New AI model improves brain-computer interface control for ALS patients

Researchers from University College London have developed a new AI-based brain signal decoding model that could improve how people with types of motor neurone diseases like amyotrophic lateral sclerosis (ALS) use brain-computer interfaces (BCIs) to predict their thoughts. Published in Neuroelectronics, the study introduces a graph attention network (GAT) that makes it easier to interpret brain signals, leading to more accurate and reliable BCI performance, achieving on average 74.06% accuracy when classifying Left and Right Hand movement intention. This improvement could help speed up the development of better assistive devices, enhancing independence and quality of life for ALS patients in a real world scenario.

Classifying brain signals in people with ALS has been a long-standing challenge due to how brain activity changes between individuals and across sessions. People with ALS lose the ability to move for reasons that remain unclear, yet they still have movement intentions—signals in the brain that reflect what they want to do. BCIs can detect these signals and translate them into actions but accurately decoding them remains difficult.

Researchers have developed a new AI model that significantly improves accuracy in decoding these signals, taking a step toward more usable and accessible BCIs. Since brain signals evolve over time, the model is designed to learn many patterns, ensuring it remains effective as users improve or decline. It accounts for plastic changes in the brain, following the principle that "neurons that fire together, wire together"—a core idea in Hebbian learning. This allows the system to maintain signal integrity while adjusting to shifts in neural activity, making BCIs more reliable over time.

This approach leverages a specialized neural network called a graph attention network (GAT) to better understand how different brain regions communicate. By analyzing brain wave patterns using phase synchrony, the model outperforms traditional deep learning methods, making it easier for people with ALS to control assistive technologies with their thoughts.

A major challenge in BCIs is that many users struggle to achieve a reliable level of accuracy. Standard models often fail to capture the complexity of brain activity, preventing about 40% of users from reaching 70% accuracy, a key threshold for effective BCI use. The new GAT model addresses this by adapting to each user’s unique brain patterns, leading to more consistent and personalized results.

“This study highlights the potential of graph-based models in helping ALS patients interact with technology more effectively” said Rishan Patel, lead author of the study. “By using advanced brain signal representations, we can make BCIs more reliable and accessible to all patients.”

Dr. Dai Jiang, a senior author of this study from the Bioelectronics Group at University College London, said, “While achieving effective personalization, this model prioritizes efficiency and portability, making it well-suited for edge computing and real-world BCI applications.”

In testing, the model demonstrated impressive performance, achieving 74.06% accuracy on an ALS dataset collected over 1–2 months and 71.89% accuracy on a widely used benchmark dataset. This marks a significant improvement over existing methods, showing that graph-based AI models can better handle the complexities of neurodegenerative conditions like ALS, especially in patients who typically do not perform as well as others.

Prof. Tom Carlson, an author of the study, Professor of Assistive Robotics at UCL and leading the Aspire CREATe research lab said, “Typically, BCIs require a period of re-calibration every time they are used, which rather limits their real-world applications. By contrast, this approach is much more resilient to changes in people’s brain signals and will pave the way towards a more practical, user-friendly solution”.

Prof. Andreas Demosthenous, an author of the study, Professor of Analogue & Biomedical Electronics and leader of the Bioelectronics Group said, “This work has great potential of providing a reliable and low-cost solution that can improve the quality of life of patients suffering from neurodegenerative conditions”.

Future work will focus on conducting tests to further evaluate the model's robustness. The team aims to validate its effectiveness and demonstrate its superiority over current state-of-the-art methods with sufficient explainability, ensuring reliable adaptation to dynamic real-world conditions, with the ultimate goal of advancing towards more effective online scenarios.

This work is supported by the Engineering and Physical Sciences Research Council (EPSRC). For more information, please contact: Rishan Patel, PhD Candidate uceerjp@ucl.ac.uk.

Biography
Amyotrophic Lateral Sclerosis is a progressive, degenerative disease impacting around 200,000 individuals worldwide. It leads to paralysis and frequently results in death due to respiratory failure within 2-5 years of onset. Brain Computer Interfaces allow the translation of thought/imagined movement (MI) into real actions like controlling a computer. ALS is often neglected in coadaptive algorithm development for BCIs due to the challenge of disease progression. This non-stationarity decreases accuracy and creates barriers for the adoption of this technology. During his PhD, he works on the resolution of this issue through novel machine learning methods which can help bridge the gap between patients and BCIs. 
Rishan is funded by the Institute of Healthcare Engineering Doctoral Training Programme (IHE-DTP) at University College London and belongs to the Bioelectronics and Aspire CREATe Labs.

Publication:

R. Patel, Z. Zhu, B. Bryson, T. Carlson, D. Jiang, and A. Demosthenous. Advancing EEG classification for neurodegenerative conditions using BCI: A graph attention approach with phase synchrony. Neuroelectronics, 2025, 2(1), DOI: 10.55092/neuroelectronics20250001