A Characterization of Brain-Computer Interface Performance Trade-Offs Using Support Vector Machines and Deep Neural...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Rehabilitation Technology

Background:

• Brain-computer interface (BCI) systems offer potential for individuals with paralysis by translating neural activity into device control.
• Key performance criteria for clinical BCI translation include accuracy, response latency, and multi-functionality.
• Optimizing these criteria is challenging due to inherent trade-offs in neural decoder design.

Purpose of the Study:

• To systematically explore the trade-offs between accuracy, response latency, and multi-functionality in BCI decoders.
• To compare the performance of a support vector machine (SVM) classifier with a proposed deep neural network (DNN) framework.
• To quantify the impact of decoder design on BCI performance for discrete movement classification.

Main Methods:

• Utilized historical intracortical recordings from a tetraplegic participant imagining hand and finger movements.
• Implemented and compared SVM and DNN decoding strategies for neural activity.
• Evaluated real-time control of functional electrical stimulation (FES) using the DNN decoder.

Main Results:

• Both SVM and DNN decoders showed increased response times and decreased accuracy with increased functionality.
• The DNN framework demonstrated less degradation in response time and accuracy compared to SVM as functionality increased.
• Data preprocessing significantly impacted decoder performance differently for SVM and DNN.

Conclusions:

• Deep neural networks offer improved performance characteristics for BCI decoders compared to traditional SVMs, particularly regarding multi-functionality.
• Understanding and quantifying decoder trade-offs is crucial for developing clinically viable BCI systems.
• Real-time evaluation and comparison with able-bodied performance establish benchmarks for BCI-FES systems.