Brain-machine interface based on transfer-learning for detecting the appearance of obstacles during...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Brain-machine interfaces (BMIs) face challenges in real-world applications due to signal artifacts, large data needs, and non-stationarity.
• Classical signal processing techniques often fall short for real-time EEG-based control.
• Deep learning offers potential solutions for these complex BMI challenges.

Purpose of the Study:

• To develop a robust BMI capable of detecting the evoked potential associated with the intention to stop when encountering an unexpected obstacle.
• To improve the accuracy and reliability of real-time obstacle detection for BMI-controlled systems.

Main Methods:

• Utilized a two-consecutive convolutional neural network architecture for intention detection and false detection correction.
• Tested the interface on a treadmill with healthy subjects, simulating obstacle detection.
• Implemented a closed-loop experiment with an exoskeleton for real-time stop commands.
• Applied transfer learning techniques for feasibility in patients with incomplete Spinal Cord Injury (iSCI).

Main Results:

• The two-network system significantly outperformed a single network, reducing false positives per minute (FP/min) from 31.8 to 3.9 and improving no false positives/true positives (NOFP/TP) from 34.9% to 60.3% in pseudo-online analysis.
• In closed-loop exoskeleton experiments, the system achieved 3.8 FP/min and 49.3% NOFP/TP in healthy subjects.
• For iSCI patients, the system demonstrated promising results with 7.7 FP/min and 37.9% NOFP/TP.

Conclusions:

• A deep learning-based BMI with a two-network approach effectively detects obstacle-avoidance intentions, enhancing real-time control.
• The developed methodology shows significant improvements in accuracy and reliability compared to single-network approaches.
• The system's adaptability through transfer learning makes it a viable tool for assisting individuals with mobility impairments, including those with iSCI.