NEUROTECHNOLOGY

MoveAgain BCI Platform

Advancing Neural Decoder Performance for Paralysis Patients

Discover how AE Studio partnered with Blackrock Neurotech to optimize the MoveAgain brain-computer interface platform, improving neural decoder accuracy and cross-session robustness for the only FDA-approved implantable BCI device.

MoveAgain BCI Platform

THE CHALLENGE

The problem.

Brain-computer interfaces face a fundamental engineering challenge: neural signals drift over time. What worked in yesterday's session may not work today. Patients with paralysis using BCIs need systems that adapt quickly and maintain accuracy across sessions without lengthy recalibration.

AE Studio partnered with Blackrock Neurotech, creator of the Utah Array, the only FDA-approved implanted BCI device, to advance their MoveAgain platform. MoveAgain enables patients with paralysis to control cursors and devices using their thoughts, but the neural decoders powering the system required optimization for real-world clinical use.

The technical challenges were significant: improving decoder accuracy for cursor movement and click classification, reducing the time required to retrain models between sessions, and building tools that clinical teams could use to analyze sessions without specialized data science support.

THE SOLUTION

What we built.

Self-Supervised Learning for Cross-Session Robustness

The breakthrough came from applying self-supervised learning techniques to neural decoder training. Traditional supervised approaches required extensive labeled data from each session, creating a cold-start problem every time a patient began a new session.

AE Studio demonstrated that self-supervised pre-training across subjects and sessions yielded robust cross-session decoding with improved accuracy. By learning general representations of neural activity patterns before fine-tuning on specific tasks, the decoder could generalize across the natural variability in brain signals, reducing the burden on patients and clinicians while maintaining high performance.

Neural Data Transformer Architecture Optimization

We optimized the Neural Data Transformer (NDT) architecture for the MoveAgain platform. This included systematic hyperparameter tuning for both fine-tuning and pre-training phases, ensuring parameters transferred reliably across runs, sessions, tasks, and subjects.

We also investigated contrastive learning approaches to determine if they could provide additional improvements to performance or robustness compared to existing NDT implementations. Model stabilization methods were developed to minimize retraining time in new sessions while maintaining decoder accuracy.

Real-Time In-Session Analysis Tools

Clinical BCI sessions are expensive and time-constrained. Losing an entire recording due to an undetected hardware issue or environmental noise is costly. We built in-session analysis tools capable of identifying experimental problems in real time.

The tool detected issues including environmental noise, errors in experimental or hardware setup, and unexpected task-locked neural responses, allowing clinical teams to fix problems immediately rather than discovering them during post-session analysis. This transformed session efficiency and data quality.

Automated Multi-Session Analytics Platform

As Blackrock scaled to multiple participants with up to two sessions per month each, manual analysis became unsustainable. We built an automated analytics system that ingested data from MoveAgain sessions, computed relevant metrics, and stored results in a structured database.

The platform provided high-level dashboards for cross-session performance tracking alongside detailed reports for within-session analysis. This reduced time between data acquisition and actionable insights, freeing clinical researchers to focus on interpretation rather than data processing.

  • Centralized and automated storage of data, analysis, and trained models across all participants and sessions
  • Automated signal quality analysis including temporal and spectral noise characterization and artifact detection
  • Behavioral data quantification with velocity and trajectory profiling for cursor control tasks
  • Cross-session performance dashboards enabling rapid comparison of decoder accuracy across the participant's measurement history

Production Integration with Rust Development

Research findings needed to translate into the production MoveAgain platform. AE Studio's Rust developers extended the MoveAgain codebase to support self-supervised and pre-trained models, including adding a linear decoder and updating the NDT architecture.

This enabled closed-loop calibration improvements and brought research-demonstrated performance gains into the hands of patients. We also supported documentation and requirements generation for FDA compliance, leveraging our experience with both the research Python codebase and production Rust implementation.

HOW IT WORKS

The details.

Teaching the Decoder to Generalise Across Sessions

The core problem was that a decoder trained on one session would not work well in the next. Every time a patient sat down, the system was starting from scratch. We applied a learning approach where the model first learns general patterns across many subjects and sessions before being fine-tuned for a specific task. This meant the decoder already had a strong foundation each time it adapted to a new session, reducing the burden on patients and clinicians.

Improving the Architecture Systematically

We worked through the underlying model architecture in detail, tuning the settings that control how it trains and how it transfers what it has learned between sessions. We also tested alternative learning approaches to see if they could improve reliability or speed. The goal was a decoder that worked consistently across the natural variability in brain signals, not just under ideal conditions.

Catching Problems During the Session, Not After

Clinical sessions are expensive and hard to reschedule. Losing an entire recording because of a hardware issue discovered after the fact is a costly mistake. We built tools that monitor the session in real time and flag problems as they happen: environmental noise, equipment errors, and unexpected signal patterns. Clinical teams could fix issues on the spot rather than finding out the data was unusable the next day.

Automating What Used to Take Hours of Manual Work

As Blackrock worked with more participants across more sessions, reviewing each one manually became unsustainable. We built an automated analytics system that processes session data, computes standard metrics, and stores results in a structured database. Dashboards let the clinical team compare performance across sessions and participants at a glance. Detailed reports are available for any session that needs closer inspection.

  • Centralised storage of data, analysis results, and trained models across all participants and sessions
  • Automated signal quality checks for noise, interference, and data anomalies
  • Cursor control behaviour profiling including speed and path accuracy
  • Cross-session dashboards for tracking decoder performance over time

Bringing Research Findings Into the Real Product

Demonstrating improvements in a research environment is only half the work. We extended the production version of the MoveAgain platform to support the new model types we had developed. This brought the performance gains into the hands of patients. We also supported the documentation required for FDA compliance, working across both the research codebase and the production software.

OUTCOMES

What shipped.

First FDA-approved implantable BCI platform supported

Superior cross-session decoder performance demonstrated

Reduced retraining time between sessions

Real-time in-session issue detection

Automated multi-session analytics platform

20+ concurrent behavioral detectors supported

Support for 2 participants with bi-monthly sessions

Python to Rust production implementation

HIPAA-compliant data handling

KEY TAKEAWAYS

What we learned.

  • Self-supervised pre-training improves BCI decoder robustness. By learning general neural activity representations before task-specific fine-tuning, models generalize better across sessions and reduce patient burden during calibration.
  • Real-time session monitoring prevents data loss. In-session tools that detect hardware issues, environmental noise, and setup errors immediately save costly clinical recordings that would otherwise be lost to post-hoc discovery.
  • Automated analytics scale clinical BCI research. As participant numbers grow, manual analysis becomes a bottleneck. Centralized data pipelines with automated quality metrics and cross-session dashboards enable faster clinical iteration.
  • Research-to-production translation requires dual expertise. Optimizing algorithms in Python research environments is only half the challenge, production deployment in Rust for embedded systems requires teams that span both domains.
  • Clinical neurotechnology demands rigorous data handling. FDA compliance, HIPAA requirements, and patient privacy necessitate secure pipelines and transparent collaboration between research and clinical teams.

IN SUMMARY

Bottom line.

In summary, AE Studio's partnership with Blackrock Neurotech demonstrates how applied machine learning and production software engineering can advance clinical brain-computer interfaces. As a result, by combining self-supervised learning techniques, real-time analysis tools, and automated analytics platforms, we helped extend the capabilities of the MoveAgain platform, bringing improved decoder performance and clinical usability to patients with paralysis. Furthermore, this collaboration exemplifies AE Studio's mission: applying cutting-edge AI and software development to neurotechnology that genuinely improves human lives.

FAQ

Frequently asked.

What is the MoveAgain platform and why is it significant?
MoveAgain is Blackrock Neurotech's commercial brain-computer interface platform designed to restore movement for people with paralysis. It's powered by the Utah Array, the only FDA-approved implantable BCI device on the market. The system enables patients with spinal cord injuries and other motor impairments to control cursors, communicate, and interact with devices using neural signals recorded directly from their motor cortex.
How did self-supervised learning improve the neural decoder performance?
Traditional supervised learning requires extensive labeled data from each session, creating a cold-start problem where patients must complete lengthy calibration procedures. Self-supervised pre-training learns general representations of neural activity patterns across subjects and sessions before fine-tuning on specific tasks. This approach yielded robust cross-session decoding because the model had already learned what neural signals 'look like' in general, making it faster and more accurate when adapting to a specific patient or session.
What types of issues did the in-session analysis tools detect?
The in-session tools identified several categories of problems that could compromise session data: environmental electrical noise that corrupts neural recordings, errors in experimental setup or hardware configuration, unexpected task-locked neural responses that indicate protocol issues, and signal quality degradation over the course of a session. By catching these issues in real time, clinical teams could make adjustments immediately rather than losing entire session recordings.
How did AE Studio ensure HIPAA compliance and data security for clinical neural data?
Clinical neurotechnology involves highly sensitive patient data requiring strict privacy protections. AE Studio implemented secure data pipelines that maintained HIPAA compliance throughout the analysis workflow. This included encrypted data transfer and storage, access controls limiting data visibility to authorized team members, and audit trails documenting data handling. The automated analytics platform was designed with clinical data confidentiality as a core requirement.
Why was Rust used for the production implementation instead of Python?
The research and algorithm development work was done in Python, which offers flexibility and access to machine learning libraries. However, the production MoveAgain platform runs on embedded systems where performance, memory safety, and reliability are critical. Rust provides memory safety without garbage collection, predictable performance characteristics, and the low-level control needed for real-time BCI applications. AE Studio's team bridged both domains, translating research findings into production Rust code.
What is the Neural Data Transformer (NDT) and how was it optimized?
The Neural Data Transformer is a deep learning architecture designed for neural signal decoding. It processes time-series neural data to predict intended movements or actions. AE Studio optimized the NDT through systematic hyperparameter tuning for both pre-training and fine-tuning phases, investigated alternative self-supervised learning approaches like contrastive learning, and developed model stabilization methods to improve robustness across sessions and reduce training time.
How does this work relate to AE Studio's broader BCI expertise?
The Blackrock collaboration is part of AE Studio's extensive neurotechnology portfolio. The team has won first place in every category of the Neural Latents Benchmark Challenge, contributed to open-source tools like the Neurotech Development Kit and Neural Data Simulator, and works with multiple BCI companies including Forest Neurotech and clinical partners. AE Studio's models achieved top rankings in international neural decoding competitions, demonstrating technical leadership in translating brain signals into actionable outputs.
What was the typical development timeline for the different project components?
The project spanned multiple statements of work with varying timelines by component. Decoder development tasks like model pre-training and hyperparameter tuning took 4-7 weeks each. Research session support required approximately 1.5 weeks per session for signal analysis, decoding analysis, and model release. The automated multi-session analysis platform required 8-11 weeks for centralized storage, automated analysis, and dashboard visualization. In-session tool development and MoveAgain software integration were ongoing efforts.

LET'S TALK

Bring us the hard problem.

We'll bring the team that ships.

Book a call Back to home
Get in touch