پایدارسازی بدون نظارت منیفولدهای عصبی برای رمزگشایی حرکت برای کاربردهای واسط مغز-رایانه
کد: G-1838
نویسندگان: Mohammadali Ganjali ℗, Alireza Mehridehnavi, Abed Khorasani *
زمان بندی: زمان بندی نشده!
برچسب: پردازش سیگنال های پزشکی
دانلود: دانلود پوستر
خلاصه مقاله:
خلاصه مقاله
Background and aims: Brain-computer interfaces (BCIs) aim to convert brain signals into commands to control external devices for assisting individuals with spinal cord injuries (SCIs). However, a challenge in real-world BCI applications is the changes in neural activity over time, which is needed for day-to-day calibration. One solution to overcome the need for calibration is projecting neural data onto low-dimensional manifolds and aligning them across sessions using alignment methods such as canonical correlation analysis (CCA). However, CCA requires the actual trajectory of subject movements, such as target labels, which is often not feasible for real-world applications. Method: In this study, an automatic algorithm named unsupervised neural manifold alignment decoding (UnMAD) is proposed to decode movement parameters from neural activity using aligned manifolds without the need for actual target labels. UnMAD integrates three main stages: (1) Dimensionality Reduction for extracting manifolds, (2) Discrete Trajectory Decoding for predicting target labels, and (3) Continuous Movement Decoding for aligning and decoding. The primary goal is to decode 2D velocity from the neural activity of the primary motor cortex of two monkeys (Monkey C and Monkey M) during a reach center-out task. Results: Results show that UnMAD compensated the variation between manifolds across two different recording sessions in two monkeys, improving the average correlation from R=0.47 before UnMAD to R=0.97 after UnMAD. The decoding performance results demonstrated that UnMAD outperformed the unsupervised distribution alignment decoding (DAD) approach. Also, UnMAD achieved 84% of the decoding performance compared to the CCA supervised method, with an average R-squared of 0.65 for UnMAD and 0.77 for CCA. Conclusion: This study suggests UnMAD, an unsupervised manifold stabilization method for decoding movement parameters. Unlike other alignment methods, UnMAD does not need true target labels, making it suitable for clinical applications such as subjects with SCIs.
کلمات کلیدی
Brain-Computer Interfaces, Neural Manifold, Movement Decoding
