Simple Algorithm Boosts Brain-Machine Interface Success

By Soham NandiJan 22 2024Reviewed by Susha Cheriyedath, M.Sc.

In an article published in the journal Plos One, researchers explored brain-machine interfaces (BMI) that connected brains to external devices, focusing on the bio-feedback control approach. The study implemented a firing-rate-to-motion rule, assigning neuron groups to control a cursor. 

Background

BMIs linking brains to machines have evolved, particularly using intracortical neuronal recordings from implanted electrodes. While previous work focused on bio-mimetic approaches with complex algorithms decoding neural activity for intuitive control, this study addressed the gap in leveraging user learning for simpler BMI control paradigms. With chronic recordings becoming more stable, the bio-feedback approach became feasible, allowing users to learn to control novel appendages without complex neural tuning.

The research introduced a group weight algorithm, summing spike counts within neuron groups for cursor control. Unlike bio-mimetic approaches, it required minimal parameter estimation, emphasizing user learning. The study involved two Rhesus monkeys learning to control a two-dimensional (2D) cursor using this bio-feedback method. The group weight method introduced a simpler control paradigm, highlighting the need for further investigation into its application in BMI design and its utility in studying neural changes during learning.

Materials and methodology 

The authors introduced a group weight algorithm for BMIs that enabled 2D cursor control by converting neuron firing rates into cursor velocity. The algorithm divided neurons into four groups corresponding to up, down, left, and right movements, facilitating intuitive control. The conversion process involved summing and normalizing firing rates within each group. Neurons were grouped based on preferred directions (PD) obtained from a linear regression model.

Stability criteria ensured long-term stable recordings and the normalization process addressed changes in firing rates due to recording instability. The study involved two Rhesus monkeys with Utah microelectrode arrays implanted in the primary motor cortex. The monkeys learned to control a cursor via the group weight BMI paradigm, with significant improvements observed over approximately one week of practice.

The algorithm's simplicity, leveraging user learning without complex parameter estimation, distinguished it from bio-mimetic approaches. A key element was the grouping of neurons based on PDs, aiding initial control. The researchers addressed the potential challenges of long-term bio-mimetic BMI control by proposing a feasible bio-feedback paradigm.

Neuronal tuning properties, trajectory analysis, and learning patterns provided insights into the algorithm's effectiveness. The normalization method ensured a balance between opposing groups and accommodated variations in firing rates. The monkeys successfully performed center-out reaching tasks, demonstrating the algorithm's practical utility. The authors emphasized the potential of user-friendly BMI designs that prioritize simplicity and user adaptability, laying the groundwork for further exploration in BMI development and user learning in real-world applications.

Results 

Researchers demonstrated the successful implementation of a group weight algorithm in BMIs for 2D cursor control using Utah microelectrode arrays implanted in the primary motor cortex of Rhesus monkeys. Over a week of training, both monkeys exhibited steady improvements in task performance, with the number of successfully finished trials and success rates significantly increasing. The success rates surpassed random baseline values, indicating that the monkeys actively participated in and learned the group weight control paradigm. Trajectory analysis revealed evolving patterns, with cursor trajectories becoming more refined and directional success rates increasing over the learning period.

Moreover, the researchers explored neural ensemble activity and revealed that the monkeys' control quality improved. Output-potent values, crucial for effective cursor movement, increased significantly, reflecting enhanced group activity effectiveness. Tuning analysis of individual neurons showed an increase in tuning quality during the learning process.

Additionally, the PD of neurons shifted toward their assigned directions, aligning more closely with the desired cursor movement. While the authors emphasized the suitability of the group weight algorithm for BMI control, they also provided insights into the neural changes associated with learning and adaptation in the context of motor tasks. Overall, the results underscored the efficacy of the proposed BMI paradigm and offered valuable insights into the mechanisms underlying neural learning in the context of cursor control tasks.

Discussion

The authors demonstrated monkeys' learning of a bio-feedback BMI, showing a 70% improvement in trial success rate over a week. Utilizing a simple group weight algorithm for 2D cursor control, monkeys successfully adapted neural ensembles for improved performance. Comparisons with related studies highlighted the unique aspects of the proposed method, focusing on group-based control with orthogonal contribution directions.

The researchers emphasized bio-feedback learning over bio-mimetic decoding, and while recording instability and short training periods present limitations, the work served as a feasibility demonstration for future refinement of multi-dimensional control methods in BMIs.

Conclusion

In conclusion, researchers introduced a bio-feedback BMI paradigm using a simple group weight algorithm for 2D cursor control, demonstrating a 70% improvement in trial success rate over a week of training in Rhesus monkeys. The algorithm leveraged user learning without complex parameter estimation, emphasizing simplicity and adaptability. Results highlighted the efficacy of the approach, paving the way for further research in user-friendly BMI designs and providing valuable insights into neural changes during motor learning tasks.