TYPE Original Research

PUBLISHED 28 August 2024

DOI 10.3389/fbioe.2024.1417497

Predicting improvement in
OPEN ACCESS biofeedback gait training using
EDITED BY
Jing-hong Liang,
Sun Yat-sen University, China
short-term spectral features from
REVIEWED BY
Federica Verdini,
minimum foot clearance data
Marche Polytechnic University, Italy
Tianyun Jiang,
China Academy of Chinese Medical Sciences,
Nandini Sengupta 1*, Rezaul Begg 2, Aravinda S. Rao 1,
China Soheil Bajelan 2, Catherine M. Said 3,4,5,6 and
*CORRESPONDENCE Marimuthu Palaniswami 1
Nandini Sengupta,
nsengupta@student.unimelb.edu.au 1
Department of Electrical and Electronic Engineering, The University of Melbourne, Parkville, VIC,
Australia, 2Institute for Health and Sport, Victoria University, Melbourne, VIC, Australia, 3Physiotherapy,
RECEIVED 15 April 2024
Melbourne School of Health Sciences, The University of Melbourne, Parkville, VIC, Australia,
ACCEPTED 13 August 2024 4
Physiotherapy Department, Western Health, St Albans, VIC, Australia, 5Australian Institute for
PUBLISHED 28 August 2024
Musculoskeletal Science (AIMSS), Melbourne, VIC, Australia, 6Physiotherapy Department, Austin Health,
CITATION Heidelberg, VIC, Australia
Sengupta N, Begg R, Rao AS, Bajelan S, Said CM
and Palaniswami M (2024) Predicting
improvement in biofeedback gait training using
short-term spectral features from minimum
foot clearance data. Stroke rehabilitation interventions require multiple training sessions and
Front. Bioeng. Biotechnol. 12:1417497.
repeated assessments to evaluate the improvements from training.
doi: 10.3389/fbioe.2024.1417497
Biofeedback-based treadmill training often involves 10 or more sessions to
COPYRIGHT
determine its effectiveness. The training and assessment process incurs time,
© 2024 Sengupta, Begg, Rao, Bajelan, Said and
Palaniswami. This is an open-access article labor, and cost to determine whether the training produces positive outcomes.
distributed under the terms of the Creative Predicting the effectiveness of gait training based on baseline minimum foot
Commons Attribution License (CC BY). The use,
clearance (MFC) data would be highly beneficial, potentially saving resources,
distribution or reproduction in other forums is
permitted, provided the original author(s) and costs, and patient time. This work proposes novel features using the Short-term
the copyright owner(s) are credited and that the Fourier Transform (STFT)-based magnitude spectrum of MFC data to predict
original publication in this journal is cited, in
the effectiveness of biofeedback training. This approach enables tracking non-
accordance with accepted academic practice.
No use, distribution or reproduction is stationary dynamics and capturing stride-to-stride MFC value fluctuations,
permitted which does not comply with these providing a compact representation for efficient processing compared to
terms.
time-domain analysis alone. The proposed STFT-based features outperform
existing wavelet, histogram, and Poincaré-based features with a maximum
accuracy of 95%, F1 score of 96%, sensitivity of 93.33% and specificity of
100%. The proposed features are also statistically significant (p < 0.001)
compared to the descriptive statistical features extracted from the MFC
series and the tone and entropy features extracted from the MFC
percentage index series. The study found that short-term spectral
components and the windowed mean value (DC value) possess predictive
capabilities regarding the success of biofeedback training. The higher
spectral amplitude and lower variance in the lower frequency zone indicate
lower chances of improvement, while the lower spectral amplitude and higher
variance indicate higher chances of improvement.

KEYWORDS

stroke rehabilitation, biofeedback, treadmill training, interventions, machine learning,

signal processing

Frontiers in Bioengineering and Biotechnology 01 frontiersin.org

Sengupta et al. 10.3389/fbioe.2024.1417497

1 Introduction from their baseline walking data before training. Gait improvements
are identified from increased MFC within an individual-specific
Stroke affects millions of people worldwide each year (Gerstl threshold. We hypothesized that biofeedback training effects on
et al., 2023) and approximately 60,000 in Australia, i.e., more than MFC could be predicted using novel features of the Short-Term
100 documented incidents daily.1 Stroke is a prevalent and Fourier Transform (STFT) magnitude spectrum of MFC data. This
significant health risk associated with ageing, with stroke patients approach was expected to enable the tracking of non-stationary
often exhibiting impaired gait dynamics of varying severity. Stroke dynamics and capturing stride-to-stride MFC fluctuations (Pachori,
survivors with impaired gait dynamics commonly experience a 2023), providing compact representations and more efficient
higher likelihood of falls (Roelofs et al., 2023). Minimum Foot processing than time-domain analysis alone.
Clearance (MFC) is the foot’s minimum vertical displacement Our main contributions to this work are threefold:
from the walking surface during the mid-swing phase of the
walking cycle. Low MFC can increase the risk of tripping-related • We propose new STFT-based magnitude spectrum features
falls (Nagano et al., 2020; Nagano et al., 2022). Assessing the derived from MFC data to predict the effectiveness of
effectiveness of gait training requires multiple training sessions, biofeedback gait training. The mean short-term magnitude
with follow-up clinical evaluations requiring major resources. Fall spectrum reveals the tendency of stride-to-stride fluctuation in
prevention programs based on exercises have proven beneficial for MFC values, providing insights into gait stability. Analyzing
the general older adult population, but they lack effectiveness when frequency-related information helps identify patterns and
applied to stroke-injured individuals (Begg et al., 2019). In one falls irregularities in movement across strides. To the best of our
intervention study of stroke patients, a home-based balance and knowledge, this is the first study to propose short-term
strength program was trialled (Batchelor et al., 2012), and in spectral features to predict improvement in biofeedback
another, an exercise program with both group and home-based gait training from baseline data. Although histogram and
balance and strength training was conducted (Dean et al., 2012), but Poincaré features (Begg et al., 2007) and wavelet-based
neither demonstrated a reduction in falls. features (Khandoker et al., 2007) were applied to recognise
These traditional stroke rehabilitation methods are hindered by the gait patterns, these methods were used primarily in the context
absence of real-time, objective feedback, limiting patient engagement of healthy adults, and their performance in the analysis of
and impeding effective progress (Teodoro et al., 2024; Spencer et al., MFC data in stroke patients remains unknown.
2021; Giggins et al., 2013). To overcome the limitations of traditional • MFC data has dynamic patterns and is non-stationary. To enable
stroke rehabilitation, our research group has pioneered treadmill-based tracking of non-stationary MFC dynamics and to capture the
biofeedback training (Begg et al., 2014) by presenting a real-time display patterns, we first decompose the baseline MFC series into their
of the forefoot marker’s trajectory on a video monitor positioned in underlying sinusoidal structures, revealing embedded stride-to-
front of the treadmill (Begg et al., 2019; van der Straaten et al., 2020). stride fluctuations. Then, we make use of three frequency
Figure 1 shows an example of real-time biofeedback treadmill training components derived from STFT-based magnitude spectrum
and the associated MFC series. The principal biofeedback variable features to predict biofeedback-based training from baseline
derived from the forefoot marker is the MFC (see Figure 1) MFC data. The three derived frequency components
(Nagano et al., 2022). MFC at mid-swing is the critical gait variable demonstrate high statistical significance (p < 0.001), making
in predicting tripping (Nagano et al., 2022; Begg et al., 2019) with low our proposed framework superior to existing approaches.
MFC leading to unanticipated, destabilizing, foot-ground contacts • We provide a comprehensive analysis of spectral features and
(Pathak et al., 2022; Best and Begg, 2008). Stroke participants who evaluate these components in combination with machine-
have difficulty in stepping over relatively low surface irregularities of learning classification models to develop predictions from
approximately 4 cm are at increased risk of falling (Said et al., 2014), and baseline data. In addition, we provide an analysis of
they often exhibit lower and more variable MFC control across multiple windowing the MFC series and the robustness of frequency
steps (Pathak et al., 2022). components under noisy conditions.
Biofeedback gait training has been shown to be effective in
stroke rehabilitation to improve MFC data by controlling swing foot
movements (Begg et al., 2014; Nagano et al., 2022), where people can 2 Related work
receive real-time visual feedback to control MFC within the target
band, determined by individuals’ swing foot motions (Nagano et al., Previous studies have primarily focused on analyzing the linear
2022). A uniform rehabilitation program may not be suitable for all statistical properties of biomechanical variables to investigate safer
individuals, and the capacity to predict the effectiveness of gait walking and lower-limb control characteristics (Khandoker et al.,
training from pre-intervention MFC data would be highly beneficial, 2016; Begg et al., 2005). Statistical features, including mean, standard
potentially saving public health costs and reducing patient deviation (s.d.), skewness, kurtosis, median, 25th and 75th percentiles,
inconvenience. The aim of this project was to predict the interquartile range, mode, minimum, maximum, and quartile coefficient
effectiveness of biofeedback treadmill training for stroke patients of dispersion were extracted (Begg et al., 2005). They also utilized
Poincaré plots to visually represent the relationship between
successive gait cycles and provide insights into the performance of
the locomotor system in controlling critical events. From the
1 https://www.aihw.gov.au/reports/heart-stroke-vascular-disease/hsvd- Poincaré plot, they extracted features corresponding to both the
facts/contents/all-heart-stroke-and-vascular-disease/stroke major and minor axes, capturing short- and long-term variability in

Frontiers in Bioengineering and Biotechnology 02 frontiersin.org

Sengupta et al. 10.3389/fbioe.2024.1417497

FIGURE 1
The diagram depicts a person engaged in treadmill training while receiving real-time biofeedback. The training aims to regulate minimum foot
clearance (MFC) data within a target band, indicated by a red dot. The target band is defined by the (mean + SD) ±(0.5*SD) (standard deviation) of MFC,
establishing the upper and lower boundaries (Nagano et al., 2022). Adapted from Nagano et al. (2022), licensed under CC-BY 4.0 and the authors have
made no changes to the original figure.

FIGURE 2
A high-level overview of the proposed approach. Baseline MFC data is processed to extract MFC features and then fed to a classifier to predict
whether there is an improvement in the MFC or not.

the MFC data. The other research work (Khandoker et al., 2016) also stability of MFC fluctuations across strides. This information is
analyzed descriptive statistics to quantify the MFC series, including the crucial for assessing mobility and functional recovery in stroke
mean, median, standard deviation (SD), 25th percentile (Q1), 75th patients, as it might reflect gait stability and muscle coordination.
percentile (Q3), and interquartile range (IQR). In addition, the
authors also introduced tone and entropy features based on the
percentage change in successive MFC observations relative to the 3 Materials and method
previous MFC, referred to as the Percentage Index (PI).
To consider the complexity and nonstationary properties of the The MFC series offers valuable insights into foot trajectory control.
MFC series, a wavelet-based multiscale exponent to capture correlations The short-term average is suitable for characterizing MFC control in
among the variances of wavelet coefficients across different scales was stroke patients because it represents the overall intensity or strength of the
employed (Khandoker et al., 2007). The MFC series underwent MFC signal within short intervals and accommodates nonstationary
decomposition using Dabaucheis wavelets of order 6, with eight levels characteristics of the MFC series (Khandoker et al., 2007). Reduced
of decomposition, resulting in a sequential list of detailed coefficients that MFC fluctuation suggests a more stable gait; in such cases, the short-
represented the correlation evolution between the series and selected term average can effectively capture the overall intensity of the MFC
frequencies within various frequency ranges. Although these methods signal. This measure is particularly useful for quantifying the stability and
have shown success in healthy adults, their performance in analysing regularity of gait in stroke patients due to the focus on average signal
MFC data in stroke patients remains unknown. magnitude rather than time-frequency characteristics. While post-training
Short-term magnitude spectrum is valuable for observing assessments are typically used to determine any improvements during
fluctuations in MFC values in consecutive strides within smaller stroke rehabilitation, our objective here was to predict training effects. This
intervals. Its analysis of stride-to-stride fluctuations provides insights is a novel problem in stroke rehabilitation and this report is the first to
into gait characteristics, specifically regarding the consistency and address this problem, Figure 2 illustrates our approach.

Frontiers in Bioengineering and Biotechnology 03 frontiersin.org

Sengupta et al. 10.3389/fbioe.2024.1417497

TABLE 1 The table summarizes the participants’ details, including the 3.4 Spectral analysis of baseline MFC data
number of subjects, age, affected lower limb, and walking speed among
19 subjects.
The short-term average provides the overall intensity or strength
Class Improved Unimproved of the MFC signal within short intervals, which covers the
No. of subjects 14 5 nonstationary characteristics of the MFC series.
The height of the MFC refers to the vertical displacement between
Age (years) 68.71 ± 12.31 69.40 ± 11.76
the lowest point of the foot (represented by the toe marker) and the
Female 6 1 ground during the swing phase of walking (Begg et al., 2007). Then the
Left affected 5 5
series of the MFC height can be represented as in Equation 1

Walking speed (km/h) 1.96 ± 1.04 2.08 ± 0.48

MFC  {MFC1 , MFC2 , MFC3 , . . . , MFCN }, (1)

where N is the number of MFC data points (Khandoker et al., 2016).

3.1 Participants We normalize the MFC series to reduce the effects of lengthy series
and between-subject variability.
This study included 19 patients over 18 years of age at least 6 months We can now define the STFT of the MFC series as in Equation 2
after a single stroke (ischemic or hemorrhagic). They could walk
N−1
independently for 50 m and were able to provide informed consent MFCf [k, m] 
MFC[n]w[n − m]e−j N kn ,
2π
(2)
(Begg et al., 2019). Patients were excluded if they had an ankle orthosis, n0
any other medical condition that prevented them from walking on a
where w[n] is the window function, m is the shift parameter, k is the
treadmill, visual deficits, or body mass exceeding 158 kg (Begg et al., 2019),
frequency bin index, and N is the length of the signal.
participant characteristics have been presented in Table 1. Participants
The average magnitude across frames can be calculated as in
were carefully briefed and their consent was secured to ensure informed
Equation 3
participation. The study was included in the Australian and New Zealand

L−1  
Clinical Trials Registry - trial ACTRN12617000250336 and approved by MFCf (k, m)
the Human Research Ethics Committees of Victoria University, Australia MFCf (k) 
, (3)
m0
L
and Austin Hospital, Melbourne, Australia.
where L is the number of frames.
Figure 3 file shows the MFC series for improved and unimproved
3.2 Data collection patients with their frequency domain amplitude spectrum below. From
Figure 3, it is evident that the lower frequency range appears to exhibit
We employed a three-dimensional motion analysis system notable discriminant characteristics, but we can employ F-tests to more
(Optotrak®, NDI, Canada) to capture kinematic data at 100 Hz. reliably confirm the separability of data (Nicholson et al., 1997;
Following a standardized protocol (Begg et al., 2007), participants Sengupta et al., 2016). The F statistic is the ratio of the between-
were outfitted with a cluster of three active markers, including one class and within-class variance of magnitude spectrum coefficients for a
affixed to the big toe. The forefoot’s imaginary position was digitized particular frequency component. Figure 4 shows the F-ratio plot, with a
using an active digitizing probe. To ensure safety and adherence to the higher F-ratio indicating more separation between the classes. For a
protocol, all participants were secured by a safety harness and given frequency component k, the F-ratio is defined as in Equation 4
instructed to walk on a motorized treadmill at their self-selected
Ci1 Si (k) − S(k)
2
walking speed for up to 10 min, with rest breaks as needed. During F(k)  (4)
Si (k) − Si (k)
j 2
subsequent biofeedback gait training sessions, the real-time sagittal Ci1 1
Ci
Cj1
i

trajectory of the big toe marker was displayed on a screen positioned in

j
front of the treadmill (see Figure 1). This display featured toe clearance, where Si (k) is the magnitude spectrum coefficient of the j-th sample
associated MFC events and the individual patient’s training-target (j  1, 2, . . . , Ci ) of the i-th class (i  1, 2, . . . , C) at frequency k.
MFC from their baseline MFC data, depicted as a horizontal line on the Si (k) and S(k)
 are the mean of the k-th frequency component of the
screen (Begg et al., 2019). Participants were then tasked with adjusting i-th class and all the classes, respectively. We calculated F-ratios for
their MFC height to match the monitored range. Patients underwent a different frequency components to analyze the separability of classes
total of 10 biofeedback training sessions, with faded biofeedback in detail. This F-ratio analysis involved using all the MFC series
introduced after the initial six sessions. Detailed information about belonging to the two classes. The variation of the F-ratio concerning
the biofeedback training sessions is available in Begg et al. (2019). frequency is depicted in Figure 4. The short-term magnitude
spectrum of each MFC series was obtained by dividing it into
eight segments with 50% overlap using a Hamming window.
3.3 Assessment The primary finding from Figure 4 is that the short-term
magnitude spectrum in the lowest-frequency region (i.e., 0 to ≈
Gait assessment tests were scheduled at the baseline and 0.2 cycles/stride) carries more discriminative information compared
immediately after the final training session (with a minimum gap to higher-frequency regions. While there are non-zero data of
of 20 min). The class labels were based on post-training MFC change F-ratio in the higher frequency range as well, they are lower than
from baseline MFC data by which participants could be categorized those in the lower frequency zone. Therefore, it is quite evident that
as either improved or unimproved following training. the lower frequency zone is more discriminative than the higher

Frontiers in Bioengineering and Biotechnology 04 frontiersin.org

Sengupta et al. 10.3389/fbioe.2024.1417497

FIGURE 3
Typical baseline normalized MFC height series for an arbitrarily chosen stroke patient (upper panel) who improved after training and another patient
who did not improve after training (middle panel). The magnitude (absolute amplitude) spectrum was computed from the normalized baseline MFC data
above (lower panel).

FIGURE 4
The F-ratio of magnitude-spectrum coefficients demonstrates the separability between improved and unimproved.

frequency zone. The MFC data represents a discrete series since it is different frequencies. Figure 5 represents a basic diagram of the
generated at specific time points within the strides. Considering that feature extraction technique comprised of an STFT of the MFC series
each patient maintained a self-preferred consistent speed on the and then averaging across the stride axis to obtain the feature vector.
treadmill, the occurrence of MFC values at equal time intervals We chose the short-term magnitude corresponding to the lowest
enables us to define the frequency as cycles per stride. Our primary frequencies as our feature since it demonstrated the maximum
focus is on observing the changes in MFC values within each stride, differentiation between the two groups as observed in Figures 3,
rather than the time gap between consecutive values. To facilitate 4 and can be computed as MFCf (k). In the following section, we
this analysis, we employed a default normalized sampling frequency, consider the lower frequency zone as our main focus for feature
where the term ‘frequency’ refers to a normalized measure. selection because of its superior discrimination power and
prominent amplitude values.
3.4.1 Spectral feature extraction
The feature calculation involved dividing each MFC series into eight
segments with 50% overlap and applying the Hamming window. A 256- 3.5 Feature significance and selection
point Fast Fourier Transform (FFT) was performed, and by averaging the
magnitude over stride frames, we obtained 129 frequency components To determine the statistical significance of the spectral
(including the 0th frequency bin) representing magnitude values at characteristics, we conducted a Mann-Whitney U-Test (Gibbons

Frontiers in Bioengineering and Biotechnology 05 frontiersin.org

Sengupta et al. 10.3389/fbioe.2024.1417497

FIGURE 5
The figure showcases the computation of spectral features from the MFC series through the Short-Term Fourier Transform (STFT). This yields a
three-dimensional matrix that captures stride-frames, frequency, and magnitude. Averaging across stride frames provides the dominant frequency
component across strides.

and Chakraborti, 2020) on each characteristic of the two groups. For output) of xi , αi is the Lagrange multiplier of the i-th data point,
our sample (n = 19) nonparametric estimation was preferred with p K(xi , x) is the kernel function, and M represents the number of
< 0.05 a suitable threshold for significance. In addition, we support vectors—data points in the margin. The sgn(·) function
computed the AUC values for each feature to investigate their returns the sign of the argument.
importance further. ROC curves aid in developing an automatic Kernel techniques facilitate class separation by projecting data
classification model by assessing the effectiveness of each feature at points into a high-dimensional space when they are not separable in
various thresholds (Fawcett, 2004). They highlight the feature’s the lower-dimensional space. Nonlinear kernels, including
ability to maximize detection probability while minimizing false polynomial kernels, radial basis functions employed in addition
alarms. AUC was calculated from the ROC curve for individual to linear kernels (Burges, 1998).
spectral characteristics, and a support vector machine (SVM) was Decision Tree is a non-parametric supervised learning method
used for this binary classification as it showed consistent and reliable utilized for both classification and regression tasks (Mitchell, 1997).
performance in the prior studies (Begg et al., 2005; Khandoker et al., It uses tree-like structure to make decisions or predictions based on
2007). Given our smaller sample size, we adopted a precaution input features. It recursively partitions the data based on feature
against overfitting by selecting only the three most significant values, creating a hierarchical structure of decision nodes and leaf
features for classification, balancing the model’s complexity and nodes. Each internal node represents a decision based on a specific
generalizability (Murphy, 2012). feature, while each leaf node represents a class label.

1. Random Forest (RF) combines multiple Decision Trees

3.6 Predicting the improvement of MFC through the use of bagging, i.e., training each tree on a
random subset of the data and considering only a random
To predict the improvement in MFC data from baseline subset of features at each split (Breiman, 2001). The outcome is
treadmill training, we use five classifiers based on previous determined by averaging or majority voting on the predictions
literature specifically used in binary classification (Caruana and generated by these trees.
Niculescu-Mizil, 2006) and our work in this area (Begg et al., 2. AdaBoost, short for Adaptive Boosting, is a boosting algorithm
2005; Khandoker et al., 2007). These classifiers include Support that sequentially combines multiple weak learners, often
Vector Machine (SVM), Random Forest (RF), AdaBoost, Ensemble Decision Trees with only one level of depth or “stumps”
Decision Tree (EDT) and Artificial Neural Network (ANN). (Freund and Schapire, 1997). Each weak learner is trained
Support Vector Machine (SVM) is a supervised machine on a weighted version of the training data, with higher weights
learning technique that is based on guaranteed risk bounds of assigned to misclassified samples. The subsequent weak
statistical learning theory known as structural risk minimization learners focus more on the previously misclassified samples,
(SRM) principle (Burges, 1998) and it is used for both classification improving the overall performance. AdaBoost iteratively
and regression. The main function of SVM is to find an optimal updates the sample weights and combines the weak learners’
hyperplane that effectively separates data points into different predictions through weighted voting (Freund and
classes and maximizes the margin between them. The decision Schapire, 1997).
function is given by Equation 5 3. Ensemble Decision Tree (EDT) with bagging, combines
M multiple Decision Trees trained on different bootstrap
⎝
αi yi K(xi , x) + b ⎞
f(x)  sgn(w.x + b)  sgn⎛ ⎠ (5) samples of the training data. Bagging aims to reduce
i1
variance and enhance stability by introducing randomness
where, f(x) represents the decision function, w is the weight vector in the training process. Each tree in the ensemble is
perpendicular to the separating hyperplane, b serves as a bias constructed independently on a randomly drawn subset of
determining the position of the hyperplane, xi represents the i-th the training data with replacement. The final prediction is
feature vector of dimension d, yi ∈ {+1, −1} is the label (target obtained by aggregating the predictions of all the individual

Frontiers in Bioengineering and Biotechnology 06 frontiersin.org

Sengupta et al. 10.3389/fbioe.2024.1417497

trees, typically through majority voting or averaging

(Dietterich, 2000).

Artificial Neural Networks (ANNs), are computational models

inspired by the structure and function of the human brain. They
consist of artificial neurons that mimic biological neurons and are
connected through synapse-like links. ANNs are organized in layers
with connections between them. The input layer receives the data to be
modelled, and the output layer produces the predicted output. The
response produced in the output node is defined as Equation 6:
m
ok  fo ⎝
⎛
wk,j hj ⎠
⎞, (6)
j0

where ok is the produced response of the k-th node of the output

layer, fo is the non-linear function at the output layer node, m is the
number of nodes in the hidden layer, wk,j is the weight connecting
FIGURE 6
the j-th hidden node and the k-th output node, and h0  1 is the bias The figure depicts the two parameters (p-value and AUC) for
term. ANN is a supervised classifier, and the weights are determined selecting the spectral features. The upper panel represents the
significance of 11 features across lower frequencies using p-values.
during the training phase. The backpropagation algorithm and
The first 11 features consist of the DC value and the absolute
mathematical optimization techniques are used to learn the amplitudes (magnitudes) corresponding to the first 10 frequencies
weights of the connections between neurons, minimizing the (frequency range of 0–0.2545 cycles/stride). The lower panel shows
the efficacy of modelling individual characteristics regarding
difference between the expected and predicted outputs. Learning
AUC values.
the proper weights is crucial, and optimization algorithms adjust
these weights through mathematical procedures. ANNs can be used
for both classification and regression problems (Haykin, 2009).
To ensure generalization accuracy (ACC), this study used three
metrics. These metrics, including sensitivity (SENS), specificity
3.7 Model training and evaluation (SPEC), and the F1 score (F1), were calculated for each class in all
subjects (Murphy, 2012). This validation method allowed evaluation
We used the proposed STFT-based features to develop an of the model’s performance while accounting for variation between
automated classification model to identify individuals who would subjects and ensuring the ability to generalize to unseen data.
experience improvement based on their baseline MFC series as a
result of biofeedback training. Due to limited samples, with five
samples in the unimproved class (see Table 1), we approached the 4 Results
cross-validation in three ways:
We present our results and analyses in five subsections
1. Leave-one-sample-out cross-validation. Each of the 19 samples highlighting the significance of spectral features, including
was taken individually as a test sample, while the remaining statistical significance of the spectral features, comparison with
samples were used for training. This process was repeated for other MFC features, comparison of features with alternative
all 19 samples and the average performance metric was classifiers, MFC window effects and performance under noisy
calculated across all samples. conditions.
2. Leave-one-fold-out cross-validation. By randomly selecting
four samples from each class, we trained the model with
these samples while using the remaining samples for testing. 4.1 Statistical significance of the
This random selection process was repeated 50 times to ensure spectral features
reliable results. We calculated the average performance metrics
for these iterations. The MFC frequency spectrum in Figure 3 indicates that the
3. 5-fold stratified cross-validation. The data set was divided into lower frequency range exhibited greater discrimination between
five folds and the model was trained in four folds, while one- improved and non-improved classes, with differentiation
fold was reserved for testing. This process was repeated five declining as frequency increments. The same trend is seen in
times, ensuring that each fold maintained the original dataset’s Figure 6, in which the first 11 spectral components are presented
class distribution. This approach allowed for a proper with Mann Whitney p-values and AUC statistics from the SVM
proportion of samples from each class in both training and classifier (refer to Section 3.5). The consistent findings are shown in
testing sets. Following a 5-fold cross-validation procedure, we Figures 3, 6 provide strong evidence that the lower frequency range
calculated the average performance metrics. This involved contains the most valuable discriminant features.
evaluating the model’s performance in multiple iterations to The leave-one-fold-out cross-validation (refer to Section 3.7
ensure robustness and reliability. second approach) showed that the p-values of the first three

Frontiers in Bioengineering and Biotechnology 07 frontiersin.org

Sengupta et al. 10.3389/fbioe.2024.1417497

TABLE 2 Mean ± SD Values of each of three proposed spectral features, classifier due to its performance in using the features presented in
tone-entropy parameters, and MFC descriptive statistics (Khandoker et al.,
2016) for the improved and unimproved groups of stroke patients. Begg et al. (2005), Khandoker et al. (2007).
We evaluated the performance of the features using a linear
Feature Unimproved Improved p-value kernel, a radial basis function (RBF), and a polynomial kernel
Mean 0.7148 ± 0.0537 0.5164 ± 0.1282 0.005 (degree 3) kernel with SVM. Upon examining the overall
performance across all metrics, it is evident that the wavelet-
Median 0.7083 ± 0.0556 0.5068 ± 0.1316 0.007
based features (Khandoker et al., 2007) consistently
STD 0.0933 ± 0.0186 0.1351 ± 0.0319 0.014 underperformed in the three classification approaches, regardless
Q1 0.6492 ± 0.0620 0.4163 ± 0.1373 0.003
of the SVM kernel used. On the contrary, the proposed spectral
features outperformed the features based on histograms and
Q3 0.7751 ± 0.0472 0.5990 ± 0.1305 0.014 Poincaré plots (Begg et al., 2005) and wavelets (Khandoker
IQR 0.1258 ± 0.0224 0.1827 ± 0.0566 0.055 et al., 2007).
In addition, the AUCs for wavelet-based features were poor,
Tone −0.9642 ± 0.2923 −5.715 ± 4.1928 0.007
indicating that the classifier struggled to establish appropriate
Entropy 2.6811 ± 0.3523 2.6181 ± 0.3399 0.516 boundaries for these features. In contrast, both the spectral and
Proposed 16.6300 ± 1.2512 11.4983 ± 2.2320 0.0003
histogram-Poincaré features yielded comparable AUC values. On
multiple occasions, we observed a specificity of 100% for all three
All statistical comparisons were accepted as significant when p < 0.001.
Bold value represents the best performance.
types of features (histogram features, Poincaré features, and the
proposed spectral features).

features [MFCf (0), MFCf (1) and MFCf (2)] were 0.03(< 0.05),
indicating strong statistical significance, i.e., below the 4.3 Performance of proposed feature with
0.05 threshold, but p-values exceeded 0.05 after the fourth different classifiers
feature. The initial three values contributed to approximately
0.98 AUC but subsequently decreased. Table 6 provides an overview of the performance achieved by
We used statistical analysis to compare the results, similar to five machine learning models after fine-tuning their
Khandoker et al. (2016). We calculated the following descriptive hyperparameters.
statistics to quantify the MFC series: mean, median, standard We experimented with two key hyperparameters for RF and
deviation (SD), Q1 (25th percentile), Q3 (75th percentile), IQR EDT: minimum leaf size (1, 5, and 10) and the number of trees
(Q1-Q3, interquartile range); additionally, we used tone and entropy (randomly chosen between 5 and 100). In the case of the AdaBoost
estimates from the MFC Percentage Index (PI) series (Khandoker classifier, we focused on the learning rate (0.001, 0.01, and 0.1) and
et al., 2016). The Mann-Whitney U test (Gibbons and Chakraborti, the number of weak learners (15, 20, 25, and 30). Regarding the
2020) was used to determine statistically reliable differences in these artificial neural network (ANN), we varied the learning rate (0.1,
statistics between the two groups of stroke patients, that is, improved 0.01, and 0.001) and the number of hidden nodes (10, 20, and 50)
and unimproved. within a single hidden layer. We also performed hyperparameter
Table 2 displays the mean and standard deviation (std) values optimization for the ANN, including optimization functions
for all the features mentioned in Khandoker et al. (2016), as well as (Gradient Descent Backpropagation, Fletcher-Reeves Conjugate
our proposed feature for two groups: improved and unimproved. Gradient Descent, Polak-Ribiére Conjugate Gradient Descent),
On examination, it is evident that all features, except the entropy transfer functions (tan-sigmoid and log-sigmoid), a fixed number
feature, demonstrated statistical significance when considering a of epochs (1,000), and a ridge regularization value of 0.01.
threshold of p < 0.05. However, the proposed spectral feature stood Due to the limited sample size mentioned earlier in the
out as it remained significant even when considering a more classification section, we employed three cross-validation
stringent threshold of p < 0.001 (Singh, 2013). This highlights approaches to assess the effectiveness of these features.
the notable significance and efficacy of our proposed feature (all
three spectral features) in effectively distinguishing between the 4.3.1 Leave-one-sample-out approach
two groups. We observed that SVM, EDT, and AdaBoost classifiers achieved
the best performance, each with an F1-score of over 96%, and an
accuracy of 94.73%. Notably, AdaBoost achieved 100% sensitivity,
4.2 Comparison of MFC-based features while SVM achieved 100% specificity. RF also performed well with
an F1-score of 92.85%. However, ANN exhibited poor performance
Tables 3–5 present the results of the features previously used in with a specificity of 0%.
Begg et al. (2005), Khandoker et al. (2007) and our proposed spectral
feature evaluated using an SVM classifier. In Begg et al. (2005), the 4.3.2 Leave-one-fold-out method
authors used a total of 24 features, including statistical features The SVM classifier demonstrated the best performance,
extracted from the histogram representation of MFC data and achieving an F1-score of 92.95%, an accuracy of 88.18%, a
features derived from the Poincaré plot of MFC data (Begg et al., specificity of 94.00%, and a sensitivity of 87.60%. RF also
2005). In Khandoker et al. (2007), six features were derived from the performed well with an F1-score of 92.39%. Conversely, ANN,
MFC values from wavelet decomposition. We chose to use the SVM EDT, and AdaBoost classifiers performed poorly in this scenario.

Frontiers in Bioengineering and Biotechnology 08 frontiersin.org

Frontiers in Bioengineering and Biotechnology

Sengupta et al.
TABLE 3 The classification performance of the SVM classifier uses the leave-one-sample-out cross-validation method with different kernels (linear, Gaussian RBF, and polynomial) for different regularization
parameters (C).

Parameters Spectral feature Histogram-Poincaré based feature Wavelet-based feature

Kernel C ACC SENS SPEC F1 AUC ACC SENS SPEC F1 AUC ACC SENS SPEC F1 AUC
Linear 0.01 89.47 92.85 80.00 92.85 92.85 73.68 100.00 0.00 84.85 78.57 73.68 100.00 0.00 84.85 35.71

0.1 84.21 85.72 80.00 88.89 94.28 73.68 100.00 0.00 84.85 77.14 73.68 100.00 0.00 84.85 35.71

1 94.73 92.85 100 96.29 94.28 63.15 85.71 0.00 77.42 82.85 68.42 85.71 20.00 80.00 41.43

10 89.47 92.85 80.00 92.85 92.85 84.21 85.71 80.00 88.89 92.85 63.15 78.57 20.00 75.86 27.14

100 89.47 92.85 80.00 92.85 77.14 84.21 85.71 80.00 88.89 92.85 52.63 64.28 20.00 66.67 20.00

RBF 0.01 73.68 100.00 0.00 84.84 82.85 73.68 100.00 0.00 84.85 78.57 73.68 100.00 0.00 84.85 41.43

0.1
09

73.68 100.00 0.00 84.84 82.85 73.68 100.00 0.00 84.85 80.00 73.68 100.00 0.00 84.85 42.85

1 84.21 92.85 60.00 89.65 84.28 73.68 100.00 0.00 84.85 80.00 73.68 100.00 0.00 84.85 42.85

10 84.21 92.85 60.00 89.65 72.85 63.15 85.71 0.00 77.42 87.14 63.15 78.57 20.00 75.86 48.57

100 68.42 78.57 40.00 78.57 74.28 84.21 85.71 80.00 88.89 92.85 63.15 78.57 20.00 75.86 48.57

Polynomial 0.01 89.47 92.85 80.00 92.85 77.14 73.68 100.00 0.00 84.85 84.28 73.68 100.00 0.00 84.85 40.00

0.1 89.47 92.85 80.00 92.85 74.28 73.68 100.00 0.00 84.85 74.28 68.42 92.85 0.00 81.25 41.42

1 89.47 92.85 80.00 92.85 71.43 73.68 100.00 0.00 84.85 77.14 63.15 78.57 20.00 75.86 42.85

10 84.21 85.71 80.00 88.89 87.14 68.42 92.85 0.00 81.25 75.71 52.63 57.14 40.00 64.00 27.14

100 84.21 85.71 80.00 88.89 88.57 78.94 85.71 60.00 85.71 92.85 42.10 42.85 40.00 52.17 25.71

The results reported are for the three spectral features. Best performance is obtained with the spectral feature using the linear kernel (C 1): 94.73% Accuracy, 92.85% Sensitivity, 100% Specificity, 96.29% F1-score. (ACC, accuracy; SENS, sensitivity; SPEC, specificity; F1,
F1-score; AUC, AREA UNDER the ROC CURVE).

10.3389/fbioe.2024.1417497
Bold values represent the best performance.
frontiersin.org
Frontiers in Bioengineering and Biotechnology

Sengupta et al.
TABLE 4 The classification performance of the SVM classifier uses the leave-one-fold-out cross-validation method with different kernels (linear, Gaussian RBF, and polynomial) for different regularization
parameters (C).

Parameters Spectral feature Histogram-Poincaré based feature Wavelet-based feature

Kernel C ACC SENS SPEC F1 AUC ACC SENS SPEC F1 AUC ACC SENS SPEC F1 AUC
Linear 0.01 82.91 81.20 100.00 89.40 98.20 33.09 26.40 100.00 37.23 69.60 53.81 57.00 22.00 64.09 33.60

0.1 83.45 81.80 100.00 89.73 98.20 33.81 27.20 100.00 38.56 74.60 54.55 56.80 32.00 66.62 43.40

1 85.64 84.20 100.00 91.11 98.20 62.18 58.40 100.00 71.31 84.20 52.00 52.60 46.00 63.93 51.00

10 85.64 84.80 94.00 91.14 98.20 72.73 70.80 92.00 81.78 83.60 48.90 49.80 40.00 61.43 44.00

100 85.64 84.80 94.00 91.14 98.20 70.00 68.00 90.00 79.51 82.60 48.90 49.60 42.00 61.73 45.20

RBF 0.01 78.36 77.60 86.00 85.80 93.40 34.72 28.20 100.00 38.92 81.60 54.72 56.00 42.00 67.17 55.20
10

0.1 78.36 77.60 86.00 85.80 93.40 34.72 28.20 100.00 38.92 81.60 54.72 56.00 42.00 67.17 55.20

1 78.55 77.80 86.00 85.72 92.20 34.90 28.40 100.00 39.28 81.40 54.00 55.20 42.00 66.37 53.80

10 78.91 78.80 80.00 85.95 88.20 56.91 52.80 98.00 66.48 85.00 48.90 49.40 44.00 61.71 47.40

100 75.45 75.60 74.00 83.50 81.80 71.45 69.00 96.00 80.36 81.60 48.90 49.40 44.00 61.71 47.40

Polynomial 0.01 88.18 87.60 94.00 92.95 98.20 27.45 20.60 96.00 30.26 58.20 68.36 72.80 24.00 71.72 37.40

0.1 87.82 87.20 94.00 92.73 98.00 27.45 20.60 96.00 30.26 62.20 62.00 65.80 24.00 67.89 43.40

1 86.91 86.20 94.00 92.13 97.80 29.64 23.00 96.00 34.14 67.20 48.36 49.80 34.00 59.32 38.20

10 86.18 85.40 94.00 91.65 98.20 36.55 30.60 96.00 44.69 77.00 42.90 43.20 40.00 54.72 36.60

100 86.00 85.20 94.00 91.48 98.20 67.09 65.20 86.00 76.34 80.80 42.91 43.20 40.00 54.72 36.60

The results reported are for the three spectral features. Best performance is obtained with spectral feature in Polynomial kernel (C 0.01): 88.18% Accuracy, 87.60% Sensitivity, 94% Specificity, 92.95% F1-score.

10.3389/fbioe.2024.1417497
Bold value represents the best performance.
frontiersin.org
Frontiers in Bioengineering and Biotechnology

Sengupta et al.
TABLE 5 The classification performance of the SVM classifier uses the stratified five-fold cross-validation method with different kernels (linear, Gaussian RBF, and polynomial) for different regularization
parameters (C).

Parameters Spectral feature Histogram-Poincaré based feature Wavelet-based feature

Kernel C ACC SENS SPEC F1 AUC ACC SENS SPEC F1 AUC ACC SENS SPEC F1 AUC
Linear 0.01 90.00 93.33 80.00 93.14 80.00 73.33 100.00 0.00 84.57 83.33 73.33 100.00 0.00 84.57 33.33

0.1 83.33 83.33 80.00 86.47 80.00 73.33 100.00 0.00 84.57 83.33 73.33 100.00 0.00 84.57 33.33

1 95.00 93.33 100.00 96.00 80.00 73.33 83.33 40.00 80.76 83.33 75.00 93.33 20.00 84.76 33.33

10 90.00 93.33 80.00 93.14 80.00 78.33 70.00 100.00 79.33 83.33 65.00 80.00 20.00 75.62 26.67

100 90.00 93.33 80.00 93.14 80.00 83.33 76.67 100.00 85.33 83.33 60.00 73.33 20.00 70.28 26.67

RBF 0.01 73.33 100.00 0.00 84.57 80.00 73.33 100.00 0.00 84.57 90.00 73.33 100.00 0.00 84.57 46.67
11

0.1 73.33 100.00 0.00 84.57 80.00 73.33 100.00 0.00 84.57 90.00 73.33 100.00 0.00 84.57 53.33

1 85.00 93.33 60.00 90.28 80.00 73.33 100.00 0.00 84.57 90.00 68.33 93.33 0.00 80.76 53.33

10 85.00 93.33 60.00 90.28 53.33 78.33 90.00 40.00 84.76 90.00 56.67 70.00 20.00 67.62 50.00

100 71.67 73.33 60.00 70.28 53.33 88.33 83.33 100.00 89.33 73.33 56.67 70.00 20.00 67.62 50.00

Polynomial 0.01 90.00 93.33 80.00 93.14 80.00 73.33 100.00 0.00 84.57 76.67 73.33 100.00 0.00 84.57 33.33

0.1 90.00 93.33 80.00 93.14 80.00 73.33 100.00 0.00 84.57 83.33 68.33 93.33 0.00 80.76 40.00

1 90.00 93.33 80.00 93.14 80.00 73.33 100.00 0.00 84.57 90.00 70.00 86.67 20.00 80.95 33.33

10 90.00 93.33 80.00 93.14 80.00 73.33 100.00 0.00 84.57 90.00 60.00 66.67 40.00 68.47 13.33

100 90.00 93.33 80.00 93.14 80.00 78.33 83.33 60.00 83.62 83.33 53.33 66.67 20.00 64.47 13.33

The results reported are for the three spectral features. Best performance is obtained with spectral feature in linear kernel (C 1): 95% Accuracy, 93.33% Sensitivity, 100% Specificity, 96% F1-score.

10.3389/fbioe.2024.1417497
Bold value represents the best performance.
frontiersin.org
Sengupta et al. 10.3389/fbioe.2024.1417497

TABLE 6 Classification performance metrics of spectral feature with three different cross-validation approaches using five classifiers.

Classifier Approach

Leave-one-sample-out Leave-one-fold-out Five-fold-stratified-cross-

validation

ACC SENS SPEC F1 ACC SENS SPEC F1 ACC SENS SPEC F1

SVM 94.73 92.86 100.00 96.29 88.18 87.60 94.00 92.95 95.00 93.33 100.00 96.00

ANN 73.68 100.00 0.00 84.85 90.91 100.00 0.00 95.24 73.33 100.00 0.00 84.57

RF 89.47 92.85 80.00 92.85 87.27 87.00 90.00 92.39 95.00 93.33 100.00 96.00

EDT 94.73 92.85 100.00 96.29 61.45 64.00 36.00 60.95 90.00 93.33 80.00 93.14

AdaBoost 94.73 100.00 80.00 96.55 9.00 0.00 100.00 0.00 90.00 100.00 60.00 94.28

Using ACC and F1 as comparable metrics across three cross-validation approaches, SVM and RF classifiers using the proposed spectral features consistently perform well.
Metrics: Accuracy (ACC), Sensitivity (SENS), Specificity (SPEC), and F1-score (F1).

The limited number of training samples in the leave-one-fold-out 4.5 Performance under noisy conditions
approach (only four samples from each class) may have hindered the
ability of ANN, AdaBoost, and EDT to differentiate between the In real-world scenarios, foot clearance measurements obtained
classes effectively. from stroke patients undergoing treadmill therapy can be subject to
noise due to various factors. These factors include biological
4.3.3 Five-fold stratified cross-validation differences, such as variations in fitness levels and coping
All classifiers performed reasonably well except for ANN. Both abilities, which can introduce stochasticity even among patients
SVM and RF achieved the best performance with an F1-score of with similar lesions. Additionally, small and unpredictable
96%. The poor performance of ANN across all these approaches can fluctuations in the internal state, such as temporary increases in
be attributed to its inability to train effectively with the limited fatigue or shifts in effort and motivation, contribute to noise (Jin
number of samples available. et al., 2022). On the external front, minor and unpredictable
In summary, SVM demonstrated the best performance for most environmental disturbances introduce randomness into the
classification decision scenarios, while RF performance exceeded the measurements, including distractions, temperature variations, or
other classifiers. Limited samples contributed to ANN’s poorer gait perturbations caused by the moving treadmill belt. To this end,
performance, and reduced training samples can also account for we evaluated the performance of the proposed spectral features and
suboptimal discrimination using AdaBoost and EDT in the leave- baseline features (Begg et al., 2005) in the presence of white Gaussian
one-fold-out method. noise. We introduced noise to the test data while training the model
on clean data. This evaluation helps us understand the robustness of
the model and its features, considering that real-life test data can
4.4 Influence of windowing become noisy for various reasons. We added white Gaussian noise at
different levels, ranging from 0% (representing clean data) to 10%,
The presentation in Figure 6 and the statistical analysis Table 2 20%, and 30% (representing increasing noise levels).
indicate that the three selected spectral features are equally powerful Figure 7 presents the comparative results of cross-validating the
in distinguishing the two classes (p < 0.001, considering the entire leave-one-sample-out method using an SVM classifier. Spectral
database). In particular, the first component among these features features exhibited greater robustness than histogram and
corresponds to the DC value (magnitude at 0-th frequency bin) of Poincaré-based features (represented as baseline features) (Begg
the windowed MFC series. The DC value represents the average of a et al., 2005). The AUC values of the baseline features decreased
j−1
windowed segment, written as n0 w(n) · x(n), where x(n) as more noise was added to the test data, whereas the spectral
represents a frame of the MFC series, w(n) represents the features remained relatively robust across different percentages of
window function, and j is the frame length. noise. Interestingly, the spectral features declined abruptly, seen
The window function w(n) is multiplied element-wise with the clearly in specificity, when 10% noise was introduced, possibly due
corresponding samples of the frame x(n) and the resulting products to the regularization effect of noise contamination (Bishop, 1995).
are summed together to obtain the DC value for that window Considering the overall performance, we can conclude that the
segment. Surprisingly, the average DC of all frames plays an proposed spectral features outperformed the baseline features
equally important role in differentiating the two groups. Previous under noisy conditions.
studies in the literature discussed using descriptive statistical
features, including the average of the MFC series, as one of the
features (Begg et al., 2005). However, windowing followed by 5 Discussion
calculating the mean allows for localized averages within each
window, capturing more granular MFC fluctuations and proving Accurately predicting the improvement in MFC after
more significant than the global mean of the entire MFC series. biofeedback training sessions based solely on their baseline data

Frontiers in Bioengineering and Biotechnology 12 frontiersin.org

Sengupta et al. 10.3389/fbioe.2024.1417497

FIGURE 7
The figure presents performance metrics (decimal value: 0–1) with varying percentages of noise contamination in the test data, comparing the (A)
baseline (Begg et al., 2005) with the (B) proposed spectral features.

is of paramount importance in stroke rehabilitation. This predictive Table 2 reveals that the lower frequency values, including the DC
ability empowers healthcare professionals to create personalized component, were more influential in differentiating the improved
treatment plans that are tailored to each individual’s needs. By and unimproved classes.
identifying patients who are likely to benefit from such MFC values fluctuate rather than being strictly regular across
interventions, resources can be allocated more efficiently, strides. When we looked at its frequency spectrum in Figure 3, it
optimizing patient outcomes while minimizing unnecessary showed a dominant cluster of lower frequencies. These lower
expenditures. In this context, the MFC value serves as a widely frequencies reflected the underlying rhythmic dynamics that
recognized marker to determine the appropriate range of foot helped shape the fluctuations in the MFC series data. It is worth
clearance during walking (Begg et al., 2019) to prevent falls and noting that a higher DC value in the spectrum indicated that the
improve gait quality. signal amplitudes exhibited minimal fluctuations over the strides. In
Considering the potential discomfort or limitations experienced the case of the unimproved class, the higher amplitude spectrum
by stroke patients when walking on a treadmill for extended periods, showed that the fluctuations across the strides were potentially more
this study aimed to analyze MFC series data, encompassing stable/sustained over the window (across strides) for the high-
approximately 200 strides. We aimed to predict the improvement amplitude signal. On the contrary, in the case of the improving
in stroke patient’s condition, allowing targeted real-time class, the MFC values were less stable.
biofeedback training for those who would benefit the most. The Figure 8 shows the mean magnitude spectrum of the MFC series
results indicate that using the MFC series and its frequency domain across windows in all stroke patients, as well as the standard
characteristics is crucial in achieving the desired objective. deviation for the improved and unimproved classes. The
Furthermore, using features extracted from the frequency domain improved class shows more variance (or differences) in their
can contribute to developing a subject-independent model capable frequency spectrum across windows compared to the
of automatically predicting patients who will experience unimproved class. This meant that the dominant frequencies
improvement following the sessions. representing the oscillation in the MFC values across the strides
The average spectrogram of the MFC values measured over changed more and were less consistent. The unimproved class had
multiple strides provides a characterization of the average frequency less variance, and their dominant frequencies stayed more similar to
properties of MFC fluctuations. Averaging reduces variability and the MFC values across strides. A more dominant mean frequency
noise in the spectral estimates, highlighting the dominant component and lower variation in MFC values between strides
frequencies across stride sequences, with mean spectrum peaks indicate a tendency to maintain a consistent stable pattern,
revealing rhythmic patterns linked to MFC fluctuations. The potentially suggesting a lesser likelihood of improvement in
statistical analysis of the spectral features shown in Figure 6 and the future.

Frontiers in Bioengineering and Biotechnology 13 frontiersin.org

Sengupta et al. 10.3389/fbioe.2024.1417497

FIGURE 8
Figure represents the mean and standard deviation of magnitude spectrum across frames from the normalized baseline MFC data of all the
stroke patients.

The target band of biofeedback training was designed to spectral features compared to features based on the histogram,
increase and maintain the MFC within a specific range, aiming Poincaré, and wavelet.
for an elevated MFC value and reduced variation in MFC across Although the wavelet transform is generally suitable for
strides compared to the baseline. For patients exhibiting analyzing biological signals with various frequency components,
comparatively lower MFC values and higher fluctuations across it should be noted that in this study, the wavelet-based features
strides (see, Mean, STD Feature of Table 2 and Figure 8), (Khandoker et al., 2007) did not perform well. It is quite prominent
biofeedback training is found to be more effective in stabilizing that the variation of MFC values might be much higher in the case of
their MFC values by increasing the mean value and reducing older people than in the case of younger people (Khandoker et al.,
fluctuations. Conversely, when patients already have a relatively 2007). The multiscale exponent of the MFC signals, which captures
higher mean MFC and lower fluctuations, there is limited scope for correlations among variances of wavelet coefficients at different
enhancing the mean and decreasing fluctuations. This indicates scales (multiresolution), proved more effective for that purpose
that individuals who are more susceptible to tripping, (young vs. old) (Khandoker et al., 2007). However, in the context
characterized by lower MFC and greater variability in MFC of stroke patients, the variation among different scales may not be as
(Sadeghi et al., 2000; Said et al., 2014), are likely to achieve pronounced as that observed between younger and older
superior outcomes with biofeedback training within this individuals.
targeted range compared to less at risk of tripping, i.e., who The proposed features showed consistent performance across
possess higher MFC values and less variability. Regarding the most classifiers, except for the leave-one-fold-out ensemble decision
second category of stroke patients, these individuals might tree and AdaBoost (refer to Table 6). This could be attributed to the
benefit from additional training sessions or alternative target limited number of samples for training (four per class) of these
strategies to achieve successful outcomes. Another factor to classifiers. Limited diversity and representation in low sample sizes
consider could be the potential presence of spasticity in stroke might lead to overfitting, causing poor performance in ensemble
patients (Park et al., 2021) whose MFC did not improve after decision tree and AdaBoost, whereas Random Forests’ additional
completing 10 sessions of biofeedback training, indicating a randomization might mitigate these issues, resulting in better
promising area for exploration in future studies. results. SVMs might perform well in lower sample sizes with
Table 2 demonstrates the superiority of window-based spectral linear, RBF, and polynomial kernels due to their margin-based
features. In previous work, we used linear descriptive statistical optimization and regularization. In addition, the lower
features (Khandoker et al., 2016) in older adults’ MFC data, which performance observed in the ANN across each of the cross-
were calculated from the entire MFC series without considering validation approaches is likely to be attributed to the smaller
the non-stationarity of the MFC series. Tone and entropy features sample size creating overfitting. The results indicate that the
were computed from the time series of the percentage index, proposed spectral feature demonstrated good performance with
considering deviations between consecutive MFC data points. In most classifiers. In some cases, the features based on the
the present dataset, to address non-stationarity, we have used histogram and Poincaré performed equivalent to the proposed
windowing technique comprising more than two samples. spectral feature. However, regardless of the classifier used, the
Findings presented Tables 3–5 demonstrate the improved wavelet-based feature consistently exhibited lower performance
performance in stroke patients’ MFC data using short-term compared to the other two features.

Frontiers in Bioengineering and Biotechnology 14 frontiersin.org

Sengupta et al. 10.3389/fbioe.2024.1417497

6 Conclusion Author contributions

In this article, novel features using the Short-term Fourier NS: Formal Analysis, Methodology, Writing–original draft. RB:
Transform (STFT)-based magnitude spectrum of MFC data are Conceptualization, Data curation, Formal Analysis, Funding
used to predict the effectiveness of biofeedback training. With acquisition, Methodology, Project administration, Supervision,
our proposed approach, we can predict the effectiveness of real- Writing–review and editing. AR: Conceptualization, Formal
time biofeedback training for a group of stroke patients solely Analysis, Methodology, Supervision, Validation, Writing–review
based on their baseline data. The study revealed that short-term and editing. SB: Data curation, Validation, Writing–review and
spectral components and the windowed mean value (DC value) editing. CS: Data curation, Software, Validation, Writing–review
carry significant information to predict the success of and editing. MP: Funding acquisition, Investigation, Project
biofeedback training. The findings indicate that patients administration, Resources, Supervision, Writing–review and editing.
with high spectral amplitude and low variance in the lower
frequency zone are less likely to show improvement,
whereas patients with comparatively low spectral amplitude Funding
and high variance are more likely to show improvement
after training. The authors declare that financial support was received for the
Future research will focus on measuring changes after each research, authorship, and/or publication of this article. This research
training session for individual patients and identifying patterns was supported partially by the Australian Government through the
of change specifically for those who show improvement after ten Australian Research Council’s Discovery Projects funding scheme
sessions, compared to those who do not. Applying long short- (DP190101248). Data collection was supported by the National
term memory (LSTM) neural networks (Zaroug et al., 2020) to Health and Medical Research Council (NHMRC) grant
predict kinematics of lower limb trajectories during biofeedback GNT1105800.
training would be useful to forecast MFC changes following
biofeedback training session. Additionally, we plan to create a
large database of patients to further evaluate our
Conflict of interest
proposed features.
The authors declare that the research was conducted in the
absence of any commercial or financial relationships that could be
construed as a potential conflict of interest.
Data availability statement The authors declared that they were an editorial board member
of Frontiers, at the time of submission. This had no impact on the
The raw data supporting the conclusions of this article will be
peer review process and the final decision.
made available by the authors, without undue reservation.

Publisher’s note
Ethics statement
All claims expressed in this article are solely those of the authors
The studies involving humans were approved by the and do not necessarily represent those of their affiliated organizations,
Australian and New Zealand Clinical Trials Registry - trial or those of the publisher, the editors and the reviewers. Any product
ACTRN12617000250336 and approved by the Human that may be evaluated in this article, or claim that may be made by its
Research Ethics Committees of Victoria University, Australia manufacturer, is not guaranteed or endorsed by the publisher.
and Austin Hospital, Melbourne, Australia. The studies were
conducted in accordance with the local legislation and
institutional requirements. The participants provided their Author disclaimer
written informed consent to participate in this study. Written
informed consent was obtained from the individual(s) for the The views expressed herein are those of the authors and are not
publication of any potentially identifiable images or data necessarily those of the Australian Government or the Australian
included in this article. Research Council.

References
Batchelor, F. A., Hill, K. D., Mackintosh, S. F., Said, C. M., and Whitehead, C. H. randomized controlled trial protocol. Trials 20, 317–7. doi:10.1186/s13063-019-
(2012). Effects of a multifactorial falls prevention program for people with stroke 3404-6
returning home after rehabilitation: a randomized controlled trial. Archives Phys. Med.
Begg, R. K., Palaniswami, M., and Owen, B. (2005). Support vector machines for
rehabilitation 93, 1648–1655. doi:10.1016/j.apmr.2012.03.031
automated gait classification. IEEE Trans. Biomed. Eng. 52, 828–838. doi:10.1109/tbme.
Begg, R., Best, R., Dell’Oro, L., and Taylor, S. (2007). Minimum foot clearance during 2005.845241
walking: strategies for the minimisation of trip-related falls. Gait & posture 25, 191–198.
Begg, R. K., Tirosh, O., Said, C. M., Sparrow, W. A., Steinberg, N., Levinger, P., et al.
doi:10.1016/j.gaitpost.2006.03.008
(2014). Gait training with real-time augmented toe-ground clearance information
Begg, R., Galea, M. P., James, L., Sparrow, W. T., Levinger, P., Khan, F., et al. (2019). decreases tripping risk in older adults and a person with chronic stroke. Front.
Real-time foot clearance biofeedback to assist gait rehabilitation following stroke: a Hum. Neurosci. 8, 243. doi:10.3389/fnhum.2014.00243

Frontiers in Bioengineering and Biotechnology 15 frontiersin.org

Sengupta et al. 10.3389/fbioe.2024.1417497

Best, R., and Begg, R. (2008). A method for calculating the probability of tripping Nagano, H., Said, C. M., James, L., and Begg, R. K. (2020). Feasibility of using
while walking. J. biomechanics 41, 1147–1151. doi:10.1016/j.jbiomech.2007.11.023 foot–ground clearance biofeedback training in treadmill walking for post-stroke gait
rehabilitation. Brain Sci. 10, 978. doi:10.3390/brainsci10120978
Bishop, C. M. (1995). Training with noise is equivalent to tikhonov regularization.
Neural Comput. 7, 108–116. doi:10.1162/neco.1995.7.1.108 Nagano, H., Said, C. M., James, L., Sparrow, W. A., and Begg, R. (2022).
Biomechanical correlates of falls risk in gait impaired stroke survivors. Front.
Breiman, L. (2001). Random forests. Mach. Learn. 45, 5–32. doi:10.1023/a:
physiology 13, 833417. doi:10.3389/fphys.2022.833417
1010933404324
Nicholson, S., Milner, B., and Cox, S. (1997). “Evaluating feature set performance
Burges, C. J. (1998). A tutorial on support vector machines for pattern recognition.
using the f-ratio and j-measures,” in Fifth European Conference on Speech
Data Min. Knowl. Discov. 2, 121–167. doi:10.1023/a:1009715923555
Communication and Technology, Rhodes, Greece, September 22–25, 1997, 413–416.
Caruana, R., and Niculescu-Mizil, A. (2006). “An empirical comparison of supervised doi:10.21437/Eurospeech.1997-152
learning algorithms,” in Proceedings of the 23rd international conference on Machine
Pachori, R. B. (2023). Time-frequency analysis techniques and their applications. Boca
learning, Pittsburgh, PA, June 25–29, 2006, 161–168. doi:10.1145/1143844.
Raton, FL: CRC Press.
1143865Proc. 23rd Int. Conf. Mach. Learn. - ICML ’06.
Park, C., Oh-Park, M., Bialek, A., Friel, K., Edwards, D., and You, J. S. H. (2021).
Dean, C. M., Rissel, C., Sherrington, C., Sharkey, M., Cumming, R. G., Lord, S. R., et al.
Abnormal synergistic gait mitigation in acute stroke using an innovative
(2012). Exercise to enhance mobility and prevent falls after stroke: the community
ankle–knee–hip interlimb humanoid robot: a preliminary randomized controlled
stroke club randomized trial. Neurorehabilitation neural repair 26, 1046–1057. doi:10.
trial. Sci. Rep. 11, 22823. doi:10.1038/s41598-021-01959-z
1177/1545968312441711
Pathak, P., Moon, J., Roh, S.-g., Roh, C., Shim, Y., and Ahn, J. (2022). Application of
Dietterich, T. G. (2000). An experimental comparison of three methods for
vibration to the soles reduces minimum toe clearance variability during walking. Plos
constructing ensembles of decision trees: bagging, boosting, and randomization.
one 17, e0261732. doi:10.1371/journal.pone.0261732
Mach. Learn. 40, 139–157. doi:10.1023/a:1007607513941
Roelofs, J. M., Zandvliet, S. B., Schut, I. M., Huisinga, A. C., Schouten, A. C.,
Fawcett, T. (2004). Roc graphs: notes and practical considerations for researchers.
Hendricks, H. T., et al. (2023). Mild stroke, serious problems: limitations in balance
Mach. Learn. 31, 1–38.
and gait capacity and the impact on fall rate, and physical activity. Neurorehabilitation
Freund, Y., and Schapire, R. E. (1997). A decision-theoretic generalization of on-line neural repair 37, 786–798. doi:10.1177/15459683231207360
learning and an application to boosting. J. Comput. Syst. Sci. 55, 119–139. doi:10.1006/
Sadeghi, H., Allard, P., Prince, F., and Labelle, H. (2000). Symmetry and limb
jcss.1997.1504
dominance in able-bodied gait: a review. Gait & posture 12, 34–45. doi:10.1016/
Gerstl, J. V., Blitz, S. E., Qu, Q. R., Yearley, A. G., Lassarén, P., Lindberg, R., et al. s0966-6362(00)00070-9
(2023). Global, regional, and national economic consequences of stroke. Stroke 54,
Said, C. M., Galea, M., and Lythgo, N. (2014). Obstacle crossing following stroke
2380–2389. doi:10.1161/strokeaha.123.043131
improves over one month when the unaffected limb leads, but not when the affected
Gibbons, J. D., and Chakraborti, S. (2020). Nonparametric statistical inference. limb leads. Gait & Posture 39, 213–217. doi:10.1016/j.gaitpost.2013.07.008
Chapman and Hall/CRC. doi:10.1201/9781315110479
Sengupta, N., Sahidullah, M., and Saha, G. (2016). Lung sound classification using
Giggins, O. M., Persson, U. M., and Caulfield, B. (2013). Biofeedback in rehabilitation. cepstral-based statistical features. Comput. Biol. Med. 75, 118–129. doi:10.1016/j.
J. neuroengineering rehabilitation 10, 60–11. doi:10.1186/1743-0003-10-60 compbiomed.2016.05.013
Haykin, S. (2009). Neural networks and learning machines, 3/E. New Delhi, India: Singh, P. (2013). P value, statistical significance and clinical significance. J. Clin. Prev.
Pearson Education India. Cardiol. 2, 202–204.
Jin, Y., Sano, Y., Shogenji, M., and Watanabe, T. (2022). Fatigue effect on minimal toe Spencer, J., Wolf, S. L., and Kesar, T. M. (2021). Biofeedback for post-stroke gait
clearance and toe activity during walking. Sensors 22, 9300. doi:10.3390/s22239300 retraining: a review of current evidence and future research directions in the context of
emerging technologies. Front. Neurology 12, 637199. doi:10.3389/fneur.2021.637199
Khandoker, A. H., Lai, D. T., Begg, R. K., and Palaniswami, M. (2007). Wavelet-based
feature extraction for support vector machines for screening balance impairments in the Teodoro, J., Fernandes, S., Castro, C., and Fernandes, J. B. (2024). Current trends in
elderly. IEEE Trans. Neural Syst. Rehabilitation Eng. 15, 587–597. doi:10.1109/tnsre. gait rehabilitation for stroke survivors: a scoping review of randomized controlled trials.
2007.906961 J. Clin. Med. 13, 1358. doi:10.3390/jcm13051358

Khandoker, A. H., Sparrow, W. A., and Begg, R. K. (2016). Tone entropy analysis of van der Straaten, R., Tirosh, O., Sparrow, W. A. T., and Begg, R. (2020). Effects of
augmented information effects on toe-ground clearance when walking. IEEE Trans. visually augmented gait training on foot-ground clearance: an intervention to reduce
Neural Syst. Rehabilitation Eng. 24, 1218–1224. doi:10.1109/tnsre.2016.2538294 tripping-related falls. J. Appl. biomechanics 36, 20–26. doi:10.1123/jab.2018-0291

Mitchell, T. (1997). Decision tree learning. Mach. Learn. 414, 52–78. Zaroug, A., Lai, D. T., Mudie, K., and Begg, R. (2020). Lower limb kinematics
trajectory prediction using long short-term memory neural networks. Front. Bioeng.
Murphy, K. P. (2012). Machine learning: a probabilistic perspective. MIT press. Biotechnol. 8, 362. doi:10.3389/fbioe.2020.00362

Frontiers in Bioengineering and Biotechnology 16 frontiersin.org