PreprintPDF Available

Deep Learning-Based Approaches for Decoding Motor Intent from Peripheral Nerve Signals

Authors:

Abstract and Figures

The ultimate goal of an upper-limb neuroprosthesis is to achieve dexterous and intuitive control of individual fingers. Previous literature shows that deep learning (DL) is an effective tool to decode the motor intent from neural signals obtained from different parts of the nervous system. However, it still requires complicated deep neural networks that are inefficient and not feasible to work in real-time. Here we investigate different approaches to enhance the efficiency of the DL-based motor decoding paradigm. First, a comprehensive collection of feature extraction techniques is applied to reduce the input data dimensionality. Next, we investigate two different strategies for deploying DL models: a one-step (1S) approach when big input data are available and a two-step (2S) when input data are limited. With the 1S approach, a single regression stage predicts the trajectories of all fingers. With the 2S approach, a classification stage identifies the fingers in motion, followed by a regression stage that predicts those active digits' trajectories. The addition of feature extraction substantially lowers the motor decoder's complexity, making it feasible for translation to a real-time paradigm. The 1S approach using a recurrent neural network (RNN) generally gives better prediction results than all the classic machine learning (ML) algorithms with mean squared error (MSE) ranges from 10 ⁻³ to 10 ⁻⁴ for all finger while variance accounted for (VAF) scores are above 0.8 for the most degree of freedom (DOF). This result reaffirms that DL is more advantageous than classic ML methods for handling a large dataset. However, when training on a smaller input data set as in the 2S approach, ML techniques offers a simpler implementation while ensuring comparably good decoding outcome to the DL ones. In the classification step, either ML or DL models achieve the accuracy and F1 score of 0.99. Thanks to the classification step, in the regression step, both types of models result in comparable MSE and VAF scores as those of the 1S approach. Our study outlines the trade-offs to inform the implementation of real-time, low-latency, and high accuracy DL-based motor decoder for clinical applications.
Content may be subject to copyright.
Deep Learning-Based Approaches for Decoding Motor Intent from
Peripheral Nerve Signals
Diu Khue Luu1,2, Anh Tuan Nguyen1,2,6, Ming Jiang3, Jian Xu2,
Markus W. Drealan2, Jonathan Cheng4,5, Edward W. Keefer5,6, Qi Zhao3and Zhi Yang2,6
1Co-first authors
2Biomedical Engineering, University of Minnesota, Minneapolis, MN, USA
3Computer Science and Engineering, University of Minnesota, Minneapolis, MN, USA
4Plastic Surgery, University of Texas Southwestern Medical Center, Dallas, TX, USA
5Nerves Incorporated, Dallas, TX, USA
6Fasikl Incorporated, Minneapolis, MN, USA
Correspondence: Diu Khue Luu (Email: luu00009@umn.edu)
Abstract
The ultimate goal of an upper-limb neuroprosthesis is to achieve dexterous and intuitive
control of individual fingers. Previous literature shows that deep learning (DL) is an effective
tool to decode the motor intent from neural signals obtained from different parts of the nervous
system. However, it still requires complicated deep neural networks that are inefficient and not
feasible to work in real-time. Here we investigate different approaches to enhance the efficiency of
the DL-based motor decoding paradigm. First, a comprehensive collection of feature extraction
techniques is applied to reduce the input data dimensionality. Next, we investigate two different
strategies for deploying DL models: a one-step (1S) approach when big input data are available
and a two-step (2S) when input data are limited. With the 1S approach, a single regression stage
predicts the trajectories of all fingers. With the 2S approach, a classification stage identifies the
fingers in motion, followed by a regression stage that predicts those active digits’ trajectories.
The addition of feature extraction substantially lowers the motor decoder’s complexity, making
it feasible for translation to a real-time paradigm. The 1S approach using a recurrent neural
network (RNN) generally gives better prediction results than all the ML algorithms with mean
squared error (MSE) ranges from 103to 104for all finger while variance accounted for (VAF)
scores are above 0.8 for the most degree of freedom (DOF). This result reaffirms that DL is more
advantageous than classic ML methods for handling a large dataset. However, when training on
a smaller input data set as in the 2S approach, ML techniques offers a simpler implementation
while ensuring comparably good decoding outcome to the DL ones. In the classification step,
either machine-learning (ML) or DL models achieve the accuracy and F1 score of 0.99. Thanks
to the classification step, in the regression step, both types of models result in comparable MSE
and VAF scores as those of the 1S approach. Our study outlines the trade-offs to inform the
implementation of real-time, low-latency, and high accuracy DL-based motor decoder for clinical
applications.
Keywords: convolutional neural network, deep learning, feature extraction, motor decoding,
neuroprosthesis, neural decoder, peripheral nerve interface, recurrent neural network.
1
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted February 19, 2021. ; https://doi.org/10.1101/2021.02.18.431483doi: bioRxiv preprint
1 Introduction
Upper-limb amputation affects the quality of life and well-being of millions of people in the United
States, with hundreds of thousand new cases annually (Ziegler-Graham, 2008). Neuroprosthetic
systems promise the ultimate solution by developing human-machine interfaces (HMI) that could
allow amputees to control robotic limbs using their thoughts (Schultz, 2011; Cordella, 2016). It is
achieved by decoding the subject’s motor intent with neural data acquired from different parts of
the nervous system. Proven approaches include surface electromyogram (EMG) (Sebelius, 2005;
Jiang, 2012; Fougner, 2012; Amsuss, 2013; Zuleta, 2019), electroencephalogram (EEG) (Hu, 2015;
Zeng, 2015; Sakhavi, 2018; Kwon, 2019), cortical recordings (Mollazadeh, 2011; Hochberg, 2012;
Irwin, 2017), and peripheral nerve recordings (Micera, 2011; Davis, 2016; Vu, 2017; Wendelken,
2017; Zhang, 2017; Nguyen & Xu, 2020).
However, implementing an effective HMI for neuroprostheses remains a challenging task. The
decoder should be able to predict the subject’s motor intents accurately and satisfy certain criteria
to make it practical and useful in daily lives. Some criteria include dexterity, i.e. controlling
multiple degrees-of-freedom (DOF) such as individual fingers; intuitiveness, i.e. reflecting the true
motor intent in mind, and real-time, i.e. having minimal latency from thoughts to movements.
In recent years, deep learning (DL) techniques have emerged as strong candidates to overcome
this challenge thanks to their ability to process and analyze biological big data (Mahmud, 2018).
Our previous work (Nguyen & Xu, 2020) shows that neural decoders based on the convolutional
neural network (CNN) and recurrent neural network (RNN) architecture outperform other “classic”
machine learning counterparts in decoding motor intents from peripheral nerve data obtained with
an implantable bioelectric neural interface. The DL-based motor decoders can regress the intended
motion of fifteen degrees-of-freedom (DOF) simultaneously, including flexion/extension and abduc-
tion/adduction of individual fingers state-of-the-art performance metrics, thus complying with the
dexterity and intuitiveness criteria.
Here we build upon the foundation of (Nguyen & Xu, 2020) by exploring different strategies
to optimize the motor decoding paradigm’s efficiency. The aim is to lower the neural decoder’s
computational complexity while retaining high accuracy predictions to make it feasible to translate
the motor decoding paradigm to real-time operation suitable for clinical applications, especially
when deploying in a portable platform.
First, we utilize feature extraction to reduce data dimensionality. By examining the data
spectrogram, we learn that most of the signals’ power concentrates in the frequency band 25-
600 Hz (Nguyen & Xu, 2020). Many feature extraction techniques (Zardoshti-Kermani, 1995;
Phinyomark, 2009; Rafiee, 2011; Phinyomark, 2012) have been developed to handle signals with
similar characteristics. Feature extraction aims not only to amplify the crucial information, lessen
the noise but also to substantially reduce the data dimensionality before feeding them to DL models.
This could simultaneously enhance the prediction accuracy and lower the DL models’ complexity.
Here we focus on a comprehensive list of fourteen features that consistently appear in the field of
neuroprosthesis.
Second, we explore two different strategies for deploying DL models: the two-step (2S) and the
one-step (1S) approaches. The 2S approach consists of a classification stage to identify the active
fingers and a regression stage to predict the trajectories of digits in motion. The 1S approach only
has one regression stage to predict the trajectories of all fingers concurrently. In practice, the 2S
approach should be marginal more efficient because not all models are inferred at a given moment.
The models in the 2S approach are only trained on the subset where a particular is active, while all
models in the 1S approach are trained on the full dataset. Here we focus on exploring the trade-offs
between two approaches to inform future decisions of implementing the DL-based motor decoder
2
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted February 19, 2021. ; https://doi.org/10.1101/2021.02.18.431483doi: bioRxiv preprint
in real-world applications.
The rest of this paper is organized as follows: section ”Data Description” introduces the human
participant of this research, the process of collecting input neural signals from the residual periph-
eral nerves of the participant, and establishing the ground-truth for the motor decoding paradigm
using DL models. Section ”Data Preprocessing” elaborates on how to cut raw input neural data
into trials and extract their main features in the temporal domain before feeding to DL decoding
models. Section ”Proposed Methods and Deep Learning Models for Motor Intent Decoding” dis-
cusses the two approaches to efficiently translate motor intent from the residual peripheral nerves
of the participant into motor control of the prosthesis as well as the architecture and the hyper-
parameters of the DL models used in each approach. Section ”Experimental Setup” is about the
three ML models used as the baseline and how to input neural data are allocated to the training
and validation set. Section ”Metrics and Results” presents the metrics to measure the performance
of all models used in the motor intent decoding process and discusses the main results of both
proposed approaches. Section ”Discussion” discusses the role of feature extraction in reducing the
DL motor decoders’ complexity for real-time applications, how to further apply it in future works,
and the advantages of ML and DL motor decoders in different scenarios where input dataset’s size
varies. Finally, section ”Conclusion” is to summarize the main contributions of this paper.
2 Data Description
2.1 Human Participant
Figure 1: Photo of the (A) amputee and (B) the data collection software during a training session.
The patient performs various hand movements repeatedly during the training session. Nerve data
and ground-truth movements are collected by a computer and displayed in real-time on the monitor
for comparison.
The participant is a transradial male amputee who has lost his hand for over five years (Fig. 1).
Among seven levels of upper-limb amputation, transradial is the most common type that accounts
for about 57% of upper-limb loss in the U.S.(Schultz, 2011; Cordella, 2016). Like most amputees,
the subject still has phantom limb movements; however, such phantom feelings fade away over
3
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted February 19, 2021. ; https://doi.org/10.1101/2021.02.18.431483doi: bioRxiv preprint
time. By successfully decoding neural signals from the residual nerves of an amputee who has lost
his limb for a long time, we would offer a chance to regain upper-limb motor control for those who
are sharing the same conditions.
The human experiment protocols are reviewed and approved by the Institutional Review Board
(IRB) at the University of Minnesota (UMN) and the University of Texas Southwestern Medical
Center (UTSW). The patient voluntarily participates in our study and is informed of the methods,
aims, benefits, and potential risks of the experiments before signing the Informed Consent. Patient
safety and data privacy are overseen by the Data and Safety Monitoring Committee (DSMC) at
UTSW. The implantation, initial testing, and post-operative care are performed at UTSW by Dr.
Cheng and Dr. Keefer, while motor decoding experiments are performed at UMN by Dr. Yang’s
lab. The clinical team travels with the patient in each experiment session.
Figure 2: (A) Overview of the human experiment setup and data acquisition using the mirrored
bilateral training. The patient has four FAST-LIFE microelectrode arrays implanted in the residual
ulnar and median nerve (Overstreet, 2019). Peripheral nerve signals are acquired by two Scorpius
neural interface devices (Nguyen & Xu, 2020). The ground-truth movements are obtained with a
data glove. (B) Neural data are cut using a sliding window to resemble online decoding.
The patient undergoes an implant surgery where four longitudinal intrafascicular electrode
(LIFE) arrays are inserted into the residual median and ulnar nerves using the microsurgical fas-
cicular targeting (FAST) technique (Fig. 2(A)). The electrode array’ design, characteristics, and
surgical procedures are reported in (Cheng, 2017; Overstreet, 2019). The patient has the electrode
arrays implanted for 12 months, during which the conditions of the implantation site is regularly
monitored for signs of degradation.
The patient participates in several neural stimulations, neural recording, and motor decoding
experiment sessions. He initially has weak phantom limb movements due to reduced motor control
signals in the residual nerves throughout the years. However, the patient reports that the more
experiment sessions he takes part in, the stronger his phantom control and sensation of the lost
hand become. This suggests that training may help re-establish the connection between the motor
cortex and the residual nerves, resulting in better motor control signals.
4
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted February 19, 2021. ; https://doi.org/10.1101/2021.02.18.431483doi: bioRxiv preprint
2.2 Nerve Data Acquisition
Nerve signals are acquired using the Scorpius neural interface (Fig. 2(A)) - a miniaturized, high-
performance neural recording system developed by Yang’s lab at UMN. The system employs the
Neuronix chip family, which consists of fully-integrated neural recorders designed based on the fre-
quency shaping (FS) architecture (Xu, 2014; Yang, 2016, 2018, 2020; Xu, 2020). The specifications
of the Scorpius system are reported in (Nguyen & Xu, 2020). The system allows acquiring nerve
signals with high-fidelity while suppressing artifacts and interference. Here two Scorpius devices
are used to acquire signals from 16 channels across four microelectrode arrays at a sampling rate
of 40 kHz (7.68 Mbps, 480 kbps per channel). The data are further downsampled to 5 kHz before
applying a bandpass filter in 25-600 Hz bandwidth to capture most of the signals’ power. This
results in a pre-processed data stream of 1.28 Mbps (80 kbps per channel).
2.3 Ground-truth Collection
The mirrored bilateral training paradigm (Sebelius, 2005; Jiang, 2012) is used to establish the
ground-truth labels needed for supervised learning (Fig. 2(A)). The patient performs various hand
gestures with the able hand while simultaneously imagining doing the same movement with the
phantom/injured hand. The gestures include bending the thumb, index, middle, ring, little finger,
index pinch, tripod pinch, and grasp/fist. Each gesture is repeated 100 times, altering between
resting and flexing. Peripheral nerve signals are acquired from the injured hand with the Scorpius
system, while ground-truth movements are captured with a data glove (VMG, 30, Virtual Motion
Labs, TX) from the able hand. The glove can acquire up to 15 DOF; however, we only focus on
the main 10 DOF (MCP and PIP) corresponding to the flexion/extension of five fingers.
3 Data Preprocessing
3.1 Cutting Raw Neural Data
In this paper, raw neural data are cut using a sliding window to resemble online motor decoding
(Fig. 2(B)). Here the window’s length is set to 4 sec with an incremental step of 100 msec. At
any instant of time, the decoder can only observe the past neural data. The pseudo-online dataset
contains overlapping windows from a total of 50.7 min worth of neural recordings. Each of these
4-sec neural data segments serves as an input trial of the motor decoding process later.
3.2 Feature Extraction
Previous studies have shown that feature extraction is an effective gateway to achieve optimal
classification performance with signals in the low-frequency band by highlighting critical hidden
information while rejecting unwanted noise and interference. Here we select fourteen of the most
simple and robust features that are frequently used in previous motor decoding studies (Zardoshti-
Kermani, 1995; Phinyomark, 2009; Rafiee, 2011; Phinyomark, 2012). They are chosen such that
there is no linear relationship between any pair of features. All features can be computed in
the temporal domain with relatively simple arithmetic, thus aiding the implementation in future
portable systems.
Table 1 summaries the descriptions and formula of the features. xiis the 4-sec neural data
segments, which are further divided into windows of 100 milliseconds with Nis the window length.
Two consecutive windows are 80% overlapped, which is equivalent to a 20 milliseconds time step.
This results in a data stream of 224 features over 16 channels with a data rate of 179.2 kbps (11.2
5
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted February 19, 2021. ; https://doi.org/10.1101/2021.02.18.431483doi: bioRxiv preprint
Table 1: List of features, descriptions, and formula
Feature Description Formula
F1 Zero crossing
(ZC)
The number of times the demeaned
data change sign.
ΣN
i=2sgn(xi1xi)
F2 Slope sign
changes (SSC)
The number of times the differential
data change sign.
ΣN
i=3sgn[(xixi1)(xi1xi2)]
F3 Waveform
length (WL)
The summation of the absolute values
of the differential data.
ΣN
i=2|xixi1|
F4 Wilson
amplitude
(WA)
The number of times the change in the
signal amplitudes of two consecutive
samples exceeds the standard devia-
tion.
ΣN
i=2sgn(|xixi1| − xstd)
F5 Mean absolute
(MAB)
The average of the absolute values of
the data.
1
NΣN
i=1|xi|
F6 Mean square
(MSQ)
The average of the square values of the
data.
1
NΣN
i=1x2
i
F7 Root mean
square (RMS)
The root of MSQ or v-order 2. q1
NΣN
i=1x2
i
F8 V-order 3
(V3)
The cubic root of the average of the
cube of the data.
3
q1
NΣN
i=1x3
i
F9 Log detector
(LD)
The exponential of the average of the
log data.
exp 1
NΣN
i=1log|xi|
F10 Difference
absolute
standard
deviation
(DABS)
Standard deviation of the absolute of
the differential data. q1
N1ΣN
i=2(xixi1)2
F11 Maximum
fractal length
(MFL)
Equivalent to the log of DABS minus
an offset that is equal to 1/2log(N1)
log qΣN
i=2(xixi1)2
F12 Myopulse
percentage
rate (MPR)
The number of times the absolute of
the data exceeds the standard devia-
tion.
ΣN
i=1sgn(|xi| − xstd)
F13 Mean absolute
value slope
(MAVS)
A modified version of MAV that is
the difference between the MAV of the
first half of a signal window and the
second half.
ΣbN/2c
i=1 |xi|−ΣN
i=bN/2c+1|xi|
bN/2c
F14 Weighted
mean absolute
(WMA)
A modified version of the MAB where
the first and last 25% of a signal win-
dow is given less weight than the mid-
dle 50%.
1
NΣN
i=1wi|xi|where wi= 1 if i
[0.25N, 0.75N], and wi= 0.5 other-
wise
6
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted February 19, 2021. ; https://doi.org/10.1101/2021.02.18.431483doi: bioRxiv preprint
Figure 3: An example of feature data in one trial which show clear correlation with the finger’s
movement. The amplitude of each feature is normalized by a fixed value.
kbps per channel), which is more than 40 times lower than the raw data rate. Fig. 3 presents an
example of the feature data in one trial that shows a clear correlation between the changes of the
14 extracted features and the finger’s movement. The amplitude of each feature is normalized by
a fixed value before feeding to the DL models.
4 Proposed Methods and Deep Learning Models For Motor Intent
Decoding
4.1 Two-Step (2S) Strategy
Each finger exists in a binary state: active or inactive, depending on the patient’s intent to move
it or not. There are 32 different combinations of five fingers corresponding to 32 hand gestures.
Only a few gestures are frequently used in daily living activities, such as bending a finger (“10000”,
“01000”,... , “00001”), index pinching (“00011”), or grasp/fist (“11111”). Therefore, classifying
the hand gesture before regressing the fingers’ trajectories would significantly reduce the possible
outcome and lead to more accurate predictions. The movement of inactive fingers could also be set
to zero, which lessens the false positives when a finger “wiggles” while it is not supposed to.
Fig. 4(A) shows an illustration of the 2S strategy. In the first step, the classification output
is a [1×5] vector encoding the state of five fingers. While the dataset used in this study only
includes nine possible outcomes, the system can be easily expanded in the future to cover more
hand gestures by appending the dataset and fine-tuning the models. In contrast, many past studies
focus on classifying a specific motion, which requires modifying the architecture and re-training the
models to account for additional gestures.
In the second step, the trajectory of each DOF is regressed by a DL model. Ten separate
models regress the trajectory of ten DOF (two per finger). The models associated with inactive
7
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted February 19, 2021. ; https://doi.org/10.1101/2021.02.18.431483doi: bioRxiv preprint
Figure 4: Illustration of the (A) two-step (2S) and (B) one-step (1S) strategy for deploying DL
models.
fingers are disabled, and the prediction outputs are set to zero. As a result, the dataset used
to train each model is only a subset of the full dataset where the corresponding DOF is active.
While all models use the same architecture, they are independently optimized using different sets
of training parameters such as learning rate, minibatch size, number of epochs, etc., to achieve the
best performance. An advantage of this approach is that if one DOF fails or has poor performance,
it would not affect the performance of others.
4.2 One-Step (1S) Strategy
Fig. 4(B) shows an illustration of the 1S strategy. It is the most straightforward approach where
the trajectories of each DOF are directly regressed regardless of the fingers’ state. As a result,
the full dataset, which includes data when the DOF is active (positive samples) and idle (negative
samples), must be used to train each DOF. Because the number of negative samples often exceeds
the number of positive samples from 5:1 to 10:1, additional steps such as data augmentation and/or
weight balancing need to be done during training. This also leads to more false-positives where an
idle DOF still has small movements that could affect the overall accuracy.
Moreover, it is worth noting that the time-latency and efficiency of the 1S approach do not
necessarily better than the 2S. In the second step of the 2S approach, several DL models are
disabled depending on the hand gesture, resulting in lower overall latency and computation in most
implementations where there is only one processing unit (GPU or CPU). The actual performance
differences are difficult to be quantitatively measured because the proportion of individual hand
8
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted February 19, 2021. ; https://doi.org/10.1101/2021.02.18.431483doi: bioRxiv preprint
gestures largely depends on the user and scenarios. Nevertheless, by comparing the decoding
outcomes of the 1S and 2S strategy, we can better understand the trade-off between simplicity and
accuracy.
4.3 Deep Learning Models
Figure 5: Architecture of the DL models: (A) CNN for classification, (B) RNN for classification,
(C) CNN for regression, (D) RNN for regression.
Table 2: Comparison between this work and Nguyen & Xu et al. (2020)
No. of layers No. of parameters
Conv. LSTM Fully conn.
Nguyen & Xu et al. (2020) 21 2 3 25,927,050
This work (classification) 1 0 3 1,465,749
This work (regression) 3 2 0 767,200
Fig. 5 shows the architecture of the DL classification and regression models. They include
standard building blocks such as convolutional, long-short term memory (LSTM), fully-connected,
and dropout layers of different combinations, order, and set of parameter values. The architecture
is optimized by gradually adding layers and tuning their parameters while tracking the decoder’s
efficacy using 5-fold cross-validation. As the performance converges, additional layers would tend
to result in over-fitting.
There are ten copies of the regression model for ten DOF, each of which is trained separately.
We use Adam optimizer with the default parameters β1= 0.99, β2= 0.999, and a weight decay
regularization L2= 105. The mini-batch size is set to 38, with each training epoch consists of
10 mini-batches. The learning rate is initialized to 0.005 and reduced by a factor of 10 when the
training loss stopped improving for two consecutive epochs.
9
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted February 19, 2021. ; https://doi.org/10.1101/2021.02.18.431483doi: bioRxiv preprint
Table 2 shows a rough comparison between the DL models used in this and our previous work.
Note that for regression, the number of learnable parameters is a total of ten models for ten different
DOF. The addition of feature extraction, thus dimensional reduction, allows significantly lowing the
DL models’ size and complexity. This is essential for translating the proposed decoding paradigm
into a real-time implementation for portable systems.
5 Experimental Setup
In this research, we investigate the performance of two main DL architectures: the CNN and
RNN for both classification and regression tasks. Besides, the DL models are benchmarked against
“classic” supervised learning techniques as the baseline. They include support vector machine
(SVM), random forest (RF), and multi-layer perceptron (MLP).
For baseline techniques, the input of the classification task is the average of 224 features across
200 time-steps, while the input of the regression task is the 30 most important PCA components.
The SVM models use the radial basis function (RBF) for classification and polynomial kernel
of degree three for regression, with parameter C= 1. The RF models use five and ten trees
for classification and regression, respectively, with a max depth of three. The MLP model for
classification is created by replacing the convolutional layer of the CNN model with a fully-connected
layer of 200 units. The MLP model for regression has four layers with 300, 300, 300, and 50 units,
respectively.
The 5-fold cross-validation is used to compare the performance of the classification task. For
the regression task, the dataset is randomly split with 80% for training and 20% for validation. The
split is done such that no data windows from the training set to overlap with any data windows
from the validation set.
6 Metrics and Results
6.1 Metrics
This subsection introduces the metrics to measure the performance of five models, including the two
discussed DL models and three other classic supervised learning techniques as benchmarks in both
classification and regression tasks. The performance of the classification task is evaluated using
standard metrics including accuracy and F1 score derived from true-positive (TP), true-negative
(TN), false-positive (FP), and false-negative (FN) as follows:
Sensitivity = TP/(TP + FN) (1)
Specificity = TN/(TN + FP) (2)
Precision = TP/(TP + FP) (3)
Accuracy = (Sensitivity + Specificity)/2 (4)
F1 Score = 2·Sensitivity ·Precision
Sensitivity + Precision (5)
We use the definition of accuracy with an equal weight of sensitivity and specificity because the
occurrence of class-1 (active finger) is largely outnumbered by the occurrence of class-0 (inactive
finger).
The performance of the regression task is quantified by two metrics: mean squared error (MSE)
and variance accounted for (VAF). MSE measures the absolute deviation of an estimated from the
10
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted February 19, 2021. ; https://doi.org/10.1101/2021.02.18.431483doi: bioRxiv preprint
actual value of a DOF while VAF reflects relative deviation from the actual values of several DOF.
They are defined as follows:
MSE(y, ˆy) = 1
NΣN
i=1(ˆyiyi)2(6)
VAF(y, ˆy)=1ΣN
i=1(ˆyiyi)2
ΣN
i=1(yi¯yi)2(7)
where Nis the number of samples, yis the ground-truth trajectory, ¯yis the average of y, and ˆyis
the estimated trajectory. The value of yand ˆyare normalized in a range [0, 1] in which 0 represents
the resting position.
Although the MSE is the most common metric and effectively measures absolute prediction
errors, it cannot reflect the relative importance of each DOF to the general movements. For
example, if the average magnitude of DOF A is hundreds of times smaller than that of DOF B, a
bad estimation of A still can yield lower MSE than a reasonable estimation of B. The VAF score is
more robust in such scenarios; thus, it could be used to compare the performance between DOF of
different magnitude. The value of the VAF score ranges from (-, 1], the higher, the better. For
practicality, negative VAF values are ignored when presenting the data.
Table 3: Classification performance
Accuracy F1 score
Thumb Index Middle Ring Little Thumb Index Middle Ring Little
SVM >0.99 >0.77 0.92 0.90 0.93 >0.99 0.81 0.91 0.77 0.68
RF >0.99 0.97 >0.99 0.99 0.99 >0.99 0.98 >0.99 0.98 0.98
MLP >0.99 0.96 0.97 0.97 0.97 >0.99 0.97 0.97 0.95 0.95
CNN >0.99 >0.99 >0.99 >0.99 >0.99 >0.99 >0.99 >0.99 >0.99 >0.99
RNN >0.99 0.95 0.99 0.97 0.99 >0.99 0.96 0.99 0.95 0.99
6.2 Classification Results
Table 3 shows the average 5-fold cross-validation classification results of all the techniques. The
predictability of each finger is largely different from one another. The thumb, which produces
strong signals only on the median nerve (first eight channels), is easily recognized by all techniques.
Overall, CNN offers the best performance with accuracy and an F1 score for all finger exceeding
99%. RF closely follows with performance ranging from 98% to 99%. While DL still outperforms
classic techniques, it is worth noting that RF could also be a prominent candidate for real-world
implementation because RF can be more efficiently deployed in low-power, portable systems.
6.3 Regression Results
Fig. 6 presents the MSE and VAF scores for both strategies. In the 1S approach, the DL models,
especially RNN, significantly outperform the other methods in both MSE and VAF, as shown in Fig.
6(A, C). In the 2S approach, the performance is more consistent across all methods, where classic
methods even outperform DL counterparts in certain DOF as shown in Fig. 6(B, D). Between the
two strategies, the 1S approach generally gives better results; however, the high performance can
only be achieved with RNN.
11
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted February 19, 2021. ; https://doi.org/10.1101/2021.02.18.431483doi: bioRxiv preprint
Figure 6: Regression performance in term of MSE (A, B) and VAF (C, D).
7 Discussion
7.1 Feature Extraction Reduces Decoders’ Complexity
Both DL architectures investigated in this study, namely CNN and RNN, delivers comparable mo-
tor decoding performance to our previous work (Nguyen & Xu, 2020) while require much lower
computational resources to implement. The average VAF score for most DOF is 0.7, with some
exceeding 0.9. Such results are achieved with relatively shallow DL models with 4-5 layers, a sig-
nificant reduction from the previous implementation with 26 layers. While an extra step of feature
extraction is required, all feature extraction techniques are specifically designed to be efficiently
computed with conventional arithmetic processors (e.g., CPU) and/or hardware accelerators (e.g.,
FPGA, ASIC).
Another direction for future works would be including additional features. Here we apply the
most common features in the temporal domain for their simplicity and well-established standing
12
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted February 19, 2021. ; https://doi.org/10.1101/2021.02.18.431483doi: bioRxiv preprint
for neuroprosthesis applications. There are other features such as mean power, median frequency,
peak frequency, etc. in the frequency domain that have been explored in previous literature.
The purpose of this study is not to exhaustively investigate the effect of all existing features in
motor decoding but to prove the effectiveness of the feature extraction method in achieving decent
decoding outcomes. The success of this method will open up to future research on simultaneously
applying more features in multiple domains for better outcomes.
Furthermore, it is worth noting that the use of feature extraction excludes most of the high-
frequency band 600-3000 Hz, which is shown in our previous work that could contain additional
nerve information associated with neural spikes. A future direction would be extracting that
information using spike detection and sorting technique and combine them with the information of
the low-frequency band to boost the prediction accuracy. However, the computational complexity
must be carefully catered to not to hinder the real-time aspect of the overall system.
7.2 Classic Machine Learning v.s. Deep Learning
The classification task can be accomplished with high accuracy using most classic ML techniques
(e.g., RF) and near-perfect with DL approaches (e.g., CNN). Along with other evidence in (Nguyen
& Xu, 2020), this suggests that nerve data captured by our neural interface contain apparent neural
patterns that can be clearly recognized to control neuroprostheses. While the current dataset only
covers 9/32 different hand gestures, these are still promising results that would support future
developments, including expanding the dataset to cover additional gestures.
The regression results are consistent with the conclusion of many past studies that DL techniques
only show clear advantages over classic ML methods when handling a large dataset. This is evident
in the 1S strategy, where each DOF is trained with the full dataset consisting of all possible hand
gestures. In contrast, in the 2S strategy, where the dataset is divided into smaller subsets, DL
techniques lose their leverage. However, as the dataset is expanded in the future, we generally
believe that DL techniques should emerge as the dominant approach.
8 Conclusion
This work presents several approaches to optimize the motor decoding paradigm that interprets
the motor intent embedded in the peripheral nerve signals for controlling the prosthetic hand. The
use of feature extraction largely reduces the data dimensionality while retaining essential neural
information in the low-frequency band. This allows achieving similar decoding performance with
DL architectures of much lower computational complexity. Two different strategies for deploying
DL models, namely 2S and 1S, with a classification and a regression stage, are also investigated.
The results indicate that CNN and RF can deliver high accuracy classification performance, while
RNN gives better regression performance when trained on the full dataset with the 1S approach.
The findings layout an important foundation for the next development, which is translating the
proposed motor decoding paradigm to real-time applications, which requires not only accuracy but
also efficiency.
Acknowledgments
The surgery and patients related costs were supported in part by the Defense Advanced Research
Projects Agency (DARPA) under Grants HR0011-17-2-0060 and N66001-15-C-4016. The tech-
nology development and human experiments were supported in part by internal funding from the
13
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted February 19, 2021. ; https://doi.org/10.1101/2021.02.18.431483doi: bioRxiv preprint
University of Minnesota, in part by the NIH under Grant R21-NS111214, in part by NSF CAREER
Award No. 1845709, and in part by Fasikl Incorporated.
Zhi Yang is co-founder of, and holds equity in, Fasikl Inc., a sponsor of this project. This
interest has been reviewed and managed by the University of Minnesota in accordance with its
Conflict of Interest policy. J. Cheng and E. W. Keefer have ownership in Nerves Incorporated, a
sponsor of this project.
References
Amsuss, S., Goebel, P. M., Jiang, N., Graimann, B., Paredes, L., & Farina, D. (2013). Self-
Correcting Pattern Recognition System of Surface EMG Signals for Upper Limb Prosthesis Con-
trol. IEEE Transactions on Biomedical Engineering, 61(4), 1167-1176.
Cheng, J., Tillery, S. H., Miguelez, J., Lachapelle, J., & Keefer, E. (2017). Dexterous Hand Control
Through Fascicular Targeting (HAPTIX-DEFT): Level 4 Evidence. Journal of Hand Surgery,
42(9), S8-S9.
Cordella, F., Ciancio, A. L., Sacchetti, R., Davalli, A., Cutti, A. G., Guglielmelli, E., & Zollo, L.
(2016). Literature Review on Needs of Upper Limb Prosthesis Users. Frontiers in Neuroscience,
10, 209.
Davis, T. S., Wark, H. A. C., Hutchinson, D. T., Warren, D. J., O’Neill, K., Scheinblum, T., Clark,
G. A., Normann, R. A., & Greger, B. (2016). Restoring Motor Control and Sensory Feedback
in People with Upper Extremity Amputations Using Arrays of 96 Microelectrodes Implanted in
The Median and Ulnar Nerves. Journal of Neural Engineering, 13(3), 036001.
Fougner, A., Stavdahl, O., Kyberd, P. J., Losier, Y. G., & Parker, P. A. (2012). Control of Upper
Limb Prostheses: Terminology and Proportional Myoelectric Control-A Review. IEEE Transac-
tions on Neural Systems and Rehabilitation Engineering, 20(5), 663-677.
George, J. A., Davis, T. S., Brinton, M. R., & Clark, G. A. (2020). Intuitive Neuromyoelectric
Control of a Dexterous Bionic Arm Using a Modified Kalman Filter. Journal of Neuroscience
Methods, 330, 108462.
George, J. A., Radhakrishnan, S., Brinton, M. R., & Clark, G. A. (2020). Inexpensive and Portable
System for Dexterous High-Density Myoelectric Control of Multiarticulate Prostheses. IEEE
International Conference on Systems, Man and Cybernetics (SMC), Toronto.
Harris, A., Katyal, K., Para, M., & Thomas, J. (2011). Revolutionizing Prosthetics Software Tech-
nology. 2011 IEEE International Conference on Systems, Man, and Cybernetics, Anchorage, AK,
2877-2884.
Hochberg, L. R., Bacher, D., Jarosiewicz, B., Masse, N. Y., Simeral, J. D., Vogel, J., Haddadin, S.,
Liu, J., Cash, S. S., & van der Smagt, P. (2012). Reach and Grasp by People with Tetraplegia
Using a Neurally Controlled Robotic Arm. Nature, 485(7398), 372.
Hu, S., Wang, H., Zhang, J., Kong, W., Cao, Y., & Kozma, R. (2015). Comparison Analysis:
Granger Causality and New Causality and Their Applications to Motor Imagery. IEEE Trans-
actions on Neural Networks and Learning Systems, 27(7), 1429-1444.
14
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted February 19, 2021. ; https://doi.org/10.1101/2021.02.18.431483doi: bioRxiv preprint
Irwin, Z. T., Schroeder, K. E., Vu, P. P., Bullard, A. J., Tat, D. M., Nu, C. S., Vaskov, A.,
Nason, S. R., Thompson, D. E., & Bentley, J. N. (2017). Neural Control of Finger Movement via
Intracortical Brain–Machine Interface. Journal of Neural Engineering, 14(6), 066004.
Jiang, N., Vest-Nielsen, J. L., Muceli, S., & Farina, D. (2012). EMG-Based Simultaneous and
Proportional Estimation of Wrist/Hand Kinematics in Unilateral Transradial Amputees. Journal
of Neuroengineering and Rehabilitation, 9(1), 42.
Johannes, M. S., Bigelow, J. D., Burck, J. M., Harshbarger, S. D., Kozlowski, M. V., & Van Doren,
T. (2011). An Overview of the Developmental Process for the Modular Prosthetic Limb. Johns
Hopkins APL Technical Digest, 30(3), 207-216.
Kwon, O.-Y., Lee, M.-H., Guan, C., & Lee, S.-W. (2019). Subject-Independent Brain–Computer In-
terfaces Based on Deep Convolutional Neural Networks. IEEE Transactions on Neural Networks
and Learning Systems, 31(10), 3839-3852.
Mahmud, M., Kaiser, M. S., Hussain, A., & Vassanelli, S. (2018). Applications of Deep Learning
and Reinforcement Learning to Biological Data. IEEE Transactions on Neural Networks and
Learning Systems, 29(6), 2063-2079.
Micera, S., Rossini, P. M., Rigosa, J., Citi, L., Carpaneto, J., Raspopovic, S., Tombini, M., Cipriani,
C., Assenza, G., & Carrozza, M. C. (2011). Decoding of Grasping Information From Neural
Signals Recorded Using Peripheral Intrafascicular Interfaces. Journal of Neuroengineering and
Rehabilitation, 8(1), 1-10.
Mollazadeh, M., Greenwald, E., Thakor, N., Schieber, M., & Cauwenberghs, G. (2011). Wire-
less Micro-ECoG Recording in Primates During Reach-to-Grasp Movements. IEEE Biomedical
Circuits and Systems Conference (BioCAS), San Diego, CA, 237-240.
Nguyen, A. T., Xu, J., Jiang, M., Luu, D. K., Wu, T., Tam, W.-k., Zhao, W., Drealan, M. W.,
Overstreet, C. K., Zhao, Q., Cheng, J., Keefer, E. W., & Yang, Z. (2020). A Bioelectric Neural
Interface Towards Intuitive Prosthetic Control For Amputees. Journal of Neural Engineering,
17(6), 066001.
Overstreet, C. K., Cheng, J., & Keefer, E. W. (2019). Fascicle Specific Targeting For Selective
Peripheral Nerve Stimulation. Journal of Neural Engineering, 16(6), 066040.
Phinyomark, A., Limsakul, C., & Phukpattaranont, P. (2009). A Novel Feature Extraction for
Robust EMG Pattern Recognition. arXiv:0912.3973,
Phinyomark, A., Nuidod, A., Phukpattaranont, P., & Limsakul, C. (2012). Feature Extraction
and Reduction of Wavelet Transform Coefficients for EMG Pattern Classification. Elektronika ir
Elektrotechnika, 122(6), 27-32.
Rafiee, J., Rafiee, M., Yavari, F., & Schoen, M. (2011). Feature Extraction of Forearm EMG Signals
for Prosthetics. Expert Systems with Applications, 38(4), 4058-4067.
Sakhavi, S., Guan, C., & Yan, S. (2018). Learning Temporal Information for Brain-Computer
Interface Using Convolutional Neural Networks. IEEE Transactions on Neural Networks and
Learning Systems, 29(11), 5619-5629.
Schultz, A. E., & Kuiken, T. A. (2011). Neural Interfaces for Control of Upper Limb Prostheses:
the State of The Art and Future Possibilities. PM&R, 3(1), 55-67.
15
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted February 19, 2021. ; https://doi.org/10.1101/2021.02.18.431483doi: bioRxiv preprint
Sebelius, F. C., Rosen, B. N., & Lundborg, G. N. (2005). Refined Myoelectric Control in Below-
Elbow Amputees Using Artificial Neural Networks and A Data Glove. The Journal of Hand
Surgery, 30(4), 780-789.
Vu, P. P., Irwin, Z. T., Bullard, A. J., Ambani, S. W., Sando, I. C., Urbanchek, M. G., Cederna, P.
S., & Chestek, C. A. (2017). Closed-Loop Continuous Hand Control via Chronic Recording of Re-
generative Peripheral Nerve Interfaces. IEEE Transactions on Neural Systems and Rehabilitation
Engineering, 26(2), 515-526.
Vu, P. P., Vaskov, A. K., Irwin, Z. T., Henning, P. T., Lueders, D. R., Laidlaw, A. T., Davis, A.
J., Nu, C. S., Gates, D. H., & Gillespie, R. B. (2020). A Regenerative Peripheral Nerve Interface
Allows Real-Time Control of an Artificial Hand in Upper Limb Amputees. Science Translational
Medicine, 12(533), eaay2857.
Wendelken, S., Page, D. M., Davis, T., Wark, H. A., Kluger, D. T., Duncan, C., Warren, D. J.,
Hutchinson, D. T., & Clark, G. A. (2017). Restoration of Motor Control and Proprioceptive
and Cutaneous Sensation in Humans with Prior Upper-Limb Amputation via Multiple Utah
Slanted Electrode Arrays (USEAs) Implanted in Residual Peripheral Arm Nerves. Journal of
Neuroengineering and Rehabilitation, 14(1), 121.
Xu, J., Wu, T., Liu, W., & Yang, Z. (2014). A Frequency Shaping Neural Recorder with 3 pF Input
Capacitance and 11 Plus 4.5 Bits Dynamic Range. IEEE Transactions on Biomedical Circuits
and Systems, 8(4), 510-527.
Xu, J., Nguyen, A. T., Luu, D. K., Drealan, M., & Yang, Z. (2020). Noise Optimization Techniques
for Switched-Capacitor Based Neural Interfaces. IEEE Transactions on Biomedical Circuits and
Systems, 15(5), 1024-1035.
Yang, Z., Xu, J., Nguyen, A. T., Wu, T., Zhao, W., & Tam, W.-k. (2016). Neuronix Enables Con-
tinuous, Simultaneous Neural Recording and Electrical Microstimulation. Annual International
Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Orlando, FL,
4451-4454.
Yang, Z., Xu, J., Nguyen, A. T., & Wu, T. (2018). System and Method for Simultaneous Stim-
ulation and Recording Using System-on-Chip (SoC) Architecture. U.S. Patent Application No.
15/876,030.
Yang, Z., Nguyen, A. T., & Xu, J. (2020). System and Method for Charge-Balancing Neurostimu-
lator with Neural Recording. U.S. Patent No. 10,716,941.
Zardoshti-Kermani, M., Wheeler, B. C., Badie, K., & Hashemi, R. M. (1995). EMG Feature Evalua-
tion for Movement Control of Upper Extremity Prostheses. IEEE Transactions on Rehabilitation
Engineering, 3(4), 324-333.
Zeng, H., & Song, A. (2015). Optimizing Single-Trial EEG Classification by Stationary Matrix
Logistic Regression in Brain–Computer Interface. IEEE Transactions on Neural Networks and
Learning Systems, 27(11), 2301-2313.
Zhang, Y., Nieveen, J., Wendelken, S., Page, D., Davis, T., Bo, A. P. L., Hutchinson, D., Clark, G.
A., Warren, D. J., & Zhang, C. (2017). Individual Hand Movement Detection and Classification
Using Peripheral Nerve Signals. International IEEE EMBS Conference on Neural Engineering
(NER), Shanghai, 448-451.
16
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted February 19, 2021. ; https://doi.org/10.1101/2021.02.18.431483doi: bioRxiv preprint
Ziegler-Graham, K., MacKenzie, E. J., Ephraim, P. L., Travison, T. G., & Brookmeyer, R. (2008).
Estimating the Prevalence of Limb Loss in the United States: 2005 to 2050. Archives of Physical
Medicine and Rehabilitation, 89(3), 422-429.
Zuleta, J., Ferro, M., Murillo, C., & Franco-Luna, R. (2019). Identification of EMG Activity
With Machine Learning in Patients With Amputation of Upper Limbs for the Development
of Mechanical Prostheses. IOP Conference Series: Materials Science and Engineering, 012010.
17
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted February 19, 2021. ; https://doi.org/10.1101/2021.02.18.431483doi: bioRxiv preprint
ResearchGate has not been able to resolve any citations for this publication.
Article
Full-text available
This paper presents the noise optimization of a novel switched-capacitor (SC) based neural interface architecture, and its circuit demonstration in a 0.13 μm\mu\text{m} CMOS process. To reduce thermal noise folding ratio, and suppress kT/CkT/C noise, several noise optimization techniques are developed in the proposed architecture. First, one parasitic capacitance suppression scheme is developed to block noise charge transfer from parasitic capacitors to amplifier output. Second, one recording path-splitting scheme is proposed in the input sampling stage to selectively record local field potentials (LFPs), extracellular spikes, or both for reducing input noise floor, and total power consumption. Third, an auto-zero noise cancellation scheme is developed to suppress kT/CkT/C noise in the neural amplifier stage. A prototype neural interface chip was fabricated, and also verified in both bench-top, and In-Vivo experiments. Bench-top testings show the input-referred noise of the designed chip is 4.8 μV\mu\text{V} from 1 Hz\text{Hz} to 300 Hz\text{Hz} , and 2.3 μV\mu\text{V} from 300 Hz\text{Hz} to 8 kHzkHz respectively, and In-Vivo experiments show the peak-to-peak amplitude of the total noise floor including neural activity, electrode interface noise, and the designed chip is only around 20 μV\mu\text{V} . In comparison with conventional architectures through both circuit measurement and animal experiments, it is well demonstrated that the proposed noise optimization techniques can effectively reduce circuit noise floor, thus extending the application range of switched-capacitor circuits.
Article
Full-text available
Peripheral nerves provide a promising source of motor control signals for neuroprosthetic devices. Unfortunately, the clinical utility of current peripheral nerve interfaces is limited by signal amplitude and stability. Here, we showed that the regenerative peripheral nerve interface (RPNI) serves as a biologically stable bioamplifier of efferent motor action potentials with long-term stability in upper limb amputees. Ultrasound assessments of RPNIs revealed prominent contractions during phantom finger flexion, confirming functional reinnervation of the RPNIs in two patients. The RPNIs in two additional patients produced electromyography signals with large signal-to-noise ratios. Using these RPNI signals, subjects successfully controlled a hand prosthesis in real-time up to 300 days without control algorithm recalibration. RPNIs show potential in enhancing prosthesis control for people with upper limb loss.
Article
Full-text available
For a brain-computer interface (BCI) system, a calibration procedure is required for each individual user before he/she can use the BCI. This procedure requires approximately 20-30 min to collect enough data to build a reliable decoder. It is, therefore, an interesting topic to build a calibration-free, or subject-independent, BCI. In this article, we construct a large motor imagery (MI)-based electroencephalography (EEG) database and propose a subject-independent framework based on deep convolutional neural networks (CNNs). The database is composed of 54 subjects performing the left- and right-hand MI on two different days, resulting in 21,600 trials for the MI task. In our framework, we formulated the discriminative feature representation as a combination of the spectral-spatial input embedding the diversity of the EEG signals, as well as a feature representation learned from the CNN through a fusion technique that integrates a variety of discriminative brain signal patterns. To generate spectral-spatial inputs, we first consider the discriminative frequency bands in an information-theoretic observation model that measures the power of the features in two classes. From discriminative frequency bands, spectral-spatial inputs that include the unique characteristics of brain signal patterns are generated and then transformed into a covariance matrix as the input to the CNN. In the process of feature representations, spectral-spatial inputs are individually trained through the CNN and then combined by a concatenation fusion technique. In this article, we demonstrate that the classification accuracy of our subject-independent (or calibration-free) model outperforms that of subject-dependent models using various methods [common spatial pattern (CSP), common spatiospectral pattern (CSSP), filter bank CSP (FBCSP), and Bayesian spatio-spectral filter optimization (BSSFO)].
Article
Full-text available
Background: Multi-articulate prostheses are capable of performing dexterous hand movements. However, clinically available control strategies fail to provide users with intuitive, independent and proportional control over multiple degrees of freedom (DOFs) in real-time. New method: We detail the use of a modified Kalman filter (MKF) to provide intuitive, independent and proportional control over six-DOF prostheses such as the DEKA "LUKE" Arm. Input features include neural firing rates recorded from Utah Slanted Electrode Arrays and mean absolute value of intramuscular electromyographic (EMG) recordings. Ad-hoc modifications include thresholds and non-unity gains on the output of a Kalman filter. Results: We demonstrate that both neural and EMG data can be combined effectively. We also highlight that modifications can be optimized to significantly improve performance relative to an unmodified Kalman filter. Thresholds significantly reduced unintended movement and promoted more independent control of the different DOFs. Gain were significantly greater than one and served to amplify participant effort. Optimal modifications can be determined quickly offline and translate to functional improvements online. Using a portable take-home system, participants performed various activities of daily living. Comparison with existing methods: In contrast to pattern recognition, the MKF allows users to continuously modulate their force output, which is critical for fine dexterity. The MKF is also computationally efficient and can be trained in less than five minutes. Conclusions: The MKF can be used to explore the functional and psychological benefits associated with long-term, at-home control of dexterous prosthetic hands.
Article
Full-text available
Deep learning (DL) methods and architectures have been the state-of-the-art classification algorithms for computer vision and natural language processing problems. However, the successful application of these methods in motor imagery (MI) brain-computer interfaces (BCIs), in order to boost classification performance, is still limited. In this paper, we propose a classification framework for MI data by introducing a new temporal representation of the data and also utilizing a convolutional neural network (CNN) architecture for classification. The new representation is generated from modifying the filter-bank common spatial patterns method, and the CNN is designed and optimized accordingly for the representation. Our framework outperforms the best classification method in the literature on the BCI competition IV-2a 4-class MI data set by 7% increase in average subject accuracy. Furthermore, by studying the convolutional weights of the trained networks, we gain an insight into the temporal characteristics of EEG.
Article
Full-text available
Background Despite advances in sophisticated robotic hands, intuitive control of and sensory feedback from these prostheses has been limited to only 3-degrees-of-freedom (DOF) with 2 sensory percepts in closed-loop control. A Utah Slanted Electrode Array (USEA) has been used in the past to provide up to 81 sensory percepts for human amputees. Here, we report on the advanced capabilities of multiple USEAs implanted in the residual peripheral arm nerves of human amputees for restoring control of 5 DOF and sensation of up to 131 proprioceptive and cutaneous hand sensory percepts. We also demonstrate that USEA-restored sensory percepts provide a useful source of feedback during closed-loop virtual prosthetic hand control. Methods Two 100-channel USEAs were implanted for 4–5 weeks, one each in the median and ulnar arm nerves of two human subjects with prior long-duration upper-arm amputations. Intended finger and wrist positions were decoded from neuronal firing patterns via a modified Kalman filter, allowing subjects to control many movements of a virtual prosthetic hand. Additionally, USEA microstimulation was used to evoke numerous sensory percepts spanning the phantom hand. Closed-loop control was achieved by stimulating via an electrode of the ulnar-nerve USEA while recording and decoding movement via the median-nerve USEA. Results Subjects controlled up to 12 degrees-of-freedom during informal, ‘freeform’ online movement decode sessions, and experienced up to 131 USEA-evoked proprioceptive and cutaneous sensations spanning the phantom hand. Independent control was achieved for a 5-DOF real-time decode that included flexion/extension of the thumb, index, middle, and ring fingers, and the wrist. Proportional control was achieved for a 4-DOF real-time decode. One subject used a USEA-evoked hand sensation as feedback to complete a 1-DOF closed-loop virtual-hand movement task. There were no observed long-term functional deficits due to the USEA implants. Conclusions Implantation of high-channel-count USEAs enables multi-degree-of-freedom control of virtual prosthetic hand movement and restoration of a rich selection of both proprioceptive and cutaneous sensory percepts spanning the hand during the short 4–5 week post-implant period. Future USEA use in longer-term implants and in closed-loop may enable restoration of many of the capabilities of an intact hand while contributing to a meaningful embodiment of the prosthesis. Electronic supplementary material The online version of this article (10.1186/s12984-017-0320-4) contains supplementary material, which is available to authorized users.
Article
Full-text available
Rapid advancement in the hardware based technologies over past decades opened up new possibilities for Biological and Life scientists to gather multimodal data from various application domains (e.g., Omics, Bioimaging, Medical Imaging, and [Brain/Body]-Machine Interfaces). Novel data intensive machine learning techniques are required to decipher these data. Recent research in Deep learning (DL), Reinforcement learning (RL), and their combination (Deep RL) promise to revolutionize Artificial Intelligence. Increasing computational power, faster data storage devices, and declining computing costs allowed scientists to apply these techniques on such enormous and complex datasets which otherwise would not have been possible. This review article provides a comprehensive survey of the applications of DL, RL, and Deep RL techniques in mining Biological data coming from various application domains. In addition, the performances of DL techniques when applied to different datasets pertaining to the various application domains have been compared. Finally, it outlines some open issues on this challenging research area and postulates possible future perspectives.
Article
Objective: While prosthetic hands with independently actuated digits have become commercially available, state-of-the-art human-machine interfaces (HMI) only permit control over a limited set of grasp patterns, which does not enable amputees to experience sufficient improvement in their daily activities to make an active prosthesis useful. Approach: Here we present a technology platform combining fully-integrated bioelectronics, implantable intrafascicular microelectrodes and deep learning-based artificial intelligence (AI) to facilitate this missing bridge by tapping into the intricate motor control signals of peripheral nerves. The bioelectric neural interface includes an ultra-low-noise neural recording system to sense electroneurography (ENG) signals from microelectrode arrays implanted in the residual nerves, and AI models employing the recurrent neural network (RNN) architecture to decode the subject's motor intention. Main results: A pilot human study has been carried out on a transradial amputee. We demonstrate that the information channel established by the proposed neural interface is sufficient to provide high accuracy control of a prosthetic hand up to 15 degrees of freedom (DOF). The interface is intuitive as it directly maps complex prosthesis movements to the patient's true intention. Significance: Our study layouts the foundation towards not only a robust and dexterous control strategy for modern neuroprostheses at a near-natural level approaching that of the able hand, but also an intuitive conduit for connecting human minds and machines through the peripheral neural pathways. (Clinical trial identifier: NCT02994160).
Article
Objective Electrical stimulation is a blunt tool for evoking neural activity. Neurons are naturally activated asynchronously and non-uniformly, whereas stimulation drives simultaneous activity within a population of cells. These differences in activation pattern can result in unintended side effects, including muddled sensory percepts and undesirable muscle contractions. These effects can be mitigated by placement of electrodes in close approximation to nerve fibers and careful selection of the neural interface's location. This work describes the benefits of placing electrodes within specific fascicles of peripheral nerve to form selective neural interfaces for bidirectional neuroprosthetic devices. Approach Chronic electrodes were targeted to individual fascicles of the ulnar and median nerves in the forearm of four human subjects. During the surgical implant procedure, fascicles were dissected from each nerve, and functional testing was used to identify the relative composition of sensory and motor fibers within each. FAST-LIFE arrays, composed of longitudinal intrafascicular arrays and fascicular cuff electrodes, were implanted in each fascicle. The location, quality, and stimulation parameters associated with sensations evoked by electrical stimulation on these electrodes were characterized throughout the 90-180 day implant period. Main Results FAST-LIFE arrays enable selective and chronic electrical stimulation of individual peripheral nerve fascicles. The quality of sensations evoked by stimulation in each fascicle is predictable and distinct; subjects reported tactile and cutaneous sensations during stimulation of sensory fascicles and deeper proprioceptive sensations during stimulation of motor fascicles. Stimulation thresholds and strength-duration time constants were typically higher within sensory fascicles. Significance Highly selective, stable neural interfaces can be created by placing electrodes within and around single fascicles of peripheral nerve. This method enables targeting electrodes to nerve fibers that innervate a specific body region or have specific functions. Fascicle-specific interfacing techniques have broad potential to maximize the therapeutic effects of electrical stimulation in many neuromodulation applications. (Clinical Trial ID NCT02994160).