BCI Weekly - March 29, 2026
2026 week 13 (March 23-29). On-Device Decoding, Portable Stimulation, and a Quietly Important Week
Not every week in neurotech delivers a headline-grabbing implant milestone (this one didn’t). But it did surface a pattern that matters: some of the most credible progress is happening in the infrastructure layer. Topics touched include on-device decoding, portable neuromodulation, better calibration, and interfaces that reduce friction rather than maximize sci-fi.
The top research news this week was a wireless bidirectional neural interface. It moves spike decoding onto the device itself. This might seem like a small detail, but it addresses a major challenge in real-world BCI systems. If decoding relies on an external computer, untethered and practical implants remain limited.
Meanwhile, the neurotech landscape continued to evolve. China approved a new commercial BCI device. Science Corp secured more funding for its retinal implant program. Cognito gained financing and shared new EEG-linked Alzheimer’s data.
Though this week had fewer “BCI company launch” announcements, it still provided valuable insights into the field’s direction.
Three themes that stood out
More computation is moving onto the device.
That’s where implantable systems need to go if they’re going to become practical outside tightly controlled lab settings.Portable neuromodulation keeps gaining clinical credibility.
This week’s tACS trial for depression reveals a trend. At-home stimulation systems are now becoming regulated products, not just experiments.The most practical interfaces may not be purely brain-first.
The best control results happened when we combine different physiological signals. A wrist-worn ultrasound system was used to decode motor intent. This method didn’t rely only on the brain.
The big story: on-device decoding gets more real
A bidirectional neural interface with direct on-device neuromorphic decoding for closed-loop optogenetics
The most important paper this week, in my view, was a new wireless bidirectional neural interface. A compact headstage combined recording, stimulation, and an on-chip neuromorphic spike decoder. The headline is architectural: the decoding happens on the device rather than being offloaded to an external computer. A lot of “closed-loop” neural systems are only closed-loop in a limited lab sense. Once the computation has to leave the device, you introduce latency, hardware bulk, and practical constraints that make the system much harder to translate.
If you care about next-generation implantable BCIs, this is the kind of paper worth paying attention to. It addresses the real bottleneck of fitting meaningful real-time decoding into miniaturized wireless hardware.
Better hardware is not enough by itself. The field also needs better deployment of decoding.
Clinical signal: portable stimulation keeps advancing
A multicenter randomized clinical trial of portable tACS for major depressive disorder
One of the more consequential translational papers this week is a multicenter randomized clinical trial. They used portable transcranial alternating current stimulation to treat major depressive disorder.
For readers interested in neurotech products and regulations, its significance goes beyond depression. The field is exploring whether non-invasive brain stimulation can move out of specialist clinics and into more practical deployment models. This evidence helps shapes that answer.
I see this as part of a larger body of evidence, not just a single turning point. However, it shows that portable neuromodulation is gaining recognition in neurotech.
The most scalable neurotech products may arrive through portable stimulation sooner than through high-profile implants.
Adaptive difficulty in motor-imagery BCI training after stroke
A small but important clinical paper looked at using distance-to-bound to change task difficulty during motor-imagery BCI training for stroke patients.
The idea is simple: many motor-imagery BCIs are too hard for users, especially in rehab settings. Here, neural signals can be weak, inconsistent, or hard to classify. If the system can gauge how distinct a user’s signal is from the classifier boundary, it can adjust difficulty. This makes the training loop more adaptive and user-friendly.
This study involves only two cases, so it shouldn’t be overstated. However, it tackles a key issue: not whether MI-BCIs work, but if enough people can actually use them.
Usability is a major hurdle in clinical BCI, and adaptive training offers a promising solution.
A broader pattern: multimodal and peripheral interfaces keep getting stronger
EEG + ECG + EOG for attention classification
A new paper on multimodal attention-classification introduces a deep-learning model. This model combines EEG, ECG, and EOG signals instead of focusing only on EEG.
This approach is significant. Non-invasive BCI has long aimed for cleaner decoding from noisy EEG alone. However, in many real-world applications, it may be better to merge signals from different physiological sources. Each type can fill in the gaps left by the others, making the overall system stronger than one focused solely on the brain.
The future of practical non-invasive BCI may rely on multimodal systems, not just EEG.
A wrist-worn ultrasound interface that decodes hand intent
A new paper on multimodal attention-classification introduces a deep-learning model. This model uses EEG, ECG, and EOG signals together, rather than focusing only on EEG.
This approach is important. Non-invasive BCI has struggled with noisy EEG data. In many real-world cases, merging signals from different sources may work better. Each signal type can compensate for the others, making the system stronger than one that relies only on brain signals.
The future of non-invasive BCI may depend on multimodal systems, not just EEG.
Field context: what happened outside of research
The research feed this week was lighter on classic company and regulatory headlines than the broader neurotech news cycle. A few developments are worth layering on top of the paper roundup:
China approved a commercially authorized BCI device. China’s NMPA has cleared the first invasive BCI medical device. This system, from Shanghai’s Borui Kang Medical Technology, aims to help quadriplegia patients regain hand grasp, including through robotic-glove control (Reuters, MobiHealthNews). This is significant for two reasons. First, it shows that commercialization can happen outside the most popular companies or regions. Second, it raises competition in the implant and neuromodulation fields.
Science Corp raised $230 million. Science Corp completed a $230 million Series C to advance its PRIMA retinal implant program. This funding will support U.S. trials and pending regulatory decisions (press release, STAT). This is important even if your focus is on BCI instead of vision restoration. It shows that investment continues in neurotechnology platforms with clear regulatory paths.
Cognito raised roughly $105 million and added EEG-linked Alzheimer’s data. Cognito Therapeutics raised about $105 million in a popular Series C to further its gamma-frequency Spectris system for Alzheimer’s disease. They presented data at AD/PD 2026 showing reduced EEG slowing in treated patients (press release, Fierce Biotech). This connects to this week’s EEG-focused research. It shows that brain-signal biomarkers are becoming part of real products and trials.
Neuralink continues to frame 2026 as a scale-up year. Reporting on Neuralink highlights plans for larger device production and more automated surgical workflows by 2026 (Reuters, Business Insider). Regardless of your view on that timeline, it shapes the conversation about implantable interfaces.
Worth tracking
Meta’s TRIBE AI: a big-tech push toward cross-individual neural decoding and foundation-model-style brain-response prediction.
Read the coverageSpike sorting evaluation gets more rigorous: a Journal of Neurophysiology paper introduces effect sizes and statistical power analysis for benchmarking sorting algorithms.
Read itAutomated TMS thresholding could reduce setup friction:
RMT-Finderproposes an automated method for determining resting motor threshold from MEP amplitudes.
Read itResting EEG + HRV can decode cognitive state: another sign that passive and multimodal decoding may be more practical than task-heavy paradigms in some settings.
Read itDBS target selection in Parkinson’s gets a network-level lens: a Molecular Psychiatry paper compares how GPi versus STN stimulation modulate inter-hemispheric dynamics.
Read itDBS for treatment-resistant depression stays on the radar: chronic implanted neuromodulation is steadily expanding beyond movement disorders.
Read the coverageTime-varying DCM for slow cortical potentials: methodologically interesting for anyone thinking about non-stationary control signals in BCI.
Read itTransformer embeddings for aphasia recovery: worth watching for speech prostheses and language-model-informed neurotech.
Read itCross-participant encoding models: another attack on the decoder-calibration problem.
Read itRepresentational drift during learning: relevant to adaptive decoders that need to survive non-stationary neural signals over time.
Read it
My take
This week wasn’t full of big BCI news. Still, it showed some important changes.
The real progress didn’t come from bold claims about mind reading or quick consumer brain interfaces. It came from better hardware, smoother stimulation, multimodal decoding, and systems that tackle control issues in practical ways.
While this may not be flashy, it’s likely more predictive.
If the last few years focused on proving neural interfaces can work, the next years will challenge us further. We need to make them portable, easy to calibrate, manufacturable, and practical outside demos.
This week felt like a glimpse into that future.
If you’re involved in this field, I’d love to hear your thoughts on the tradeoff between brain-first interfaces and peripheral-plus-multimodal systems.