protocol that runs anywhere with a phone changes the catchment a sponsor can enroll from.
Neurotech devices that use computer vision will take time to become a standard part of the healthcare system. Everyday wearables have been used to build regulated functions in neurotech before, including Rune Labs’s 510(k) for monitoring Parkinson’s symptoms via the Apple Watch. The perception stack in electronics will expand what is cheap to observe and come to change what is billable.
The asset in this space is not algorithms, as estimation backbones are open and improving on somebody else’s budget. Labeled clinical video tied to outcomes, validated protocols, and a designation attached to a specific indication are more valuable.
In this space, watch whether Apple pairs camera equipped earbuds with any FDA clearances in the coming years, whether KELVIN’s designation converts into a cleared indication with labeling for remote scoring, which companies double down on computer vision to improve their neuromodulation targeting procedures and improve their clinical workflows, and whether computer vision is explicitly described in the research publications of NeuroAI rather than kept in the job listings.
