Decoding Hand Kinematics from Local Field Potentials Using Long Short-Term Memory (LSTM) Network

Abstract

Local field potential (LFP) has gained increasing interest as an alternative input signal for brain-machine interfaces (BMIs) due to its informative features, long-term stability, and low frequency content. However, despite these interesting properties, LFP-based BMIs have been reported to yield low decoding performances compared to spike-based BMIs. In this paper, we propose a new decoder based on long short-term memory (LSTM) network which aims to improve the decoding performance of LFP-based BMIs. We compare offline decoding performance of the proposed LSTM decoder to a commonly used Kalman filter (KF) decoder on hand kinematics prediction tasks from multichannel LFPs. We also benchmark the performance of LFP-driven LSTM decoder against KF decoder driven by two types of spike signals: singleunit activity (SUA) and multi-unit activity (MUA). Our results show that LFP-driven LSTM decoder achieves significantly better decoding performance than LFP-, SUA-, and MUAdriven KF decoders. This suggests that LFPs coupled with LSTM decoder could provide high decoding performance, robust, and low power BMIs.

Decoding Kinematic Information From Primary Motor Cortex Ensemble Activities Using a Deep Canonical Correlation Analysis

“…In addition to θ, we also need to optimize the hyperparameters of DNNs, for which we employed the Bayesian optimization method (Ahmadi et al, 2019). To optimize the hyperparameters, we empirically preset the range for each parameter: the number of nodes in a layer ∈ {2 4 , 2 5 , .…”

“…The first category is a generative method that operates based on the generation of neuronal firing activities from kinematic states described by encoding models. The second category is a direct method that operates based on a direct input-output function approximation from neuronal firing activities to kinematic variables (Chapin et al, 1999;Sussillo et al, 2012;Dethier et al, 2013;Ahmadi et al, 2019).…”

“…Furthermore, owing to recent breakthroughs in deep learning, using deep neural networks (DNNs) for decoding M1 activities has become plausible (Sussillo et al, 2012;Ahmadi et al, 2019). For instance, a long short-term memory RNN (LSTM-RNN), one of the non-linear models harnessing temporal information in past neural activities, outperformed other decoding models for BMIs (Ahmadi et al, 2019). Despite their high performance, the intricate architectures of DNNs often require a much larger training data to achieve a successful decoding process.…”

Impact of referencing scheme on decoding performance of LFP-based brain-machine interface

“…For each LFP channel, six different features were extracted (one feature in the time domain and five features in the frequency domain). The time-domain feature was an average amplitude called local motor potential (LMP) whereas the frequency domain features were average spectral power in five different frequency bands: delta (1-4 Hz), theta (3-10 Hz), alpha (12-23 Hz), beta (27)(28)(29)(30)(31)(32)(33)(34)(35)(36)(37)(38) and gamma (50-300 Hz). The selection of these frequency bands followed a previous study by Stavisky et al [2].…”

“…LSTMs have demonstrated state-of-the-art performances in various applications such as language modelling and translation, speech recognition and synthesis, and video classification and captioning [29]. Our previous study has also shown that LSTM can achieve higher decoding performance when compared to a Kalman filter [30]. LSTM successfully addresses the vanishing gradient problem commonly encountered in traditional recurrent neural networks (RNNs) and is capable of learning long-term temporal dependencies.…”