Using a Long Short-Term Memory Neural Network for Generation of Control Signals From Human μECoG Data
Uloženo v:
| Vydáno v: | ProQuest Dissertations and Theses (2025) |
|---|---|
| Hlavní autor: | |
| Vydáno: |
ProQuest Dissertations & Theses
|
| Témata: | |
| On-line přístup: | Citation/Abstract Full Text - PDF |
| Tagy: |
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
|
| Abstrakt: | Objective. Brain-computer interfaces (BCIs) based on Micro-electrocorticography (µECoG) offer a promising approach for decoding motor intentions and controlling assistive technologies. While intraparenchymal BCIs have demonstrated robust control signals for communication and motor prostheses, µECoG provides a balance between invasiveness, signal fidelity, and long-term stability, making it a viable option for real-time neural interfaces. However, challenges remain in translating µECoG signals into reliable control outputs for real-world applications. This study examines the performance of Long Short-Term Memory (LSTM) networks in decoding µECoG signals for motor control, with a focus on generalization across temporal structures of the time-series data. Approach. µECoG signals were recorded from a patient performing reaching movements toward multiple targets using two 16-channel grids implanted over the primary motor cortex. Preprocessing included Common Average Referencing (CAR), and low pass filtering to reduce noise. LSTM networks were trained and evaluated under two conditions: (1) interpolated data, where temporal relationships were not preserved, and (2) extrapolated data, where the model was tested on sequentially structured time-series data to assess generalization. Additionally, a parametric evaluation was done on varying the working memory variable (length of the temporal history of the input data) for each recording session. Main Results. LSTM models achieved over 95% validation accuracy in predicting movement trajectories when trained and tested on interpolated data (time sequence shuffled). However, accuracy dropped to approximately 84% for extrapolated data (time sequence preserved). The validation accuracy for each session tended to be highs when the working memory was 1 – 250ms and then decrease with working memory was >250ms. Significance. These findings highlight the real-world challenges of generalizing across time and suggest that LSTM architectures may struggle to capture temporal dependencies in µECoG signals. We could also find that the decoding was better when patient performed movements away from the body v/s towards the body and that working memory values in the LSTM corresponded with the duration of working memory signals observed in the cerebral cortex. These findings show that matching the spatial topology of BCI electrodes and the temporal depth of the sampling of neural signal should be guided by the known spatial and temporal structure of neural signal in the human cerebral cortex, and can be used to guide the implementation of AI models. |
|---|---|
| ISBN: | 9798314876169 |
| Zdroj: | ProQuest Dissertations & Theses Global |