Measurement 165 (2020) 108184

Contents lists available at ScienceDirect

Measurement
journal homepage: www.elsevier.com/locate/measurement

Foot type classification using sensor-enabled footwear and 1D-CNN

Zhanyong Mei a, Kamen Ivanov b,c, Guoru Zhao b, Yuanyuan Wu a, Mingzhe Liu a, Lei Wang b,⇑
a
College of Cyber Security, Chengdu University of Technology, Chengdu 610059, China
b
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen 518055, China
c
Shenzhen College of Advanced Technology, University of Chinese Academy of Sciences, Shenzhen 518055, China

a r t i c l e i n f o a b s t r a c t

Article history: Poor selection of footwear, underestimation of foot health, sedentary life, and lack of accessible foot
Received 30 March 2020 screening can have significant long-term adverse effects on the health of lower limbs. Unobtrusive, per-
Received in revised form 30 June 2020 vasive methods for automated foot screening have the potential to allow for timely detection of foot
Accepted 2 July 2020
abnormalities. In the present study, we describe a proof-of-concept where data collected through
Available online 6 July 2020
sensor-enabled insoles and processed through one-dimensional convolutional neural networks were
used to distinguish normal, cavus, and planus feet. We explored several combinations of sensor modal-
Keywords:
ities to find the one that reflects foot types optimally. The highest accuracy of classification of 99.26% was
Foot type classification
Sensor insole
achieved when angular velocity and force sensing were combined. Based on results, we suggest that
1D CNN sensor insoles, combined with optimal classification techniques, could be used for foot screening.
Inertial sensor Ó 2020 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
Force sensor (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction appropriate for screening due to the associated radiation exposure.

Specialized systems for gait analysis, such as force plates [9], plan-
According to the structure of the foot arch, there are three types tar pressure platforms [10], and motion capturing systems using a
of feet: normal, planus, and cavus foot. The planus foot exhibits a camera and optical makers [11], allow for accurate classification or
wholly or partly collapsed arch, whereas cavus foot characterizes clustering of the foot types. For instance, a camera-based Vicon MX
with a high longitudinal arch. Both of these abnormal foot types Motion Capture System was used to evaluate the effect of foot
are highly prevalent. The proportion of the population with flat feet orthoses for the correction of flexible flat foot [12]. The 3D foot
is about 2.2% among all age groups [1]. The low arch foot is found shape, acquired using a 3D laser scanner, can also indicate the loca-
in 10.3% of the population in the age range 7–14 [2]. The preva- tion and severity of abnormal foot regions [13].
lence of cavus foot is about 10.5% in the age group 16–65 years However, these are stationary systems confined to use in labo-
[3]. Due to the gait alteration that occurs as a result of these foot ratory settings and not being suitable for mass foot screening and
deformities, they are also associated with a higher incidence of long-term monitoring. Therefore, new types of devices are needed
lower limb injuries and increased risk for the development of med- to serve such purposes. Our primary concern behind conducting
ial tibial stress syndrome and patellofemoral pain [4]. Individuals this research was to design and validate a method for unobtrusive
with cavus foot may have unstable gait, ankle sprains, or pain in screening and monitoring of foot function based on a sensor-
the metatarsal heads [6]. With the progression of the cavus foot enabled insole. In the current study, we integrated inertial sensor
condition, surgery will be necessary [5]. In the case of inappropri- and force sensors into an insole. In this design, the inertial sensor
ately selected shoes or foot orthotics, the condition of feet and allows acquiring the kinematic information, and the force sensors
ankles will worsen [7]. capture the kinetic information. The fusion of those two kinds of
Determining foot status requires morphological measurement, information allows reflecting the motion characteristics of each
which typically involves anthropometric methods, visual inspec- foot type entirely. Besides, our design involves wearable system
tion, footprint measurement, and radiographic evaluation [8]. technology. It thus facilitates experiments outside laboratories, in
When applying such techniques, the validity of the measurement natural daily life settings, that are very close to the real conditions
heavily depends on the experience of the podiatrist or orthopedist. of mass screening of the foot types.
Also, despite its diagnostic value, X-ray imaging of the foot is not Gait analysis takes place for analysis of locomotion ability,
clinical evaluation, and monitoring of the daily activities of the
patients [14]. For the sensing, force sensors, and inertial sensors
⇑ Corresponding author.
incorporating accelerometer and gyroscope are typically used.
E-mail address: [email protected] (L. Wang).

https://doi.org/10.1016/j.measurement.2020.108184
0263-2241/Ó 2020 The Authors. Published by Elsevier Ltd.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
2 Z. Mei et al. / Measurement 165 (2020) 108184

Features can be explicitly defined, i.e., handcrafted, or, when using surements with an insole pressure system Pedar-X, it was deter-
the latest machine learning approaches, be automatically deter- mined that the contact area in the medial midfoot, as well as the
mined. While conventional machine learning methods heavily maximum force and peak pressure in the lateral forefoot, are dif-
depend on the handcrafted features, the latter might not reflect ferent between the normal foot and planus foot [20].
essential characteristics appropriately and thus lead to bias and Kerr et al. [32] captured motion data using a tracking system
lower the performance of the classifier. Recently, one- with reflective markers and infrared cameras. They found differ-
dimensional convolutional neural networks (1D CNN) have gained ences in the inversion-eversion of the hindfoot and abduction–ad-
popularity for their unprecedented ability to extract discriminative duction of the forefoot between symptomatic and asymptomatic
features from raw data automatically. They have successfully been flat feet. Also, Buldt et al. [27] analyzed kinematic data captured
applied for human activity recognition [15], extraction of gait from a camera-based motion capture system using optical mark-
parameters [16], and classification of gait type [18]. In the present ers, and found that during initial contact and mid-stance the cavus
study, we hypothesized that during the foot motion, each foot type foot exhibited less motion in the sagittal and transverse planes;
exhibits characteristic ‘‘marks” in the sensor signals. To explore the also, lower midfoot range of motion during the pre-swing of planus
validity of this assumption empirically, we developed a full wear- foot. Kruger et al. [28] also analyzed the kinematic characteristics
able gait analysis framework. It involves a multi-modal sensor of foot and ankle among normal, planus, and cavus feet. For that,
insole for signal collection and a set of 1D-CNNs neural networks the differences in the tibial angle, hindfoot angle, and forefoot
for automatic classification of foot types. Our aims were: angle were observed among the three foot types throughout the
gait cycle. Differences were also found in the coronal forefoot
(1) To prove whether our designed sensor insole could serve angular velocity.
foot type classification. As to the automatic classification of the foot types based on
(2) To construct 1D-CNNs, optimize the network parameters, plantar sensing, some progress has already been made through sta-
and compare the performance of the classifiers constructed tionary devices. Barton et al. [24] classified foot types into normal
using individual modalities and combinations of sensor data. foot, cavus foot, and hallux valgus using the maximum pressure
(3) Based on the performance comparison, to determine the during the stance phase. Xu et al. [25] extracted from plantar pres-
optimal sensor configuration that would allow for a high sure four foot-arch indexes as features: Staheli index, Chippaux-
accuracy while using a lower volume of gait data. Smirak index, arch index, and modified arch index. They used a
(4) To challenge the benefits of deep learning methods for foot classifier combining fuzzy logic and neural network to classify
type classification by comparing them with conventional cavus foot and planus foot. Ramirez-Bautista et al. [48] divided
machine learning methods. the plantar pressure area into fourteen sub-areas, including five
toes, five metatarsal heads, lateral and medial midfoot, and lateral
The rest of this work is organized as follows: In Section 2, the and medial heel. Cavus foot and planus foot were classified by the
related works on foot type evaluation and classification are intro- use of fuzzy cognitive maps with a genetic algorithm for weight
duced. Section 3 provides information about the design of the cus- learning. For the classification of the normal foot and flexible pla-
tom smart insole and the acquisition of experimental data. The nus foot based on the ground reaction force, Bertani et al. [9] used
algorithms used in this work are described in Section 4. In Sec- a heuristic optimization technique named Discard-Insert-Exchange
tion 5, the results of the algorithms and performance comparison to extract the most discriminative features automatically. In this
are presented. Discussion and conclusion are given in Sections 6 method, three linear discriminative functions (LDEs) were used,
and Section 7, respectively. where each function maps one of the GRF components in the form
of a one-dimensional feature vector. A fourth LDF was used to
obtain the final foot type classification result. Böhm et al. [11] used
2. Related work principal component analysis to extract the three independent
eigenvectors from the kinematic data of flexible flatfoot. Then,
Each foot type is associated with a specific locomotion pattern. the K-means algorithm was applied to cluster transformed data
The exploration of the motion characteristics of each foot type into two types of flexible flatfoot. In this process, the inversion data
allows building methods for automatic recognition of the foot type. during the push-off phase were the most discriminant. Table 1
Technically, the data required for such analyses are obtained contains a list of the main works dedicated to foot type classifica-
through plantar pressure measurement systems, force plates, and tion using gait analysis devices.
optoelectronic motion capture systems. So far, there was no study to classify foot types using inertial
Based on a plantar pressure measurement system emedÒ-x 400, sensing. However, some authors studied the locomotion character-
Buldt et al. [19] found that the lowest average and minimum cen- istics of each foot type, Grech et al. [26] found that there was a sig-
ter of pressure (CoP) velocity during the terminal stance phase nificant difference in the acceleration at heal strike between cavus
were lower in cavus foot compared to planus foot. Cavus foot foot and normal foot, as well as between cavus foot and planus
had the highest maximum CoP velocity during the pre-swing foot. In another study [31], no difference was found in the acceler-
stance [19]. The planus foot had the smallest index range of ation between the normal foot and planus foot. Also, in a compar-
lateral-medial force during the terminal phase [19]. Among normal ison between cavus foot and planus foot difference was observed
foot, planus foot and cavus foot, the cavus foot exhibited the high- in forefoot adduction excursion and forefoot abduction velocity
est center of pressure excursion index (CPEI); the planus foot [30].
showed the smallest CPEI values, and the normal foot exhibited The above studies of motion patterns of each foot type confirm
the lowest peak pressure [17]. The peak pressures under the fourth that foot types can be differentiated based on kinematic and
and fifth metatarsophalangeal joints of the planus foot are lower kinetic characteristics. We thus hypothesized that obtaining those
than the ones of normal foot and cavus foot [29]. In previous stud- two kinds of information would allow for foot type screening.
ies, the stability of planus and normal feet was evaluated based on Hence, in this study, we suggest a proof-of-concept of a wearable
kinetic data from force plates. The force under the first hallux of system with integrated inertial and force sensors that can provide
pes planus is higher than that of normal foot [23]. In that, signifi- kinematic and kinetic information of foot motion, respectively. In
cant differences were found upon quantification through sample previous studies, one-dimensional convolutional neural network
entropy [21], CoP velocity, and total velocity [22]. Based on mea- (1D-CNN) has been successfully applied for processing motion sen-
 Z. Mei et al. / Measurement 165 (2020) 108184 3

Table 1
The devices and methods related to foot type classification.

Authors (Ref.) Platform Foot types Samples # (training#, Models Result

testing#)
Barton et al [24] plantar pressure Normal, cavus, hallux (28,13) neural network 92%
platform valgus (10,8) 100%
(EMED F)
Xu et al [25] plantar pressure cavus, planus (300,200) adaptive neuro-fuzzy inference 95%
platform system
Bertani et al [9] force plate (Kistler) Normal, planus 112(leave- LDF-LC 84.8%
One-out test)
Böhm et al [11] Motion capture system Two types of flexible 255 k-means effective
(Vicon Nexus) flatfeet cluster
Ramirez-Bautista et al plantar pressure Cavus, planus 151 GA-FCM 91%
[48] platforms
(FreeMed)

sor time-series, in, e.g., golf swing classification [33], automatic locomotion type is walking at normal speed. The electronic control
detection of freezing gait of Parkinson disease [34] and real-time board, also holding the inertial sensor, was attached to the top of
activity classification [35]. CNNs outperform conventional machine the shoe over the forefoot area. Force sensors allowed capturing
learning methods based on handcrafted features thanks to their kinetic information; inertial sensor provided kinematic informa-
ability to automatically extract features from the low to the high tion. The combination of the two modalities was expected to offer
level of representation. Based on this evidence, we considered the highest accuracy of foot type classification. Sensor data were
combining the wearable gait analysis system with a 1D CNN for transferred wirelessly using Bluetooth technology to a custom
the extraction of discriminative features. The suggested concept receiver connected to a personal computer. Detailed information
could eventually reflect the functional differences among the three about the developed insole can be found in [36,37].
types of foot. In practice, such systems could allow for real-time
monitoring of foot status and early warning of foot deformities.
3.2. Data acquisition
To the best of our knowledge, the proposed concept is the first
one for foot type classification based on a wearable system.
The data collection procedure was conducted in accordance
with the Declaration of Helsinki. Each subject was informed about
3. Experimental platform and data acquisition the experimental procedure, and his/her agreement to participate
in the experiment was obtained. A total of 80 subjects were
3.1. Experimental platform recruited for the experiment. Only data of the left feet were pro-
cessed in this study. Among all subjects, there were 44 subjects
For the present study, we developed a full data acquisition with normal feet, 16 with planus feet, and 20 with cavus feet.
framework, including a custom sensor insole. Fig. 1 illustrates the The type of each foot was determined and labeled by a podiatrist.
prototype of the insole and its main components. Subjects’ demographics data are listed in Table 2. Subjects with
The device incorporates nine force sensors of type FlexiForce pain or injury in a foot, ankle, or knee, or who underwent surgery
A301 (Tekscan, Inc., USA) and an inertial sensor of type BMI160 in these areas within the last six months before the data collection,
(Bosch Sensortec GmbH). Force sensors were attached to a thin were not recruited. Before the experiment, each subject was
and flexible printed circuit board and located under the main informed about the experimental procedure and asked to fill out
weight-bearing areas of the foot, namely the hallux, five metatarsal a questionnaire about age, gender, height, weight, and health sta-
heads, lateral midfoot, the lateral heel and medial heel. The inertial tus related to the locomotor function. The latter required reporting
sensor contains a three-axis accelerometer and gyroscope. All sen- of pain in the heel, mid-foot, forefoot or hallux, or problems in the
sor signals were sampled at 100 Hz, a rate acceptable when the back, knee, or hip, as well as a history of falls. For each foot, photos

Fig. 1. The custom insole (a) implemented prototype (b) components.

4 Z. Mei et al. / Measurement 165 (2020) 108184

Table 2
Demographic characteristics of the subjects.

Foot type Gender (male/female) Age (years old ± SD) Height (cm ± SD) Weight (kg ± SD) Foot size (cm ± SD)
Normal (44) 31/13 21 ± 1.20 169.1 ± 8.3 61.7 ± 10.3 25.1 ± 1.3
Cavus foot (20) 15/5 21 ± 0.78 169.9 ± 5.9 61.0 ± 9.2 24.9 ± 1.1
Planus foot (16) 10/6 21 ± 0.78 168.7 ± 8.6 63.9 ± 10.5 25.1 ± 1.1
Total (80) 56/24 21 ± 1.19 169.2 ± 7.8 60.6 ± 10.1 25.0 ± 1.2

SD: standard deviation.

were taken of the medial longitudinal arch to examine the arch, the benefits of the CNNs: the local connection, shared weights,
and of the heel and tibia to examine the foot alignment during sta- pooling, and the use of multiple layers [38]. The typical structure
tic standing. Each subject wore shoes with inserted sensor soles of of a CNN involves stacked convolutional layers, a pooling layer,
an appropriate size. Data were acquired while each subject walked and fully connected layers. Each convolutional layer represents a
straight over a 7-meter walkway completing in this way seven or feature abstraction in space and/or time domain of the input data
eight steps until arriving at the end of the walkway. Each trial cov- or the previous feature maps. The stacked convolutional layers lead
ered the phases of gait initiation, gait stabilization, and gait termi- to progressively extracted feature maps from the low-level
nation. We hypothesized that collected data contain discriminative abstraction to the high-level ones. Each pooling layer downsam-
information for foot type classification. During the test, sudden ples the features from the preceding maps without loss of informa-
changes in the gait were not allowed. Each subject performed tion and, thus, reduces the dimension of the preceding map. Max
approximately twenty trials. The data of some trials were consid- pooling is generally used [38]. The activation function performs
ered invalid due to operator or subject mistakes. After discarding the non-linear mapping from the output of the previous neuron
invalid trials, the data of 1488 trials were deemed valid. These of the convolutional layer; this allows the network to learn a com-
included 827 records of normal feet, 343 of cavus feet, and 318 pla- plex model [39]. Non-saturated activation functions, e.g., rectified
nus feet records. The experimental configuration is illustrated in linear unit (ReLU), which solve exploding or vanishing gradient,
Fig. 2. The waveforms of a valid representative trial are shown in and speed up convergence, are widely used in deep learning [40].
Fig. 3. To determine the appropriative activation function, several satu-
rated activation functions, including ReLU, Leaky ReLU (LReLU),
parametric ReLU (PReLU), randomized ReLU (RReLU), respectively,
4. Data analysis
were explored in this study.
In the present study, to address the problem of the covariate
4.1. 1D CNNxxx
shift, batch normalization was used in the 1D CNN architecture.
The normalization accelerates the speed of network training and
CNN is a feed-forward neural network, inspired by the physio-
relieves the need to take care of the parameter initialization [41].
logical structure and signal processes of the human visual cortex.
1D-CNN, 2D-CNN, and 3D-CNN are used to process 1D time series,
2D images, and image series, respectively. Four properties allow 4.2. The proposed architecture

The basic processing model (BPM) that defines the 1D CNN

architecture applied in this study is presented in Fig. 4 (a), and
we applied it for all variants of input data. Data were first normal-
ized with Z-scores and then fed to the corresponding BPM. Each
module consisted of two convolutional layers and one max pool
layer; each layer was followed by batch normalization and activa-
tion. The architecture of the 1D-CNN that fused all three modalities
is shown in Fig. 4(b). Four modules were stacked during the stage
of feature learning. Then the learned features from the three pipeli-
nes were concatenated and passed to a network consisting of a flat-
ten layer, two fully connected layers, and one softmax layer with
softmax activation. For every single type of data, there was no con-
catenating step for feature fusion.
We aimed to investigate which individual type of data and
which fusion scheme would serve foot type classification better.
For that purpose, we trained an individual model using accelera-
tion, angular velocity, and force data, respectively. Then, we
designed four feature fusion schemes: (1) acceleration and angular
velocity, (2) acceleration and force (3) angular velocity and force,
and (4) accelerometer, angular velocity, and force.
To find an appropriate pool size, we tested the sizes from 1  3
to 1  8. We also explored convolutional kernel sizes from 1  3 to
1  7, tested several activation functions, and batch sizes from 16
to 64 with an interval of 8. Because of the restricted computing
power, we were not able to find a global optimization of the
parameters and, as a workaround, adopted a greedy search. First,
we adjusted the convolutional kernel size; then, based on the opti-
mized kernel size, we adjusted the pool size. We adopted a similar
Fig. 2. Experimental configuration. strategy to adjust the activation function and the batch size.
 Z. Mei et al. / Measurement 165 (2020) 108184 5

Fig. 3. Waveforms of a valid representative recording.

4.3. Model training Windows 10. Additionally, a GPU NVIDIA TITAN Xp, with 12 GB
GDDR5X and 3840 cores was used to accelerate the processing.
We trained each of the models for 100 epochs. Categorical
Cross-Entropy served as a loss function, and Adadelta optimizer
4.5. Comparison with conventional machine learning methods
was applied to update the model parameters. A specific character-
istic of the Adadelta optimizer is that the learning rate decreases
A third-order Butterworth low pass filter with a cutoff fre-
monotonically with the progression of learning [42]. The accuracy
quency of 16 Hz was applied to each of the three types of data
and loss of training and testing were saved for each epoch, and the
[14]. The handcrafted features obtained from insole data were
model of the last epoch was saved for the prediction.
the maximum, minimum, mean, range, zero-crossing rate, root-
5-fold cross-validation: To evaluate the generalization capabil-
mean-square value, variance, standard variance, skewness, and
ity of the 1D-CNNs and the conventional models for foot type clas-
kurtosis. For the comparison, in the role of conventional methods,
sification, we performed a stratified 5-fold cross-validation, in
Random forest, support vector machine (SVM), and decision tree
which each fold kept the same proportion of samples for each
were used to classify the foot types based on the handcrafted
given category. For that, all samples were partitioned into five sub-
features.
sets. One subset of data was used to test the trained model; from
Random forest is an ensemble learning method suitable to pro-
the rest subsets, 20% were used for validation and 80% for training.
cess foot motion data [44]. It consists of a collection of trees, each
This process was repeated five times until using each subset as a
of which builds on randomly sampled data that have the same sta-
test set. The performance of each classifier was evaluated based
tistical distribution as ones of all other trees in the forest. For clas-
on the resulting confusion matrix. For each confusion matrix, each
sification in the forest tree, voting is typically used to determine
row represents the instances of the actual class, while each column
the class labels. We used a grid search to determine optimal values
represents the ones of the predicted class. Four average metrics:
of the number of trees, maximum depth, and the number of fea-
recall, precision, f-measure, and accuracy, obtained from the
tures to split each node [47].
matrix, were also used for evaluating the performance of the mod-
The SVM is a supervised machine learning technique. It is based
els. The 5-fold cross-validation was performed over each of the
on the statistical learning theory of Vapnik–Chervonenkis dimen-
input data variants.
sion and structural risk minimization. An SVM constructs the
hyperplanes with maximum margin to separate the data into dif-
ferent classes [46]. A kernel function could be introduced to project
4.4. Technical specification the data into a higher dimension, which allows making the data
linearly separable. Among kernel functions, SVM with radial basis
The algorithm was implemented using Python 3.6 and Keras function showed the best accuracy in this study. The parameter
2.2.4 with TensorFlow-GPU 1.12.0. It was executed on a computer of the decision function shape was selected as ‘‘one against the
provided with IntelÒ Core (TM) i7-7800X CPU, 32 GB RAM, and rest”. The parameters C and c of the SVM were explored using a
6 Z. Mei et al. / Measurement 165 (2020) 108184

Fig. 4. The architecture of the 1D-CNN with a fusion of features from the three types of data. (a) The basic feature extraction module (BPM); it consists of two stacked
convolutional layers and one pooling layer. Each convolutional layer was followed by batch normalization and function activation. (b) A representative fusion model. After the
last pooling layer, the features of the separate maps were concatenated; they were then passed to a network consisting of a flatten layer, two fully connected layers, and a
softmax layer. We used four modules similar to the representative one, to extract features of each variant of input data, respectively.

grid search. PCA was used to project the handcrafted features to function PReLU; kernel size, and pool size of 1  3; a batch size
improve the performance of the SVM. of 32. We first searched the optimal kernel size with default set-
A decision tree is a tree-like classifier based on data partitioning tings; then, using the optimal kernel size and default batch size
[45]. Each internal node splits the data into subsets according to and activation function, we searched for the optimal pool size.
the features. Each edge represents the outcome of the partitioning. The rest parameters were optimized using similar steps. The opti-
Each leaf node represents a class label. The decision tree is an algo- mized parameters for all seven 1D CNN models that cover the three
rithm that does not require domain knowledge or parameter set- types of individual data and the modality combinations are shown
ting. It can handle different types of data without transformation, in Fig. 5(a)–(d). The optimized parameters in the final models for
and can also process multidimensional data. In this study, a grid each type of data and the combination of the data are listed in
search was used to find the appropriate values for the maximum Table 4.
depth and the number of features. The ranges of the hyper- After the 5-fold cross-validation, the model built on accelera-
parameters for the three algorithms are provided in Table 3. All tion data had the following optimal parameters: convolutional ker-
the conventional machine learning algorithms used in the study nel size of 1  5, pool size of 1  7, ELU activation function, and a
are a part of the Python library Scikit-learn 0.19.1. batch size of 32. The best average accuracy of the model on the test
set reached 95.16%. The accuracy and loss on the training set and
validation set for acceleration data and optimized kernel size are
5. Results shown in Fig. 6(a)–(e).

5.1. Parameter optimization 5.2. Recognition of foot types

For finding the optimal parameters of the 1D CNNs, grid search Fig. 7a–g present the confusion matrices of each five-fold cross-
was performed, which started with defaults as follows: activation validation for the individual and combined modality models. With
 Z. Mei et al. / Measurement 165 (2020) 108184 7

Table 3
Ranges of the hyper-parameters.

Classifier Parameters Range

Random Forest Number of trees 20–128
Maximum depth Acc, AV: [None, 10, 20, 30, 40, 50, 60]
Acc + AV: [None, 10, 20, 30, 40, 50, 60, 80, 100, 120]
Force:[None, 10, 20, 30, 40, 50, 60, 80, 100, 120, 140, 160, 180]
Acc + force, AV + force: [None, 10, 20, 30, 40, 50, 60, 80, 100, 120, 140, 160, 180, 200, 220, 240]
Acc + AV + force: [None, 10, 20, 30, 40, 50, 60, 80, 100, 120, 140, 160, 180, 200, 220, 240, 260, 280, 300]
Number of features [None, ’sqrt’]
Decision tree Maximum depth Acc, AV: [None, 10, 20, 30, 40, 50, 60]
Acc + AV: [None, 10, 20, 30, 40, 50, 60, 80, 100, 120]
Force:[None, 10, 20, 30, 40, 50, 60, 80, 100, 120, 140, 160, 180]
Acc + force, gyr + force: [None, 10, 20, 30, 40, 50, 60, 80, 100, 120, 140, 160, 180, 200, 220, 240]
Acc + AV + force: [None, 10, 20, 30, 40, 50, 60, 80, 100, 120, 140, 160, 180, 200, 220, 240, 260, 280, 300]
Number of features [None, ’sqrt’]
SVM C [0.01, 0.1, 1, 10, 100 ]
c [0.00001, 0.001, 0.01, 0.1, 1 ]

None: The search for the best split was performed using all features.
Sqrt: The search for the best split was performed using the square root of all features.

force data being an individual modality, only 16 samples were mis- tively. All average accuracies after the five-fold cross-validation
classified, which is lower than the other two individual modalities. were above 95%. The model using force achieved an average accu-
The model misclassified 11 samples when using the combination racy of 98.92%. As to the sensor combinations, acceleration and
of the acceleration (Acc) and force data, angular velocity (AV) angular velocity determined the lowest average accuracy of
and force data, and all types of data, respectively. 97.91%, which was lower than using force sensor data only. The
The metrics for evaluating the performance of individual data other combinations led to almost identical average accuracy, but
models and those constructed using different modality combina- with different precision, recall, and F-measure. Angular velocity
tions are illustrated in Table 5. The metrics of the model obtained and force data allowed for better recognition of cavus foot with a
from force data were higher than those of the single-modality recall of 0.992. The model using all types of data together deter-
models obtained from acceleration and angular velocity, respec- mined a recall of 0.996 for the normal foot, which is better than

Fig. 5. Performance of 1D CNNs with optimal parameters. Average accuracies with different (a) kernel size (b) pool size (c) activation function and (d) batch size.
8 Z. Mei et al. / Measurement 165 (2020) 108184

Table 4 when using combinations of two modalities. Among the classified

Optimal parameters were used for the final models. samples, the best precision of 100% was achieved for cavus foot
Data Kernel size Pool size Activation function Batch size using acceleration and force data. For planus foot, the best preci-
Acc 1  5 1  7 ELU 32 sion was 0.992 using the combination of three modalities. For nor-
AV 1  4 1  3 PReLU 16 mal foot, the best precision of 0.994 was obtained using angular
Force 1  7 1  4 PReLU 16 velocity and force data.
Acc + AV 1  4 1  3 PReLU 32
Acc + force 1  7 1  5 PReLU 32
AV + force 1  5 1  4 ReLU 16
All data 1  7 1  7 ReLU 32

Fig. 6. Representative accuracy and loss of the model built on Acc data with optimal convolutional kernel size during the 100 epochs.
 Z. Mei et al. / Measurement 165 (2020) 108184 9

Fig. 7. Accumulative confusion matrix for each type of sensor data and their combination.

5.3. Comparison with conventional methods with handcraft features; the results are presented in Table 6. The
average accuracies on all individual modalities and their combina-
To demonstrate the characteristics of the deep learning method, tions were above 93%, 95%, and 79% for the three methods, respec-
random forest, SVM with PCA, and decision tree were performed tively. Among the three conventional methods and the 1D-CNNs,
10 Z. Mei et al. / Measurement 165 (2020) 108184

Table 5
Average metrics for individual modalities and modality combinations.

Data Foot type Precision Recall F-measure Accuracy

Mean SD Mean SD Mean SD Mean SD
Acc Normal foot 0.946 0.0185 0.976 0.012 0.962 0.0075 0.9516 0.0080
Cavus foot 0.97 0.0179 0.942 0.0213 0.956 0.0162
Planus foot 0.95 0.0316 0.902 0.0264 0.922 0.0098
AV Normal foot 0.986 0.0049 0.976 0.0162 0.982 0.0098 0.9758 0.0147
Cavus foot 0.982 0.0223 0.97 0.0329 0.974 0.0242
Planus foot 0.944 0.0361 0.98 0.0190 0.962 0.0286
Force Normal foot 0.992 0.0075 0.988 0.004 0.988 0.0074 0.9892 0.0065
Cavus foot 0.988 0.0098 0.992 0.016 0.988 0.0098
Planus foot 0.982 0.0098 0.986 0.012 0.986 0.008
Acc + AV Normal foot 0.98 0.0167 0.984 0.0136 0.984 0.0049 0.9791 0.0112
Cavus foot 0.978 0.016 0.972 0.0214 0.974 0.0150
Planus foot 0.978 0.0223 0.966 0.0377 0.972 0.0204
Acc + force Normal foot 0.992 0.0075 0.994 0.0049 0.992 0.004 0.99262 0.0044
Cavus foot 1 0 0.992 0.0117 0.996 0.0049
Planus foot 0.986 0.012 0.986 0.012 0.986 0.008
AV + force Normal foot 0.994 0.008 0.99 0.0063 0.994 0.0049 0.99262 0.0053
Cavus foot 0.99 0.0155 0.994 0.012 0.992 0.0075
Planus foot 0.988 0.0098 0.992 0.0098 0.99 0.0063
Acc + AV + force Normal foot 0.992 0.0075 0.996 0.0049 0.992 0.004 0.99261 0.0032
Cavus foot 0.998 0.004 0.984 0.0162 0.99 0.0063
Planus foot 0.992 0.0098 0.992 0.0098 0.992 0.004

SD: standard deviation.

Table 6
The average accuracy of each classification method.

Data Random forest PCA + SVM Decision tree Proposed method

Mean SD Mean SD Mean SD Mean SD
Acc 0.9475 0.0089 0.9771 0.0081 0.7910 0.0062 0.9516 0.0080
AV 0.9369 0.0103 0.9583 0.0033 0.7957 0.0207 0.9758 0.0147
force 0.9845 0.0063 0.9832 0.0061 0.8818 0.0172 0.9892 0.0065
Acc + AV 0.9718 0.0061 0.9885 0.0078 0.8193 0.0298 0.9791 0.0112
Acc + force 0.9913 0.0055 0.9858 0.0039 0.8972 0.0128 0.99262 0.0044
AV + force 0.99260 0.0033 0.9886 0.0027 0.8757 0.0163 0.99262 0.0053
Acc + AV + force 0.9940 0.0039 0.9913 0.0017 0.8751 0.0193 0.99261 0.0032

SD: standard deviation.

decision tree had the lowest performance when individual modal- ported by Hsu et al. [43], where an ink footprint was compared
ities, as well as their combinations, were used. SVM with PCA with low-spatial-resolution force sensor data, and a high correla-
showed the best performance among four methods with an aver- tion was found for the medial arch in static testing and the lateral
age accuracy of 0.977 on Acc data, and 0.9885 on the combination arch in dynamic testing. However, the authors only found the sta-
of Acc and force data, respectively. Random forest outperformed tistical correlation, while no classifier was constructed.
other classifies with an average accuracy of 99.40 on all types of As for the classification of foot types based on the acceleration
data. and angular velocity data, to our best knowledge, no study
explored this topic. In several studies [26,30,31], only statistical
difference was explored. However, those differences indicate that
6. Discussion
the locomotion patterns of the foot types are different, and it is
possible to classify the foot types based on inertial data. The pre-
In this study, we proposed a set of 1D-CNNs to effectively clas-
sent study confirms this assumption. Among classifiers con-
sify three types of feet using acceleration, angular velocity, and
structed in this study, the models that combined features of
force data. Each type of data allows reaching an average accuracy
multiple types of data achieved better results. During a gait cycle,
of over 95%. This result proves that the CNN model can extract
the foot performs dorsiflexion/plantar flexion, eversion/inversion,
the most discriminative features from our insole data. However,
and abduction/adduction. The acceleration and angular velocities
the performances of models constructed on different combinations
reflect the kinematic characteristics of these motions, while the
of sensor data were different. When used individually, force data
force sensor indicates the vertical ground reaction force under
allowed achieving the best average accuracy. This result implies
the main plantar areas. Each modality represents different motion
that the forces acquired from the main anatomical positions of
characteristics of the foot; thus, the fusion strategy allows obtain-
the foot can serve as a reliable indicator of the differences in the
ing the highest number of discriminative features for best classifi-
gait patterns among the three explored foot types. So far, foot type
cation performance.
classification was always based on analysis of plantar pressure dis-
For automatic classification of foot types, Xu et al. [25] used
tribution measured using force sensor arrays. Such an approach
adaptive neuro-fuzzy inference system as a classifier; they
requires large volumes of force sensor data to be collected, pro-
extracted four kinds of arch indexes from plantar pressure and
cessed and stored. Also, it is associated with a high cost of the data
used these as features; the obtained model reached an accuracy
collecting system, especially when compared with a portable
of 95% for classification of planus and cavus feet. Also, Ramirez-
device with just a few force sensors. These conclusions are sup-
 Z. Mei et al. / Measurement 165 (2020) 108184 11

Bautista et al. [48] divided the area of the plantar pressure into solutions are very needed as the in-lab mass screening is challeng-
fourteen sub-areas. To classify planus and cavus feet, they applied ing to implement; on the other side, the target group of the pro-
Fuzzy Cognitive Maps with a genetic algorithm for weight learning posed system is individuals with limited access to regular
and achieved an accuracy of 91%. In this study, we classify foot podiatrist care who need foot monitoring or those who are una-
types into normal, cavus, and planus foot based on data collected ware of possible issues with their foot function. The first group is
through our insole and classified using a 1D CNN. Through our usually represented by the elderly or patients who had surgery
approach, we reached an accuracy of more than 99.26%. That result or are under rehabilitation and need regular foot function monitor-
was possible thanks to the integration of both inertial and force ing; also, the users of foot orthoses. On the other side, youngsters
sensors, which allows obtaining richer motion information in con- are often unaware of possible issues with their foot function and
trast with systems using solely plantar pressure. Besides, 1D CNN the necessary corrective measures. In that case, an early warning
offers the advantage of extracting discriminative features automat- would allow avoiding the long-term complications of untreated
ically, in contrast to using hand-crafted features that might be sub- foot deformity. Long-term observations of foot type could also con-
optimal. A possible drawback of our solution in comparison with tribute to proper footwear selection.
traditional classification methods is that the classification algo- As to the possible disadvantages of the proposed concept, these
rithm requires more computational power, especially in the train- are mostly related to the technical possibilities to build an unob-
ing phase. trusive, convenient device with high user acceptance. We believe
In the study, not all individual modality deep learning models that most of the technical challenges are solvable with the current
outperformed the conventional methods. One of the possible rea- level of technologies, e.g., force sensors and matrices can be cus-
sons is that features that were automatically extracted by the tomized relatively easily, and new force sensor materials are taking
1D-CNNs were different from those used with the conventional place, such as graphene. Also, in this study, we propose a proof-of-
methods; the second possible reason lies in the differences in the concept. However, for a real-world application, different technical
optimization methods. Because of the limitation of the computing details and the system components are to be determined. For
power of the station used to construct the deep learning models, instance, our current classification method is based on a 1D CNN
we adopted a greedy search approach that cannot guarantee an that may naturally be computationally heavy. In our future studies,
optimal global result. For the three conventional methods, we used we need to test if the classification can be executed on a smart-
a grid search for the essential parameters, and it was possibly bet- phone and whether models can be optimized for use in edge sys-
ter than the greedy search. The third possible reason is that when tems. If the processing is not possible locally, data need to be
used deep learning methods, higher volumes of data are needed transmitted to a cloud that will return the result of processing.
since these methods rely on learning a high number of parameters, Then, in case of limited connectivity to servers, problems with
typically exceeding one million. Another limitation is that the age the real-time display of foot status may also arise.
group of the subjects was restricted as all of them were university In the near future, we are going to implement improvements in
students. In our future studies, more subjects covering all age both hardware and algorithmic aspects. At the system level, we are
groups will be explored to allow for a better generalization ability going to increase the number of force sensors to obtain a more
of the deep models. detailed picture of dynamic plantar pressure patterns. With
increasing the number of sensors in the new design, the matter
of data volume optimization will become critical, which we will
7. Conclusions and future work address through the methods of compressive sensing. For the
research of foot types classification, we consider involving 3D force
Collaborative advances in sensing and textiles pave the way of and moments sensing such as Xsens ForceShoe; it will allow mea-
smart footwear. One essential use of such technologies would be suring the foot and ankle motion in medial–lateral, fore-and-aft,
the automatic early detection of foot abnormalities. It has the and vertical directions, thus contributing to obtaining a more dis-
potential to compensate for the lack of enough podiatrist resources criminative motion pattern. In terms of data processing and algo-
and difficulties to establish mass foot screening. So far, however, rithms, we will continue collecting and annotating foot type
little has been done to validate the feasibility of distinguishing foot data; with increasing the size of the training datasets, the general-
types based on a wearable device. To make a step towards filling ization ability of trained models is expected to improve.
this gap, in this study, we used a custom, fully-featured framework One of the open questions in both theoretical and practical
to collect and analyze walking data from normal, cavus, and planus aspects of artificial intelligence methods is that the complex non-
feet. It involved sensor insoles incorporating inertial and force sen- linear models are viewed as black-box classifiers. Hence,
sors for the signal collection and a set of 1D CNN models. For com- approaches are required to allow explaining the decisions of the
parative evaluation of the 1D CNN model performance, three classifier and the contribution of each input variable to the classi-
conventional machine learning methods with handcrafted features fication output. In a theoretical aspect, such techniques will allow
were also applied to classify foot types. Force data allowed for the for a more in-depth understanding of factors determining the clas-
best performance compared with individually used acceleration sifier decisions; from a practical perspective, knowing which input
and angular velocity data. In most cases, combinations of force data variables are most significant will make it possible to optimize the
with one of the other two modalities allowed for a better result whole pipeline from data collection at the sensor node to obtaining
than that of a single modality. the high-level classification result. In our future work, we will ana-
With the present study, we validate the feasibility of automatic lyze factors determining classifier decision and the contribution of
foot types classification through the use of a wearable system for input variables, applying techniques such as the Layer-Wise Rele-
gait analysis. Because of factors including the up-until-recently- vance Propagation [49].
limited availability of the required technologies to build unobtru-
sive and practical sensor-enabled footwear, applications of wear-
able system-based gait analysis are yet to make their way in CRediT authorship contribution statement
practice. Existing commercial sensor-enabled footwear products
mainly address diabetic foot monitoring, and fitness uses. In con- Zhanyong Mei: Investigation, Data curation, Software, Writing -
trast, the concept we suggest could allow for real-time monitoring original draft. Kamen Ivanov: Data curation, Software, Writing -
of the foot function and early warning of foot deformities. Such review & editing. Guoru Zhao: Validation, Visualization. Yua-
12 Z. Mei et al. / Measurement 165 (2020) 108184

nyuan Wu: Visualization, Writing - review & editing. Mingzhe Liu: [21] T.C. Chao, B.C. Jiang, A comparison of Postural Stability during Upright
Standing between Normal and Flatfooted Individuals Based on COP-Based
Conceptualization, Writing - review & editing. Lei Wang:
Measures, Entropy 19 (2) (2017) 1–13.
Supervision. [22] R. Tahmasebi, M.T. Karimi, B. Satvati, F. Fatoye, Evaluation of Standing Stability
in Individuals With Flatfeet, Foot Ankle Spec. 8 (3) (2015) 168–174.
Declaration of Competing Interest [23] W.R. Ledoux, H.J. Hillstrom, The distributed plantar vertical force of neutrally
aligned and pes planus feet, Gait Post. 15 (1) (2002) 1–9.
[24] J.G. Barton, A. Lees, Development of a connectionist expert system to identify
The authors declare that they have no known competing finan- foot problems based on under-foot pressure patterns, Clin. Biomech. 10 (7)
cial interests or personal relationships that could have appeared (1995) 385–391.
[25] S. Xu, X. Zhou, Y.N. Sun, A novel gait analysis system based on adaptive neuro-
to influence the work reported in this paper. fuzzy inference system, Expert Syst. Appl. 37 (2) (2010) 1265–1269.
[26] C. Grech, C. Formosa, A. Gatt, Shock attenuation properties at heel strike:
Acknowledgment Implications for the clinical management of the cavus foot, J. Orthopaed. 13 (3)
(2016) 148–151.
[27] A.K. Buldt, P. Levinger, G.S. Murley, H.B. Menz, C.J. Nester, K.B. Landor, Foot
This project was supported in parts by the NSFC under grant no. posture is associated with kinematics of the foot during gait: A comparison of
U19A2086 and Grant no. 61701049, Key Project 2017GZ0304 of normal, planus and cavus feet, Gait Post. 42 (1) (2015) 42–48.
[28] K.M. Kruger, A. Graf, A. Flanagan, B.D. McHenry, H. Altiok, P.A. Smith, G.F.
the Science and Technology Department of Sichuan province and Harris, J.J. Krzak, Segmental foot and ankle kinematic differences between
the young and middle-aged backbone teachers’ program of rectus, planus, and cavus foot types, J. Biomech. 94 (2019) 180–186.
Chengdu University of Technology. [29] A.K. Buldt, S. Forghany, K.B. Landorf, P. Levinger, G.S. Murley, H.B. Menz, Foot
posture is associated with plantar pressure during gait: A comparison of
normal, planus and cavus feet, Gait Post. 62 (2018) 235–240.
References [30] A. Barnes, J. Wheat, C.E. Milner, Fore- and rearfoot kinematics in high- and
low-arched individuals during running, Foot Ankle Int. 32 (7) (2011) 710–716.
[1] N. Shibuya, L.J. Ciliberti, D.C. Jupiter, V. VanBuren, J.L. Fontaine, Characteristics [31] W.R. Ledoux, H.J. Hillstrom, Acceleration of the calcaneus at heel strike in
of Adult Flatfoot in the United States, J. Foot Ankle Surg. 49 (4) (2010) 363–368. neutrally aligned and pes planus feet, Clin. Biomech. 16 (7) (2001) 608–613.
[2] E. Sadeghi-Demneh, J.M.A. Melvin, K. Mickle, Prevalence of pathological [32] C.M. Kerr, A.B. Zavatsky, T. Theologis, J. Stebbins, Kinematic differences
flatfoot in school-age children, Foot 37 (2018) 38–44. between neutral and flat feet with and without symptoms as measured by the
[3] V. Sachithanandam, B. Joseph, The influence of footwear on the prevalence of Oxford foot model, Gait Post. 67 (2019) 213–218.
flat foot. A survey of 1846 skeletally mature persons, J. Bone Joint Surg. Br. 77 [33] L.B. Jiao, H. Wu, R.F. Bie, A. Umek, A. Kos, Multi-sensor Golf Swing Classification
(2) (1995) 254–257. Using Deep CNN, Procedia Comput. Sci. 129 (2018) 59–65.
[4] B.S. Neal, I.B. Griffiths, G.J. Dowling, G.S. Murley, S.E. Munteanu, M.M.F. Smith, [34] J. Camps, A. Samà, M. Martín, D. Rodríguez-Martín, C. Pérez-López, J.M. Moreno
N.J. Collins, C.J. Barton, Foot posture as a risk factor for lower limb overuse Arostegui, J. Cabestany, A. Català, S. Alcaine, B. Mestre, A. Prats, M.C. Crespo-
injury: a systematic review and meta-analysis, J. Foot Ankle Res. 7 (55) (2014) Maraver, T.J. Counihan, P. Browne, L.R. Quinlan, G. Laighin, D. Sweeney, H.
1–13. Lewy, G. Vainstein, Alberto Costa, R. Annicchiarico, À. Bayés, A. Rodríguez-
[5] P. Wicart, Cavus foot, from neonates to adolescents, Orthopaed. Traumatol. Molinero, Deep learning for freezing of gait detection in Parkinson’s disease
Surg. Res. 98 (7) (2012) 813–828. patients in their homes using a waist-worn inertial measurement unit,
[6] A.S. Eleswarapu, B. Yamini, R.J. Bielski, Evaluating the Cavus Foot, Pediatr. Ann. Knowl.-Based Syst. 139 (2018) 119–131, https://doi.org/10.1016/
45 (6) (2016) e218. j.knosys.2017.10.017. https://linkinghub.elsevier.com/retrieve/pii/
[7] A. Manoli, B. Graham, Clinical and new aspects of the subtle cavus foot: A S0950705117304859.
review of an additional twelve year experience, Fuß & Sprunggelenk 16 (1) [35] D. Ravi, C. Wong, B. Lo, G.Z. Yang, A deep learning approach to on-node sensor
(2018) 3–29. data analytics for mobile or wearable devices, IEEE J. Biomed. Health. Inf. 21
[8] M. Razeghi, M.E. Batt, Foot type classification: a critical review of current (1) (2017) 56–64.
methods, Gait Post. 15 (3) (2002) 282–291. [36] K. Ivanov, Z.Y. Mei, L. Lubich, N. Guo, X.L. Deng, Z.C. Zhao, O.M. Omisore, D. Ho,
[9] A. Bertani, A. Cappello, M.G. Benedetti, L. Simoncini, F. Catani, Flat foot L. Wang, Design of a Sensor Insole for Gait Analysis, ICIRA 2019 (2019) 433–
functional evaluation using pattern recognition of ground reaction data, Clin. 444.
Biomech. 14 (7) (1999) 484–493. [37] K. Ivanov, Z.Y. Mei, L. Lubich, L. Wang, Recognition of Pes Cavus Foot Using
[10] A.D. Cock, T. Willems, E. Witvrouw, J. Vanrenterghem, D.D. Clercq, A functional Smart Insole: A Pilot Study, ICIRA 2019 (2019) 654–662.
foot type classification with cluster analysis based on plantar pressure [38] Y. Lecun, Y. Bengio, G. Hinton, Deep Learning, Nature 521 (7553) (2015) 436–
distribution during jogging, Gait Post. 23 (3) (2006) 339–347. 444.
[11] H. Böhm, C. Oestreich, R. Rethwilm, P. Federolf, L. Döderlein, A. Fujak, C.U. [39] Y. Zheng, Q. Liu, E.H. Chen, Y. Ge, J.L. Zhao, Time series classification using
Dussa, Cluster analysis to identify foot motion patterns in children with multi-channels deep convolutional neural networks, Web-Age Information
flexible flatfeet using gait analysis—A statistical approach to detect Management (2014) 298–310.
decompensated pathology?, Gait Post. 71 (2019) 151–156. [40] B. Xu, N.Y. Wang, T.Q. Chen, M. Li, Empirical Evaluation of Rectified Activations
[12] A. Jafarnezhadgero, S.H. Mousavi, M. Madadi-Shadd, J.M. Hijmans, Quantifying in Convolutional, Network (2015). https://arxiv.org/pdf/1505.00853.
lower limb inter-joint coordination and coordination variability after four- [41] S. Ioffe, C. Szegedy, Batch Normalization: Accelerating Deep Network Training
month wearing arch support foot orthoses in children with flexible flat feet, by Reducing Internal Covariate Shift, 2015. https://arxiv.org/abs/1502.03167.
Hum. Mov. Sci. 70 (2020) 102593. [42] S. Ruder, An overview of gradient descent optimization algorithms, 2017.
[13] K. Stanković, T. Huysmans, F. Danckaers, J. Sijbers, B. Booth, Subject-specific https://arxiv.org/abs/1609.04747.
identification of three dimensional foot shape deviations using statistical [43] W.C. Hsu, T. Sugiarto, J.W. Chen, Y.J. Lin, The design and application of
shape analysis, Expert Syst. Appl. 151 (2020) 113372, https://doi.org/10.1016/ simplified insole-based prototypes with plantar pressure measurement for
j.eswa.2020.113372. https://linkinghub.elsevier.com/retrieve/pii/ fast screening of flat-foot, Sensors 18 (11) (2018) 1–14.
S0957417420301974. [44] O.S. Schneider, K.E. MacLean, K. Altun, I. Karuei, M.M. Wu, Real-time gait
[14] T.P. Fong, Y.Y. Chan, The Use of Wearable Inertial Motion Sensors in Human classification for persuasive smartphone apps: structuring the literature and
Lower Limb Biomechanics Studies: A Systematic Review, Sensors 10 (12) pushing the limits, IUI 2019 (2019) 19–22.
(2010) 11556–11565. [45] J.W. Han, M. Kamber, J. Pei, Data Mining: Concepts and Techniques, third ed.,
[15] C.A. Ronao, S.B. Cho, Human activity recognition with smartphone sensors Morgan Kaufmann, MA, USA, 2011.
using deep learning neural networks, Expert Syst. Appl. 59 (2016) 235–244. [46] V. Vapnik, The Nature of Statistical Learning Theory, Springer-Verlag, New
[16] J. Hannink, T. Kautz, C.F. Pasluosta, K.G. Gasmann, J. Klucken, B.M. Eskofier, York, NY, USA, 2000.
Sensor-based gait parameter extraction with deep convolutional neural [47] N.U. Ahamed, D. Kobsar, L.C. Benson, C.A. Clermont, S.T. Osisc, R. Ferber,
networks, IEEE J. Biomed. Health. Inf. 21 (1) (2017) 85–93. Subject-specific and group-based running pattern classification using a single
[17] J. Song, H.J. Hillstrom, M. Neary, K. Choe, W. Brechue, R.A. Zifchock, et al., wearable sensor, J. Biomech. 84 (2019) 227–233.
Dynamic barefoot plantar pressure in gait and foot type biomechanics, J. Foot [48] J.A. Ramirez-Bautista, J.A. Huerta-Ruelas, L.T. Kóczy, M.F. Hatwágner, S.L.
Ankle Res. 7 (1 Supplement) (2014) A77. Chaparro-Cárdenas, A. Hernández-Zavala, Classification of plantar foot
[18] S.S. Lee, S.T. Choi, S.-I.I. Choi, Classification of Gait Type Based on Deep alterations by fuzzy cognitive maps against multi-layer perceptron neural
Learning Using Various Sensors with Smart Insole, Sensors. 19 (8) (2019). network, Biocybern. Biomed. Eng. 40 (1) (2020) 404–414.
[19] A.K. Buldt, S. Forghany, K.B. Landorf, G.S. Murley, P. Levinger, H.B. Menz, Centre [49] F. Horst, S. Lapuschkin, W. Samek, K.R. Müller, W.I. Schöllhorn, Explaining the
of pressure characteristics in normal, planus and cavus feet, J. Foot Ankle Res. Unique Nature of Individual Gait Patterns with Deep Learning, Sci. Rep. 9
11 (3) (2018) 1–9. (2391) (2019).
[20] B. Chuckpaiwong, J.A. Nunley, N.A. Mall, R.M. Queen, The effect of foot type on
in-shoe plantar pressure during walking and running, Gait Post. 28 (3) (2008)
405–411.