Available via license: CC BY 4.0
Content may be subject to copyright.
IEEE TRANSACTIONS AND JOURNALS TEMPLATE 1
Development of a Wearable Sleeve-Based
System Combining Polymer Optical Fiber
Sensors and an LSTM Network for Estimating
Knee Kinematics
B. L. Pugliese, Member, IEEE, A. Angelucci, Member, IEEE, F. Parisi, Member, IEEE, S. Sapienza, E.
Fabara, G. Corniani, Member, IEEE, A. S. Tenforde, MD, A. Aliverti, Senior Member, IEEE,
D. Demarchi, Senior Member, IEEE, and P. Bonato, Senior Member, IEEE
Abstract— This study presents a novel wearable solution
integrating Polymer Optical Fiber (POF) sensors into a knee
sleeve to monitor knee flexion/extension (F/E) patterns
during walking. POF sensors offer advantages such as
flexibility, light weight, and robustness to electromagnetic
interference, making them ideal for wearable applications.
However, when one integrates these sensors into a knee
sleeve, they exhibit non-linearities, including hysteresis and
mode coupling, which complicate signal interpretation. To
address this issue, a Long Short-Term Memory (LSTM)
network was implemented to model temporal dependencies
in sensor output, hence providing accurate knee angle
estimates. Data were collected from 31 participants walking
at different speeds on a treadmill, using a camera-based
motion capture system for validation. Configurations with
multiple (up to five) sensors were considered. The best
performance was achieved using three sensors, yielding a
median root mean square error (RMSE) of 3.41° (interquartile
range: 2.50° – 5.19°). Whereas using multiple sensors
generally improved robustness, the inclusion of data from
sub-optimally placed sensors negatively affected
performance. The technology holds potential for clinical
application in knee osteoarthritis (OA) management. Future
work should focus on optimizing signal calibration and
expanding the dataset to facilitate accounting for the
different ways in which the knee sleeve conforms to the
anatomy of different individuals.
Index Terms— Digital health, kinematics, long short-term
memory network, polymer optical fiber sensors, knee
osteoarthritis, wearable technology.
This work was supported Mitsui Chemicals (Minato City, Tokyo,
Japan). (Corresponding author: B. L. Pugliese).
B. L. Pugliese is with the Department of Electronics and
Telecommunications, Politecnico di Torino, Turin, 10129 Italy, and
Department of Physical Medicine and Rehabilitation, Harvard Medical
School, Spaulding Rehabilitation Hospital, Boston, MA 02129 USA (e-
mail: bpugliese@mgh.harvard.edu).
A. Angelucci and A. Aliverti are with Dipartimento di Elettronica,
Informazione e Bioingegneria, Politecnico di Milano, Milan, 20133, Italy
(e-mail: alessandra.angelucci@polimi.it, andrea.aliverti@polimi.it).
I. INTRODUCTION
NEE osteoarthritis (OA) is a chronic disease of the joint
characterized by degenerative changes affecting bone,
cartilage, menisci, synovium, and ligaments. This
condition affects approximately 9% of men and 18% of women
over the age of 65 years [1], [2]. Individuals with knee OA
suffer from pain, stiffness, and decreased range of motion
(ROM), significantly impairing their physical abilities and
restricting daily activities [2], [3]. One of the most prominent
gait alterations observed in knee OA patients is reduced knee
flexion/extension (F/E) angular displacement during walking,
which has been linked to disease progression and symptom
severity. Studies have consistently demonstrated that patients
with knee OA have significantly lower knee ROM during both
the stance and swing phases of gait cycle compared to healthy
controls [2], [3], [4], [5]. These altered gait biomechanics may
result from joint stiffening and compensatory strategies aimed
at minimizing pain and stabilizing the knee. However, such
compensation may lead to increased impact loading on the
tibiofemoral joint, potentially accelerating disease progression
[3], [5].
Despite the potential harm caused by these compensation
strategies, maintaining physical activity, particularly through
walking programs, is crucial for managing knee OA. Regular
movement helps preserve joint function, slow disease
progression, and prevent further disability [6], [7], [8].
However, titrating walking programs is challenging. Current
methods often rely on self-report of activity, which is not only
subjective but also delayed, capturing pain after it has already
occurred [9], [10]. This delayed feedback prevents timely
F. Parisi, S. Sapienza, E. Fabara, G. Corniani, A. S. Tenforde, and P.
Bonato are with Department of Physical Medicine and Rehabilitation,
Harvard Medical School, Spaulding Rehabilitation Hospital, Boston, MA
02129 USA (e-mail: parisi.fed@gmail.com,
stefano.sapienza87@gmail.com, efabara@mgb.org,
gcorniani@mgh.harvard.edu, atenforde@mgh.harvard.edu,
pbonato@mgh.harvard.edu).
D. Demarchi is with the Department of Electronics and
Telecommunications, Politecnico di Torino, Turin, 10129 Italy (e-mail:
danilo.demarchi@polito.it)
K
This article has been accepted for publication in IEEE Transactions on Neural Systems and Rehabilitation Engineering. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2025.3540708
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
2
adjustments that could help avoid exacerbations. Real-time
monitoring of knee F/E patterns during walking could address
this issue by detecting early deviations in movement patterns.
By identifying these changes, it may be possible to make
proactive adjustments to walking programs, reducing the risk of
exacerbating symptoms and preventing disease progression.
Camera-based motion capture systems are considered the gold
standard for measuring joint kinematics. However, these
systems are expensive, cumbersome, and limited to laboratory
environments [11]. Wearable technologies have emerged as a
practical solution, offering more accessible ways to track
motion in real-world settings [12]. Among them, textile-
integrated sensors present a promising avenue for embedding
monitoring capabilities into everyday clothing, enabling
continuous data collection with minimal burden to the user.
Polymer Optical Fiber (POF) sensors [13], [14], stand out as a
particularly suitable option for integration into smart textiles
[15]. These sensors offer several advantages over other
technologies due to their flexibility, lightweight nature, and
robustness to electromagnetic interference. Their flexible
nature allows them to conform well to the knee anatomy,
enabling unobtrusive data collection during movement.
Strategic placement of these sensors at different anatomical
locations around the knee captures complementary information,
as different sensors are expected to bend differently (according
to the anatomical geometry of the knee) during F/E movements.
This distributed sensing approach is expected to provide a more
accurate estimate of the knee joint angle [16].
POF sensors operate based on the principle of detecting
variations in light transmission as the fiber undergoes bending
[17]. When light is injected into a POF, it propagates through
multiple paths, or modes, within the fiber. These modes are
characterized by distinct patterns of internal reflections. Lower-
order modes, which involve fewer reflections within the fiber,
are marked by minimal attenuation. In contrast, higher-order
modes undergo more frequent reflections, leading to signal
attenuation as the fiber bends [18]. The multimodal nature of
POF sensors causes nonlinearities in the form of mode
coupling, where light initially confined to one mode can
transfer to another as the fiber bends, thus complicating the
interpretation of the sensor’s output [19]. Another source of
nonlinearity in POF sensors is hysteresis, where the relationship
between curvature and light transmission exhibits a lag when
the fiber is deformed and returns to its original shape, leading
to discrepancies in signal readings depending on the fiber’s
deformation pattern [20].
To handle the above-described nonlinearities in POF sensor
outputs, Long Short-Term Memory (LSTM) networks [21],
[22], [23], [24], [25], [26], [27], [28] are employed in this study.
LSTMs are designed to analyze timeseries data, allowing them
to model temporal relationships. With the integration of
redundant sensor data, LSTMs can learn patterns across
multiple sensors, providing a robust approach to mitigating
individual sensor errors while enhancing the overall precision
of the knee angle estimation during walking.
In this study, we utilize a knee sleeve with integrated POF
sensors and employ LSTM models to estimate knee F/E angular
displacement patterns during walking. We discuss the
importance of sensor placement and redundancy in enhancing
the system performance. We argue that this work is an
important step towards advancing knee OA monitoring, with
the ultimate goal of developing effective tools for clinical use.
II. MATERIAL AND METHODS
The experimental setup and the pipeline for data acquisition and
processing are shown in Fig. 1. Data were collected from 31
participants walking on a treadmill while wearing the
sensorized knee sleeve. Data collected simultaneously via a
camera-based motion capture system (Vicon Motion Systems
Inc, Oxford, UK) were used as gold standard. The motion
capture data were processed and segmented to estimate the
patterns of knee F/E angular displacement for each gait cycle.
The POF sensors’ output was also segmented according to the
gait cycles. LSTM networks were trained for various sensor
configurations. The performance of the models was evaluated
to determine the impact on estimation accuracy of the position
of sensors and their number. The error distribution over the gait
cycle was derived for the best-performing model.
This article has been accepted for publication in IEEE Transactions on Neural Systems and Rehabilitation Engineering. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2025.3540708
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
1
Fig. 1. Schematic representation of data acquisition and processing pipeline. A) Gold standard (camera-based motion capture system) and knee
sleeve prototype; B) Experimental setup: healthy volunteer walking on a treadmill and simultaneously wearing the knee sleeve and the markers for
the camera-based motion capture system; C) Data processing pipeline with a LSTM network configured for five channels. The model underwent
training for various channel counts and for every possible combination for a given channel count.
A. Hardware Description
The knee sleeve, engineered by Mitsui Chemicals (Minato City,
Tokyo, Japan), incorporates five multimodal POF sensors, each
equipped with an LED at one end and a photodiode at the other.
The fibers are encased in a shrink tubing to enhance durability.
As the knee undergoes F/E, these fibers deform, altering light
transmission and generating voltage variations. The sensors are
placed medially (M), laterally (L), inside-medially (IM), inside-
laterally (IL), and centrally above the patella (C), as depicted in
Fig. 1A. These placements account for the orientation of the
Langer’s lines [29], [30], [31], [32], [33], with some sensors
that align with the Langer’s lines and others that are
approximately orthogonal to them.
The device is available in three sizes (i.e., medium, large, and
extra-large). The voltage output of the sensors ranges between
0 and 5 V. Each sensor is connected via a USB connector to a
main board, which is equipped with a dedicated amplification
circuit. The circuit consists of a transimpedance amplifier
followed by a non-inverting amplifier, with potentiometers in
the feedback loop for amplification tuning. To optimize the
signal dynamic range, voltage levels are calibrated based on
sensor placement and the expected deformation patterns.
Channels M, L, and C are set to output approximately 0.5 V
when the knee is at 90° of flexion. In contrast, channels IM and
IL, which are flexed in the sleeve and extend as the knee flexes,
are adjusted to generate an output of 4.5 V when the leg is
straight. This setup ensures full use of the voltage range, hence
maximizing signal resolution. The amplified signals are then
sampled at a rate of 1,800 Hz.
B. Sensor Characterization
A 3D-printed testing rig was developed to characterize the POF
sensor output in response to F/E movements. The rig consists
of two segments connected by a joint. The POF was positioned
in custom housings to ensure smooth bending as shown in
Fig. 2.
Fig. 2. 3D-printed testing rig used for characterizing the multimodal POF
sensor's behavior during controlled F/E movements. The optical channel
is securely positioned within the rig to ensure accurate testing
conditions.
Two reflective markers per segment were attached to the rig.
Data were collected using the camera-based motion capture
system while flexing and extending the rig. The data were then
processed in MATLAB (MathWorks, Natick, Massachusetts,
USA). The sensor data were filtered with a 4th-order
This article has been accepted for publication in IEEE Transactions on Neural Systems and Rehabilitation Engineering. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2025.3540708
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
2
Butterworth low-pass filter with a 6 Hz cut-off frequency to
attenuate high-frequency noise. The motion capture data were
filtered with the same Butterworth filter settings. To validate
the sensor's response and ensure reliable data under controlled
conditions, the rig underwent multiple testing cycles. Non-
linear behaviors, such as hysteresis, are expected due to the
nature of the POF sensor. These non-linearities are expected to
become more prominent when the sensors are integrated into
the sleeve, as they must conform to the knee geometry and
hence display different bending patterns during walking. These
considerations suggest that a simple linear model is insufficient
to capture the full range of sensor responses, thus justifying the
use of an LSTM network for further analysis due to its ability
to model non-linearities and time-dependent data.
C. Experimental Acquisition
Data were collected from 31 healthy individuals (15 males;
mean age: 29.2 ± 8.3 years; mean height: 173.2 ± 8.5 cm; mean
weight: 69.5 ± 12.0 kg; mean body mass index: 22.7 ± 2.4
kg/ ) while walking on a treadmill at three different speeds:
self-selected (SS, 3.99 ± 0.34 km/h), fast (150% of SS), and
slow (50% of SS). The choice of these three speeds was made
to ensure a good representation of the variety of walking speeds
and gait patterns displayed by different individuals. Thirty
reflective markers were positioned on the lower limbs to collect
ground truth knee F/E angle data using the motion capture
system. A standard biomechanical marker set based on the
calibrated anatomical landmark technique (CAST) proposed by
Cappozzo et al [34] was used. A graphic representation of a
study participant during the data collection is shown in Fig. 1B.
Each participant was fitted with the appropriate size of the knee
sleeve. The sleeve was calibrated using the potentiometers on
the electronic circuit as described in Section II-A. All
participants wore the sleeve on the right knee. Two walking
trials per speed were recorded for a duration of 3 min each. The
experimental procedures were approved by the Mass General
Brigham Institutional Review Board (Protocol #
2016P000424).
D. Motion Capture Data Analysis
The motion capture data were processed using the Nexus
software (Vicon Motion Systems Inc, Oxford, UK). Marker
trajectories were reconstructed and labeled, and gaps were
filled. The processed files were then imported into Visual3D
(C-Motion Inc., Germantown, Maryland, USA) that was used
to implement the CAST model. The knee F/E angle, and the
heel and toe markers were exported to MATLAB for further
processing. The heel strike gait events were identified [35],
[36], and the gait cycles were segmented from one heel strike
to the next, with each segment resampled to 100 points for each
gait cycle.
Outlier detection on the segmented knee F/E angles was
performed using three complementary methods: (i) each gait
cycle was compared to the trial mean using the Pearson
correlation coefficient, with an empirically set threshold of 0.9.
Gait cycles with a correlation coefficient below this value were
considered abnormal and excluded; (ii) a range check was
applied to detect cycles with a total ROM outside the normative
range of 35° to 90° [37], ensuring that only gait-related
movements were included; (iii) Dynamic Time Warping
(DTW) [38] was employed to assess variations in knee F/E
patterns, and cycles deviating more than three Median Absolute
Deviations (MAD) from the median were considered outliers.
This step was necessary to exclude data affected by
measurement artifacts while preserving the natural variability
of knee angles in healthy subjects. The outlier detection aimed
to retain physiologically diverse but valid gait cycles, avoiding
bias towards normative data. The goal was to ensure that noisy
or unrepresentative cycles (e.g., due to marker gaps or heel
strike identification errors) did not negatively impact the
model's training and validation. Of the 27,925 total strides
recorded during the study, 802 strides were considered outliers.
Strides considered outliers based on the analysis of the motion
capture data led to the exclusion of corresponding data across
all channels.
D. Knee Sleeve Data Analysis
The output data from the five POF sensors of the knee sleeve
were exported to MATLAB, where they were filtered as
described in Section II-B. Following filtering, the data were
segmented into strides and resampled as described above for the
motion capture data.
Following the removal of outliers in a manner consistent with
the motion capture system data, an additional step of outlier
detection was conducted on the data collected with the POF
sensors to account for aberrant sensor behaviors. One challenge
in this study was the availability of only three different sleeve
sizes which led to imperfect fitting for some participants. When
the sleeve was not optimally fitted, migration of the sleeve
during walking occurred occasionally, hence resulting in signal
artifacts.
To address this issue, outlier detection on the sensor data was
performed using the same multi-method approach described for
the motion capture data. The Pearson correlation coefficient
threshold was set to 0.5 at this stage since abnormal gait cycles
had already been excluded. This less strict criterion was applied
to retain the normal variability of the signals while effectively
filtering out instances where the hardware may not have
functioned correctly. Gait cycles displaying a sensor output
range deviating from the mean (in either the positive or negative
direction) more than three times the range standard deviation
were also excluded from the analysis. The DTW method was
used with the same settings as described above for the motion
capture data. Of the 27,123 cycles selected after removing
outliers based on the motion capture data, we identified
additional 2,762 aberrant gait cycles in the knee sleeve data. It
should be noted that, if an outlier was detected in any channel,
the corresponding data for that stride was excluded from all
channels.
Signal standardization across participants presented challenges
due to the variability introduced by manual calibration (as
described in Section II-A) and anatomical differences. The
manual adjustment of each sleeve’s sensors introduced
participant-specific variations, while anatomical differences
across subjects affected how the sensors conformed to the knee,
resulting in differences in signal dynamics. Thus, a
normalization process was required.
The fast-walking trial was selected as the reference for
normalization, as it displayed the largest range of motion. To
This article has been accepted for publication in IEEE Transactions on Neural Systems and Rehabilitation Engineering. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2025.3540708
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
3
assure consistency across subjects, the signal offset at a 10°
knee flexion angle was removed from each channel, followed
by applying the scaling factor defined in (1), calculated
individually for each sensor of each participant, using the
angular displacement values obtained from the motion capture
(1)
The scaling factor was derived using the 95th and 5th percentiles
of the data rather than the absolute range. This method was
chosen because it reduces the impact of signals with range
values at the extremes of the distribution, which might not be
representative of the typical sensor behavior.
E. Model Development
LSTM networks were implemented to address the expected
non-linear behaviors affecting the sensors’ data. The first 41
gait cycles from each of the 6 trials collected from each
participant were used for model training. The processed data
from the knee sleeve were used to train LSTM networks for
each combination of the five sensor inputs. The segmented data
were concatenated and fed to the network, with the motion
capture system data serving as the ground truth.
As in previous studies that estimated knee kinematics using
sensors data [26], [39], we structured the model with an input
layer followed by a sequence of Bidirectional LSTM (BiLSTM)
layers, each coupled with dropout layers to prevent overfitting
[40]. The network architecture included a fully connected layer,
and a regression layer aimed at estimating the knee F/E angle.
To optimize the model’s performance, a Bayesian optimizer
[41] was employed for hyperparameter tuning for
configurations with all five sensor inputs. The optimized
hyperparameters were then applied to train models for all
combinations of sensor inputs to enable a consistent
comparison of the effect of different sensor numbers and
placements, while minimizing computational cost.
The model was trained using a piecewise learning schedule,
shuffling the dataset at every iteration of the dataset, for 50
iterations. The Root Mean Square Error (RMSE) per gait cycle
was used as the primary performance metric. The model was
evaluated using the Leave-One-Subject-Out Cross-Validation
(LOSO CV) technique [42].
For each model, a Kolmogorov-Smirnov (K-S) test [43] was
performed to determine if the RMSE data followed a normal
distribution. This step was essential to establish the appropriate
statistical method for comparing the performance of different
models. Based on the non-normality of the data, the median
RMSE was chosen as the primary performance metric, as it is
less sensitive to outliers and better reflects typical performance
across trials than the mean. The impact of the position and
number of sensors on estimation accuracy was also evaluated.
Various sensor configurations were tested, and their median
RMSE values were compared. From these, we selected the five
best-performing sensor configurations—one per sensor
count—based on the lowest median RMSE values.
For the five selected configurations, a total of 38,130 RMSE
data points (7,626 per sensor count) were collected across 31
participants, with each trial including 41 gait cycles and six
trials per participant. Heteroscedasticity was evaluated using
Levene’s test [44] on the medians due to the non-normal
distribution. Given the repeated measures design of the study,
with 246 points per participant, the non-normal distribution,
and the presence of heteroscedasticity, a rank-based mixed-
effects model with heteroscedasticity-consistent standard errors
(HCSE) was selected. The ranked RMSE was modeled as the
dependent variable, with the five sensor combinations as the
fixed effect, and participant as a random effect, accounting for
the interaction between participants and sensor combinations to
capture individual variability in response to each combination
of sensors’ data.
Post-hoc analysis was conducted using Dunn’s test with
Bonferroni correction [45] to explore pairwise differences
between sensor configurations. To further assess the practical
relevance of the differences between sensor configurations, we
calculated the Common Language Effect Size (CLES) [46].
CLES provided an intuitive interpretation of the likelihood that
a randomly chosen RMSE from one sensor configuration would
outperform a randomly chosen RMSE from another
configuration, helping to contextualize the differences in
performance across sensor setups. Additionally, error
distributions were analyzed across the gait cycle for the model
with lowest median RMSE to evaluate performance consistency
throughout the gait cycle.
III. RESULTS
A. Sensor Characterization
The POF sensor was characterized using the above-described
3D-printed testing rig. The results shown in Fig. 3 reveal
several key behaviors of the POF sensor. When we positioned
the element of the rig at 0° (herein referred to as “rig angle”),
the POF sensor was not straight due to the deformation induced
by the shrink tubing. Hence, we observed a reduction in the
transmitted light (i.e., a voltage drop compared to the position
in which the sensor was straight). As the angle increased to 5°,
the POF sensor straightened, hence resulting in a peak voltage
output. For rig angles above 5°, the POF sensor began to bend,
hence causing a decrease in the voltage output. A non-linear
behavior was observed as the sensor bent further, likely due to
a combination of changes in mode propagation and reflections
within the POF sensor, along with the viscoelastic
characteristics of the sensor and its conduit. These factors
contributed to signal attenuation, and at bending angles greater
than 30°, the sensor’s mechanical behavior led to a pronounced
hysteresis effect. The maximum hysteresis was about 14.6%.
These observations underscored the need for an LSTM network
to model the complex, non-linear behavior exhibited by the
POF sensors.
This article has been accepted for publication in IEEE Transactions on Neural Systems and Rehabilitation Engineering. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2025.3540708
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
4
Fig. 3. Characterization of the POF bending sensor using the 3D-printed
rig shown in Fig. 2. The plot displays the POF sensor output (y-axis) vs.
the rig angle (x-axis). The light blue lines represent the sensor output
during flexion, while the light green lines represent the sensor output
during extension. The dark blue line indicates the mean flexion
response, and the dark green line indicates the mean extension
response. The red line marks the maximum hysteresis (i.e., 14.6%).
B. Data Collected
When the optical channels are integrated into the sleeve, the
bending patterns become significantly more complex, as
illustrated in Fig. 4. During walking, two transitions between
knee F/E were observed, one occurring during the stance phase
and the other during the swing phase. These transitions are most
likely associated with hysteresis due to the viscoelastic
properties of the sensor. Moreover, the optical channels showed
a complex bending pattern when integrated into the sleeve as
the sleeve conformed to the geometry of the knee in a manner
that varied from subject to subject. This led to distinct bending
patterns across the sensors, depending on their specific position
in the sleeve. Additionally, the variations in knee geometry
among different individuals introduced further complexity,
resulting in unique sensor responses from person to person.
These complex relationships (arising from the sleeve’s design,
the dynamic movements during walking, and individual
anatomical differences) provide motivation for using an LSTM
network.
Fig. 4. Processed data from a representative participant (female, 26 years old, weight 57.5 kg, height 1.66 m) over multiple gait cycles at SS speed
(4.6 km/h). The first five plots display the scaled output of the 5 optical channels (M, L, IM, IL, C) vs. the reference knee angle. The sixth plot shows
the knee F/E angle versus the gait cycle percentage. The blue lines represent phases of knee flexion, while the green lines indicate phases of the
knee extension.
To complement the patterns shown in Fig. 4, Fig. 5 presents the
processed outputs of the same five optical channels (M, L, IM,
IL, C) across sequential gait cycles. This figure illustrates the
consistency of sensor outputs over time and provides a clear
view of how the sensor signals evolve across multiple gait
cycles. The blue lines represent phases of knee flexion, while
the green lines indicate phases of knee extension, demonstrating
stable and repeatable signal patterns across gait cycles.
This article has been accepted for publication in IEEE Transactions on Neural Systems and Rehabilitation Engineering. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2025.3540708
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
2
Fig. 5. Processed sensor outputs from a reference participant (female, 26 years old; weight 57.5 kg; height 1.66 m) during sequential gait cycles at
SS speed (4.6 km/h). The plot displays the scaled output of the 5 optical channels (M, L, IM, IL, C) against the gait cycle percentage over multiple
cycles. The blue lines represent knee flexion, while the green lines indicate knee extension.
C. Sensors Combinations Analysis
The hyperparameters explored during the optimization process
of the LSTM network with five input channels, and the
optimized values applied to all sensor configurations, are
reported in Table I.
TABLE I
HYPERPARAMETERS OPTIMIZATION
Parameter
Range
Optimized Value
#BiLSTM-Dropout layers
[1, 3]
3
#Hidden units Layer 1
[10, 250]
134
#Hidden units Layer 2
[10, 250]
26
#Hidden units Layer 3
[10, 250]
139
Dropout Factor Layer 1
[0.1, 0.5]
0.1580
Dropout Factor Layer 2
[0.1, 0.5]
0.4839
Dropout Factor Layer 3
[0.1, 0.5]
0.1283
Initial Learning Rate
[1e-4, 0.1]
0.0029
Minibatch size
[10, 256]
71
Optimizer
[adam, sgdm,
rmsprop]
adam
Gradient Threshold
[0.1 10]
1.076
L2 Regularization
[1e-5, 1e-2]
3.4253e-4
Learning Rate Drop Period
[10, 50]
24
Learning Rate Drop Factor
[0.1 0.5]
0.1797
The performance of various configurations of optical channels
was evaluated using the RMSE across the gait cycle as the
primary metric. This analysis compared the knee angle
estimates derived using the LSTM model from the knee sleeve
data with the ground truth motion capture data.
The results of the K-S test to assess whether each model’s
RMSE distribution followed a normal distribution showed non-
normality across all configurations (p < 0.001). Given the non-
normal distribution, the median RMSE and interquartile range
(IQR) for each sensor configuration were reported in Fig. 6.
Fig. 6. Median RMSE and interquartile range (IQR) associated with
different combinations of channels used to train the LSTM neural
network. The x axis represents the number of channels, the y axis the
RMSE error, and the label is the combination of channels. The median
and IQR values were calculated across all gait cycles for each sensor
configuration. The results with the lowest median RMSE are reported in
blue for each number of channels considered, while all other data points
are represented in red.
This article has been accepted for publication in IEEE Transactions on Neural Systems and Rehabilitation Engineering. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2025.3540708
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
2
Single-channel configurations revealed that the L channel
achieved the lowest median RMSE (3.88°, IQR: 2.85° – 5.50°).
Channels IM, IL and M also showed strong performance, with
a median RMSE below 4.5°, whereas the C channel had the
highest error, showing the worst performance with a median
RMSE exceeding 5.5°, suggesting that this channel may capture
less relevant or noisier information.
Using two-channel configurations improved the performance
slightly compared to single-channel setups. The combination of
M and L channels performed best, yielding a median RMSE of
3.49° (IQR: 2.58° – 4.79°). Combinations such as L and IM
followed closely with similar median RMSE values, while
those involving the C channel performed poorly.
Three-channel setups produced the best performance with the
combination of M, L, and IL achieving the lowest median
RMSE of 3.41° (IQR: 2.50° – 5.19°). Combinations involving
the C channel exhibited worse performance, underscoring its
limited contribution in multi-channel configurations.
Four-channel configurations did not provide further
improvement, with the combination of M, L, IM, and IL
achieving a median RMSE of 3.53° (IQR: 2.51° – 5.22°).
The five-channel configuration produced a slightly worse
performance, with a median RMSE of 3.91° (IQR: 2.60° –
5.82°), indicating that the addition of the non-optimal C channel
negates the benefits observed with fewer channels.
To statistically analyze the impact of channel configurations on
performance, we examined the RMSE values derived from the
configuration with the lowest median RMSE for each number
of channels. A Levene's test was performed showing significant
heteroscedasticity (p < 0.001), hence supporting the use of a
rank-based mixed-effects model with HCSE. The ranked
RMSE values were modeled as the dependent variable, with
sensor combinations as a fixed effect and participants as a
random effect, accounting for participant-specific variability.
The rank-based mixed model demonstrated significant
differences among sensor configurations (p < 0.001 for all
channel configurations). Post-hoc pairwise comparisons using
Dunn's test with Bonferroni correction revealed significant
differences between several channel combinations (Table II).
TABLE II
OUTCOMES OF DUNN’S TEST WITH BONFERRONI CORRECTION FOR THE
OPTIMAL CHANNEL COMBINATION FOR EACH NUMBER OF CHANNELS.
Comparison
Difference in Ranks
p-value
CLES (%)
1 vs 2
3071
< 0.001
56.2
1 vs 3
3713
< 0.001
56.5
1 vs 4
2681.5
< 0.001
55.6
1 vs 5
-189.5
< 0.001
51.8
2 vs 3
642
1
50.7
2 vs 4
-389.5
1
49.8
2 vs 5
-3260.5
< 0.001
46.3
3 vs 4
-1031.5
1
49.3
3 vs 5
-3902.5
< 0.001
46
4 vs 5
-2871
< 0.001
46.5
Notably, the performance of the two-, three- and four-channel
combinations were found to be not statistically different from
each other, indicating that these configurations are likely
capturing overlapping information from the knee joint. This
suggests that adding more than two channels does not
significantly change the quality of the information provided by
the sensors. However, the performance of these configurations
was found to be statistically different when compared to the
single-channel and five-channel configurations. Specifically,
the effect sizes calculated using CLES indicated an effect
between 55.6% and 56.2% for two-, three-, and four-channel
combinations compared to single-channel setups. In contrast,
the effect sizes between these configurations and the five-
channel setup were between 46% and 46.5%. This suggests that
the single-channel setups are capturing significantly different
(and more limited) information, leading to higher RMSEs,
while the five-channel setup introduces a level of redundancy
and noise, due to the inclusion of the C channel, which
minimally worsens performance compared to the optimal two,
three, and four-channel configurations.
Finally, an analysis was conducted on the best-performing
three-channel configuration (M, L, IL) across different walking
speeds. The results demonstrated that walking speed has a
negligible impact on the system’s performance, underscoring
the system's reliability in diverse ambulatory conditions.
Detailed results of this analysis are provided in the
supplementary materials (Supplementary Section 1:
Repeatability Across Waking Speeds).
D. Error Distribution Across the Gait Cycle
The distribution of knee F/E angle errors across the gait cycle
was derived for the combination of sensors with lowest median
RMSE (M, L, IL). The top panel of Fig. 7 shows the median
error and IQR values for each point of the gait cycle.
Throughout the gait cycle, the median error (represented by
the red line) fluctuates around 0°, indicating that the system's
estimation is generally unbiased. During the stance phase, the
median error remains close to 0°, indicating minimal bias in
angle estimation. During the swing phase, the median error
shows a slight increase, especially during late swing. The IQR
error (i.e., 25th-75th percentile error range) was between -5°
and 5°. Few gait cycles exhibited larger deviations. The
variability in estimation error was greater during the swing
phase of the gait cycle compared to the stance phase.
The bottom panel of Fig. 7 displays the average knee F/E angle
throughout the gait cycle. The blue line represents the reference
motion capture data, while the dashed red line indicate the angle
estimated using the sensorized sleeve. The estimated knee angle
aligns well with the reference data throughout the gait cycle,
further demonstrating the accuracy of the system.
This article has been accepted for publication in IEEE Transactions on Neural Systems and Rehabilitation Engineering. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2025.3540708
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
3
Fig. 7. Top: Boxplot distribution of knee angle estimation error across
the gait cycle using the sensor combination with lowest median RMSE
(M, L, IL) for training the LSTM neural network. Each boxplot
corresponds to a percentage of the gait cycle, with whiskers set to 1.5
times the IQR. The median error was calculated for each point of the
gait cycle. The red line represents the median error. Bottom: Mean knee
F/E angle during the gait cycle, with the solid blue line representing the
reference motion capture data, and the red dashed line indicating the
estimated angles derived from the model.
It should be noted that variability in model performance was
observed across subjects. While the data for most subjects
displayed RMSE values close to the optimal performance
(3.41°), a few individuals exhibited larger errors, highlighting
subject dependency. A detailed analysis of the RMSE
distribution per subject is provided in the supplementary
materials (Supplementary Section 3: Impact of Interpersonal
Variability).
IV. DISCUSSION
Monitoring patients with knee OA using textile-integrated
sensors is appealing in the context of tracking physical activity
and knee ROM to facilitate the clinical management of this
condition. Wearable solutions, such as the multimodal POF
bending sensors-based solution presented in this manuscript,
offer a non-invasive and unobtrusive method for real-time
monitoring during walking. This study demonstrates the
potential of these sensors to provide valuable data for knee OA
management.
The POF sensor characterization showed notable non-
linearities, due to hysteresis and mode coupling. These non-
linear behaviors, coupled with the cyclical nature of walking,
required the use of LSTM networks for estimating knee F/E
angles. The LSTM's capability to model temporal dependencies
and capture sensor output complexities appeared to be critical
to address these challenges, thus leading to accurate knee angle
estimates.
The analysis of the error distribution during the gait cycle (Fig.
7) demonstrated that the LSTM network performed consistently
well across the gait cycle, with a slight decrease in performance
during the swing phase of the gait cycle. This is likely due to
the rapid changes in angular displacement observed during the
swing phase, which seem to amplify the non-linearities in the
sensor response and hinder accurate angle estimation. However,
the estimation error was mostly limited to the range between -
5° and 5°. This is compatible with clinical application of the
sensorized sleeve. As a case in point, a reduction (compared to
healthy controls) of approximately 8° in knee flexion is
typically observed in patients with knee OA during the stance
phase [4]. This is compatible with the system’s capability to
detect changes in knee angular displacement. In fact, during the
stance phase, the 75th percentile error was approximately 2.90°,
well below the reduction in knee flexion that one would want
to detect. Outliers showed larger error values (up to 9.96°),
hence highlighting that occasionally the performance of the
knee sleeve might not be adequate.
A key factor affecting the performance of the model seemed to
be related to how the sensors conform to the geometry of the
knee of different individuals. We empirically observed that
outliers in the error distribution were associated with subjects
whose knee geometry was underrepresented in the dataset. This
is further supported by the non-normal distribution and
heteroscedasticity observed in the data, which highlight how
different knee geometries may have led to different sensor
responses and increased RMSE. This suggests that a larger,
more diverse dataset, better representing different knee
geometries, would enable clustering of similar geometries and
the development of geometry-specific models. These
personalized models would optimize estimation accuracy for
each user by tailoring the system to the individual's knee
geometry. Integrating biomechanical models could also
enhance accuracy by adding constraints that address deviations
in sensor outputs, especially for underrepresented knee
geometries. Tools like OpenCap could provide the basis for
implementation of these models [47].
Although this study focused on healthy participants to test the
proposed sensorized knee sleeve, we are ultimately interested
in using the sleeve to monitor individuals with knee OA. We
recognize that patients with knee OA often exhibit altered gait
patterns and greater variability in knee angles. The outlier
detection process used in this study was designed to filter out
noise and measurement artifacts. It may need to be adapted to
accommodate for the increased variability expected in patients
with knee OA. As part of future research, it will be essential to
validate the system in individuals with knee OA, ensuring that
the device can capture the full range of motion characteristics
that mark their altered gait. This is a crucial step toward making
the system robust enough for clinical use in managing OA,
where real-time monitoring of knee function could aid in
guiding personalized interventions.
One limitation encountered during the study was the
availability of only three sleeve sizes (medium, large, and extra-
large), which resulted in imperfect fitting for some participants.
In cases where the sleeve did not fit snugly, we observed
migration of the sleeve during walking, leading to artifacts in
the sensor data. This migration altered the position of the
sensors relative to the knee center of rotation, causing changes
in sensor output detected as outliers and removed from the
dataset. An example of this effect is provided in the
supplementary materials (Supplementary Section 2: Effect of
Sleeve Migration on Sensor Output), which demonstrates how
sleeve migration during a fast-walking trial resulted in changes
in sensor output. Notably, adjusting the sleeve for a subsequent
trial eliminated these artifacts and stabilized the signals. Future
This article has been accepted for publication in IEEE Transactions on Neural Systems and Rehabilitation Engineering. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2025.3540708
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
4
studies will benefit from the development of a customized (i.e.
subject-specific) sleeve design to reduce the likelihood of
sleeve migration and enhance signal reliability. Improving the
sleeve's fit for diverse body types will be crucial to ensure
consistent and accurate data collection across different users.
Using more than one sensor proved to enhance model
performance. Multiple sensors placed at different anatomical
locations provided complementary information, compensating
for the limitations of single sensors and improving overall
accuracy. However, the placement of sensors was critical, as the
inclusion of the sub-optimally positioned C channel reduced
performance. This finding indicates that careful consideration
of sensor placement is essential, and the C channel should not
be included in future configurations.
The best performance, with a median RMSE of 3.41° (IQR:
2.50° – 5.19°), was achieved using three sensors. While there
was no statistically significant difference between the best two-
, three- and four-sensor configurations, redundancy is still
recommended. Redundant sensors enhance the robustness of
the system against potential sensor failures. A multiple-model
approach, along with sensor failure detection logic, could be
deployed. Multiple separate models for different sensor
configurations could be trained and, in the event of sensor
failure, an alternative model could be used to maintain high
performance, even with fewer functioning sensors.
One of the challenges in deploying this device in the home and
community is signal normalization across different sensor
placements and users. In the study, we normalized the sensor
outputs by subtracting the offset value at a 10° knee angle from
all channels and applying a scaling factor to standardize the
signal range. However, this method requires knowledge of the
knee angle at 10° of flexion, which is impractical in real-world
applications. To overcome this limitation, improvements in the
electronics could be made by replacing the second (non-
inverting) amplification stage with a differential amplifier. This
differential amplifier would utilize a known baseline voltage,
acquired during a calibration procedure when the leg is straight
(0°), ensuring a consistent sensor output for all users. For signal
range normalization, a calibration procedure with two
measurement points (using sensor data when the leg is straight
and at 90° of knee flexion) could be employed. Although this
method is effective for the M and L sensors, which are pre-
flexed in the sleeve, it may pose challenges for the IM and IL
sensors, which are hyper-extended when the leg is straight and
display a flexion pattern as the knee bends. Hence, a calibration
based on capturing the sensor output at the above-mentioned
two measurement points might not capture the actual full range
of sensor output. An alternative solution could involve
collecting sensor data over multiple gait cycles to establish the
minimum and maximum sensor output, hence standardizing the
sensor output without needing absolute angle information.
While this would limit the ability to measure absolute angles, it
would allow for the detection of relative changes in angular
displacement, enabling monitoring of ROM changes, possibly
associated with joint stiffening due to knee pain. We argue that
the potential for real-world application of the proposed
technology is high. Although challenges remain, such as signal
calibration and standardization across different users, the
proposed system holds significant potential for clinical use. The
system could be used for monitoring knee OA progression,
guiding personalized interventions, and improving patient
outcomes.
A. Limitations and Recommendations for Future Work
The data collected in this study were limited to healthy
individuals, which poses questions about the generalizability of
the results to individuals with knee OA, who are known to
display aberrant biomechanics of knee F/E. Including OA
patients in future datasets would be key to improve the model's
robustness for clinical use.
Only three sleeve sizes were available, which resulted
occasionally in poor fitting and sleeve migration, thus
introducing signal artifacts. Developing customized sleeve
designs could mitigate these issues by ensuring a better fit
across different anatomical profiles. To further reduce the
impact of migration artifacts, implementing a quality-check
algorithm that monitors signal consistency and detects
deviations indicative of sleeve movement is recommended.
Higher errors observed in a subset of participants suggest that
underrepresented knee geometries in the dataset may contribute
to inaccuracies. Collecting a larger and more anatomically
diverse dataset could address this limitation. By clustering
anatomical characteristics, geometry-specific models could be
developed to enhance system accuracy for individuals with
varied knee shapes.
The system's repeatability under donning and doffing
conditions was not evaluated. Future studies should include
tests collecting data by instructing subjects to don/doff the
sleeve hence incorporating variability in the training set to
enhance the model's robustness to variations associated with
donning and doffing the sleeve.
Lastly, the sensor signal normalization method used in the
study, reliant on a controlled calibration (i.e., using data
collected for a specific knee angle), may pose challenges in
home deployments. Alternative methods have been proposed
and warrant further exploration to enable robust performance in
real-world environments.
Addressing these limitations will enhance the reliability and
applicability of the proposed system in clinical, real-world
scenarios.
V. CONCLUSION
This study highlights the potential of integrating multimodal
POF bending sensors into a wearable knee sleeve for real-time
monitoring of knee F/E patterns. We showed that the use of
LSTM networks effectively addresses the non-linearities
affecting the sensors’ output when they are integrated into a
knee sleeve. The results show that sensor redundancy improves
estimation accuracy, with a three-channel configuration
providing the best performance. The device shows promise for
real-world applications in monitoring the biomechanics of
patients with knee OA. Future work should focus on enhancing
sensor calibration methods and expanding datasets to improve
the accuracy of the system for different knee geometries.
This article has been accepted for publication in IEEE Transactions on Neural Systems and Rehabilitation Engineering. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2025.3540708
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
5
ACKNOWLEDGMENT
We would like to express our gratitude to Mitsui Chemicals
(Minato City, Tokyo, Japan) for their invaluable support in
developing the sensorized knee sleeve used in this study. We
also thank all members of the Motion Analysis Laboratory at
Spaulding Rehabilitation Hospital for their support during the
project.
REFERENCES
[1] E. Ringdahl, “Treatment of Knee Osteoarthritis,” Am
Fam Physician, vol. 83, no. 11, pp. 1287–1292, Jun.
2011.
[2] K. R. Kaufman, C. Hughes, B. F. Morrey, M. Morrey,
and K.-N. An, “Gait characteristics of patients with knee
osteoarthritis,” J. Biomech., vol. 34, no. 7, pp. 907–915,
Jul. 2001, doi: 10.1016/S0021-9290(01)00036-7.
[3] J. D. Childs, P. J. Sparto, G. K. Fitzgerald, M. Bizzini,
and J. J. Irrgang, “Alterations in lower extremity
movement and muscle activation patterns in individuals
with knee osteoarthritis,” Clin. Biomech., vol. 19, no. 1,
pp. 44–49, Jan. 2004, doi:
10.1016/j.clinbiomech.2003.08.007.
[4] I. McCarthy, D. Hodgins, A. Mor, A. Elbaz, and G.
Segal, “Analysis of knee flexion characteristics and how
they alter with the onset of knee osteoarthritis: a case
control study,” BMC Musculoskelet. Disord., vol. 14, no.
1, p. 169, Dec. 2013, doi: 10.1186/1471-2474-14-169.
[5] S. Farrokhi, M. O’Connell, and G. K. Fitzgerald, “Altered
gait biomechanics and increased knee-specific
impairments in patients with coexisting tibiofemoral and
patellofemoral osteoarthritis,” Gait Posture, vol. 41, no.
1, pp. 81–85, Jan. 2015, doi:
10.1016/j.gaitpost.2014.08.014.
[6] K. Bennell and R. Hinman, “Exercise as a treatment for
osteoarthritis:,” Curr. Opin. Rheumatol., vol. 17, no. 5,
pp. 634–640, Sep. 2005, doi:
10.1097/01.bor.0000171214.49876.38.
[7] N. J. Bosomworth, “Exercise and knee osteoarthritis:
benefit or hazard?,” Can Fam Physician, vol. 55, no. 9,
pp. 871–8, Sep. 2009.
[8] M. Fransen, S. McConnell, A. R. Harmer, M. Van Der
Esch, M. Simic, and K. L. Bennell, “Exercise for
osteoarthritis of the knee: a Cochrane systematic review,”
Br. J. Sports Med., vol. 49, no. 24, pp. 1554–1557, Dec.
2015, doi: 10.1136/bjsports-2015-095424.
[9] K. M. Leyland et al., “Measuring the variation between
self-reported osteoarthritis pain assessments,”
Osteoarthritis Cartilage, vol. 24, p. S8, Apr. 2016, doi:
10.1016/j.joca.2016.01.044.
[10] P. W. Stratford, D. M. Kennedy, and L. J. Woodhouse,
“Performance Measures Provide Assessments of Pain and
Function in People With Advanced Osteoarthritis of the
Hip or Knee,” Phys. Ther., vol. 86, no. 11, pp. 1489–
1496, Nov. 2006, doi: 10.2522/ptj.20060002.
[11] P. Bonato, V. Feipel, G. Corniani, G. Arin-Bal, and A.
Leardini, “Position paper on how technology for human
motion analysis and relevant clinical applications have
evolved over the past decades: Striking a balance between
accuracy and convenience,” Gait Posture, vol. 113, pp.
191–203, Sep. 2024, doi: 10.1016/j.gaitpost.2024.06.007.
[12] A. I. Faisal, S. Majumder, T. Mondal, D. Cowan, S.
Naseh, and M. J. Deen, “Monitoring Methods of Human
Body Joints: State-of-the-Art and Research Challenges,”
Sensors, vol. 19, no. 11, p. 2629, Jun. 2019, doi:
10.3390/s19112629.
[13] S. Prachi, A. Rohit Kumar, P. Suraj, and S. Mandeep,
“Fibre Optic Communications: An Overview,” Int. J.
Emerg. Technol. Adv. Eng., vol. 3, no. 5, May 2013.
[14] K. Peters, “Polymer optical fiber sensors—a review,”
Smart Mater. Struct., vol. 20, no. 1, p. 013002, Jan. 2011,
doi: 10.1088/0964-1726/20/1/013002.
[15] K. Cherenack and L. Van Pieterson, “Smart textiles:
Challenges and opportunities,” J. Appl. Phys., vol. 112,
no. 9, p. 091301, Nov. 2012, doi: 10.1063/1.4742728.
[16] M. Gholami, A. Rezaei, T. J. Cuthbert, C. Napier, and C.
Menon, “Lower Body Kinematics Monitoring in Running
Using Fabric-Based Wearable Sensors and Deep
Convolutional Neural Networks,” Sensors, vol. 19, no.
23, p. 5325, Dec. 2019, doi: 10.3390/s19235325.
[17] L. Bilro, J. G. Oliveira, J. L. Pinto, and R. N. Nogueira,
“A reliable low-cost wireless and wearable gait
monitoring system based on a plastic optical fibre
sensor,” Meas. Sci. Technol., vol. 22, no. 4, p. 045801,
Apr. 2011, doi: 10.1088/0957-0233/22/4/045801.
[18] D. A. B. Miller, “Waves, modes, communications, and
optics: a tutorial,” Adv. Opt. Photonics, vol. 11, no. 3, p.
679, Sep. 2019, doi: 10.1364/AOP.11.000679.
[19] J. M. Kahn, K.-P. Ho, and M. B. Shemirani, “Mode
Coupling Effects in Multi-Mode Fibers,” in Optical Fiber
Communication Conference, Los Angeles, California:
OSA, 2012, p. OW3D.3. doi:
10.1364/OFC.2012.OW3D.3.
[20] A. G. Leal-Junior, A. Frizera-Neto, M. J. Pontes, and T.
R. Botelho, “Hysteresis compensation technique applied
to polymer optical fiber curvature sensor for lower limb
exoskeletons,” Meas. Sci. Technol., vol. 28, no. 12, p.
125103, Dec. 2017, doi: 10.1088/1361-6501/aa946f.
[21] H. Sepp and S. Jürgen, “Long Short-Term Memory,”
Neural Comput, vol. 9, no. 8, pp. 1735–1780, Nov. 1997,
doi: https://doi.org/10.1162/neco.1997.9.8.1735.
[22] M. T. N. Truong, A. E. A. Ali, D. Owaki, and M.
Hayashibe, “EMG-Based Estimation of Lower Limb
Joint Angles and Moments Using Long Short-Term
Memory Network,” Sensors, vol. 23, no. 6, p. 3331, Mar.
2023, doi: 10.3390/s23063331.
[23] R. D. Gurchiek, N. Cheney, and R. S. McGinnis,
“Estimating Biomechanical Time-Series with Wearable
Sensors: A Systematic Review of Machine Learning
Techniques,” Sensors, vol. 19, no. 23, p. 5227, Nov.
2019, doi: 10.3390/s19235227.
[24] T. T. Alemayoh, J. H. Lee, and S. Okamoto, “Leg-Joint
Angle Estimation from a Single Inertial Sensor Attached
to Various Lower-Body Links during Walking Motion,”
Appl. Sci., vol. 13, no. 8, p. 4794, Apr. 2023, doi:
10.3390/app13084794.
[25] L. Tong, R. Liu, and L. Peng, “LSTM-Based Lower
Limbs Motion Reconstruction Using Low-Dimensional
This article has been accepted for publication in IEEE Transactions on Neural Systems and Rehabilitation Engineering. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2025.3540708
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
6
Input of Inertial Motion Capture System,” IEEE Sens. J.,
vol. 20, no. 7, pp. 3667–3677, Apr. 2020, doi:
10.1109/JSEN.2019.2959639.
[26] J.-S. Tan et al., “Predicting Knee Joint Kinematics from
Wearable Sensor Data in People with Knee Osteoarthritis
and Clinical Considerations for Future Machine Learning
Models,” Sensors, vol. 22, no. 2, p. 446, Jan. 2022, doi:
10.3390/s22020446.
[27] J. Sung et al., “Prediction of Lower Extremity Multi-Joint
Angles during Overground Walking by Using a Single
IMU with a Low Frequency Based on an LSTM
Recurrent Neural Network,” Sensors, vol. 22, no. 1, p. 53,
Dec. 2021, doi: 10.3390/s22010053.
[28] D. D. Molinaro, I. Kang, J. Camargo, M. C. Gombolay,
and A. J. Young, “Subject-Independent, Biological Hip
Moment Estimation During Multimodal Overground
Ambulation Using Deep Learning,” IEEE Trans. Med.
Robot. Bionics, vol. 4, no. 1, pp. 219–229, Feb. 2022, doi:
10.1109/TMRB.2022.3144025.
[29] K. Langer, “On the anatomy and physiology of the skin.
I. The cleavability of the cutis,” Br J Plast Surg, vol. 31,
no. 1, pp. 3–8, Jan. 1978.
[30] K. Langer, “On the anatomy and physiology of the skin.
II. Skin tension,” Br J Plast Surg, vol. 31, no. 2, pp. 93–
106, Apr. 1978.
[31] K. Langer, “On the anatomy and physiology of the skin.
III. The elasticity of the cutis,” Br J Plast Surg, vol. 31,
no. 3, pp. 185–199, Jul. 1978.
[32] K. Langer, “On the anatomy and physiology of the skin.
IV. The swelling capabilities of skin,” Br J Plast Surg,
vol. 31, no. 4, pp. 273–6, Oct. 1978.
[33] K. Langer, “On the anatomy and physiology of the skin:
conclusions,” Br J Plast Surg, vol. 31, no. 4, pp. 277–8,
Oct. 1978, doi: 10.1016/s0007-1226(78)90109-1.
[34] A. Cappozzo, F. Catani, U. Della Croce, and A. Leardini,
“Position and orientation in space of bones during
movement: anatomical frame definition and
determination,” Clin. Biomech., vol. 10, no. 4, pp. 171–
178, Jun. 1995, doi: 10.1016/0268-0033(95)91394-T.
[35] S. Ghoussayni, C. Stevens, S. Durham, and D. Ewins,
“Assessment and validation of a simple automated
method for the detection of gait events and intervals,”
Gait Posture, vol. 20, no. 3, pp. 266–272, Dec. 2004, doi:
10.1016/j.gaitpost.2003.10.001.
[36] J. A. Zeni, J. G. Richards, and J. S. Higginson, “Two
simple methods for determining gait events during
treadmill and overground walking using kinematic data,”
Gait Posture, vol. 27, no. 4, pp. 710–714, May 2008, doi:
10.1016/j.gaitpost.2007.07.007.
[37] J. R. Brinkmann and J. Perry, “Rate and Range of Knee
Motion During Ambulation in Healthy and Arthritic
Subjects,” Phys. Ther., vol. 65, no. 7, pp. 1055–1060, Jul.
1985, doi: 10.1093/ptj/65.7.1055.
[38] D. J. Berndt and J. Clifford, “Using Dynamic Time
Warping to Find Patterns in Time Series”.
[39] D. Hollinger, M. Schall, H. Chen, S. Bass, and M. Zabala,
“The Influence of Gait Phase on Predicting Lower-Limb
Joint Angles,” IEEE Trans. Med. Robot. Bionics, vol. 5,
no. 2, pp. 343–352, May 2023, doi:
10.1109/TMRB.2023.3260261.
[40] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever,
and R. Salakhutdinov, “Dropout: A Simple Way to
Prevent Neural Networks from Overfitting”.
[41] R. Martinez-Cantin, “BayesOpt: A Bayesian
Optimization Library for Nonlinear Optimization,
Experimental Design and Bandits,” J Mach Learn Res,
2014, doi: 10.48550/arXiv.1405.7430.
[42] D. Gholamiangonabadi, N. Kiselov, and K. Grolinger,
“Deep Neural Networks for Human Activity Recognition
With Wearable Sensors: Leave-One-Subject-Out Cross-
Validation for Model Selection,” IEEE Access, vol. 8, pp.
133982–133994, 2020, doi:
10.1109/ACCESS.2020.3010715.
[43] H. W. Lilliefors, “On the Kolmogorov-Smirnov Test for
Normality with Mean and Variance Unknown,” J. Am.
Stat. Assoc., vol. 62, no. 318, pp. 399–402, Jun. 1967,
doi: 10.1080/01621459.1967.10482916.
[44] R. J. Carroll and H. Schneider, “A note on levene’s tests
for equality of variances,” Stat. Probab. Lett., vol. 3, no.
4, pp. 191–194, Jul. 1985, doi: 10.1016/0167-
7152(85)90016-1.
[45] A. Dinno, “Nonparametric Pairwise Multiple
Comparisons in Independent Groups using Dunn’s Test,”
Stata J. Promot. Commun. Stat. Stata, vol. 15, no. 1, pp.
292–300, Apr. 2015, doi:
10.1177/1536867X1501500117.
[46] K. O. McGraw and S. P. Wong, “A Common Language
Effect Size Statistic,” Psychol. Bull., vol. 111, no. 2, pp.
361–365, 1992.
[47] S. D. Uhlrich et al., “OpenCap: Human movement
dynamics from smartphone videos,” PLOS Comput. Biol.,
vol. 19, no. 10, p. e1011462, Oct. 2023, doi:
10.1371/journal.pcbi.1011462.
This article has been accepted for publication in IEEE Transactions on Neural Systems and Rehabilitation Engineering. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2025.3540708
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
