Recurrent Neural Network for Gait Pathology Detection

Jorge Sanchez-Casanova, Judith Liu-Jimenez, Pablo Fernandez-Lopez and Raul Sanchez-Reillo

University Group for Identiﬁcation Technologies, University Carlos III of Madrid, Leganes, Spain

Keywords:

Pathology Detection, Recurrent Neural Network, Pattern Recognition, Inertial Sensors.

Abstract:

This work presents a pathology detection system on the lower train. For this, a database of healthy subjects has

been captured. Due to the nonexistence of pathological gait databases, pathology walks have been simulated.

The users used sole padding in order to simulate clubfoot walk. The database consists of acceleration, angular

acceleration, magnetic ﬁeld signals and the angles between the joints. The algorithm extracts fragments of the

signals which are used to train a recurrent neural network (RNN). To optimize the results, hand-tuning method

was used to modify the hyperparameters. Using the best conﬁguration, we have a 97% accuracy training with

90% of the database. Although, if we train with only 50% of the data the accuracy reaches at 91%. The

results obtained show the solution feasibility, although further research should be done using real lower train

pathologies.

1 INTRODUCTION

Nowadays, in the world of medicine, there are new

techniques that help in the diagnosis. An example of

this is breast cancer detection (Fear et al., 2002; Hen-

riksen et al., 2018). However, for the case of the lower

train pathology detection, this is not so extended. In

this ﬁeld, there are some machines that perform gait

analysis. The analysis of the walk gives informa-

tion like the pressure of the tread, joint angle, the ca-

dence of the walk, etc. The information obtained is

commonly used in rehabilitation progress. Some ex-

amples are the rehabilitation in patients with motor-

impaired (Banala SK, Agrawal SK, 2007; Raveh

et al., 2019), strokes (Dickstein, 2008; M.H. et al.,

1997; Go et al., 2019) to cerebral palsy (Kainz et al.,

2017; Moreau et al., 2016; Rutovi

c et al., 2019). Nev-

ertheless, all this data can be used in order to diagno-

sis lower train problems like sprains or fractures.

The problems of these machines are that work in

a closed environment since most are optical and need

ﬁxed cameras: this compels the patient to move to

the clinic. This is a disadvantage due to the way of

walking can be affected by the environment (Del Din

et al., 2016).

One solution is to use inertial sensors. Usually,

these sensors have a small size and can be attached to

the body to record its movements. These systems only

need a HUB to connect all the devices and to save the

data.

The aim of this study is to build a system capable

of separating healthy walks from pathology walks and

the laterality of them. For this purpose, a database

has been collected with healthy users and three kinds

of walks: normal, right injury and left injury walks.

These injuries have been simulated. On this database,

a neural network has been applied to classifying the

walks. A study of the hyperparameters has been done

to ﬁnd the best conﬁguration.

This paper is divided into 6 sections: Section 2

explains gait variability and the factors that can affect

it, in section 3 the protocol of the database collection

and the capture system is described. Data processing

is deﬁned in section 4. And the results and conclu-

sions are in sections 5 and 6.

2 WALK VARIABILITY

Each person walks different from the others, even

more, the walks of the same person differ. However,

there are some characteristics intrinsic to the way of

walking. These characteristics show a pattern in the

gait of each user. This pattern can be used to recog-

nize users (Fernandez-Lopez et al., 2017), or in the

way of walking with some disease or injury.

The variability can occur due to some dis-

ease (Schaafsma et al., 2003), surgery (Khouri and

Desailly, 2013), physiological differences (walking

speed increases with the stature (Winter, 2010) ) or

Sanchez-Casanova, J., Liu-Jimenez, J., Fernandez-Lopez, P. and Sanchez-Reillo, R.

Recurrent Neural Network for Gait Pathology Detection.

DOI: 10.5220/0008910600600067

In Proceedings of the 13th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2020) - Volume 4: BIOSIGNALS, pages 60-67

ISBN: 978-989-758-398-8; ISSN: 2184-4305

different surfaces (a beach, a ﬂat surface, a park,

etc.). Some disorders like Parkinson’s, Huntington’s

or Alzheimer’s disease can increase gait variability

(Yu et al., 2009). For example, patients with Parkin-

son’s disease tend to walk with reduced gait speed and

shorter stride (Schaafsma et al., 2003). So, the change

in gait parameters can point some disease. It can also

be used as an indicator to predict the risk of falling

down in elderly people (Hausdorff et al., 2001).

Gait variability is helpful in the case of gait recog-

nition. However, in some cases, like physiological

variances or different surfaces, it is a problem at the

time to identify pathologies. For example, the walk-

ing pattern with a sprain can be quite different be-

tween people.

3 DATABASE DESCRIPTION

There are some public databases with gait signals.

But there is no database available with healthy and

pathological walks. Because of the unavailability

of proper databases, we created our own database.

In this section, the proceedings of the database, the

equipment used in the acquisition and the dataset are

explained.

3.1 Capture System

Tech-MCS V3, a portable kinematics movement ac-

quire system was used to collect the database. This

system consists of a hub to store the data and 7 in-

ertial measurement units (IMU). Each IMU has one

accelerometer, one gyroscope and one magnetometer,

all of them work in 3D. The signals of each sensor

are merged to obtain the orientation of the IMUs. The

process is used in a sensorial fusion process in order

to obtain orientation using inertial information. This

process is performed by an Extended Kalman Filter

(EKF) that runs in each IMU. Diagram of the process

is presented in Figure 1. The orientation data can be

used to obtain the angle data of the joints.

To acquire the data, the seven IMUs were attached

the following way: two at the foot, two at the middle

of the shin, two at the middle of the sank and one at

the lumbar. In each walk, 81 signals were recorded.

The angle data correspond to different leg move-

ments. In Figure 3 the movements of each joint

are presented. The ﬁrst one belongs to the sagittal

plane (x-axis), in which the ﬂexion and extension take

place. The coronal plane (y-axis) is the next one; this

consists of the movement of the leg from right to left

and vice versa. The last one is the rotation of the joints

over themselves.

Figure 1: Process of the IMU to obtain the 3D orientation.

Figure 2: Positions of the IMUs in the body.

3.2 Protocol

A database of healthy and fake-pathological gait was

collected. In order to isolate the pathology studied

in this experiment, the people that collaborated did

not have any gait impediment, that means that no per-

son in the database suffers from sprains, surgeries, ﬂat

foot, etc.

When simulating a pathology some points must be

taken into account.

1. Easiness and comfortability for the user.

2. Replicability for all users.

3. Similarity with real pathology.

Sprains can be simulated using a bandage. But is

not easy and, in some cases can cause some pain. Fi-

nally, clubfoot walk is simulated using sole padding.

In order to perform as real as a possible experi-

ment we consider the following rules:

• Comfortability: the users wear his/her own

clothes and shoes. The only restriction is not to

use high heels or ﬂip-ﬂops, due to it is not possi-

ble to use them with sole padding.

• Freedom: The users can perform the visit when

and wherever they want. The only restriction is

the walking path must be ﬂat.

Recurrent Neural Network for Gait Pathology Detection

Figure 3: Different movements per plane.

Each user made three sessions, 15 days passed

from session one and two, and two months between

two and three. By doing this we eliminate variables

such as exhaustion or temporal pain in the users. Each

visit is made up of:

• 8 free walks.

• 4 walks with the sole padding in the right feet

• 4 walks with the sole padding in the left feet

3.3 Dataset

For this study, 31 healthy people were recruited. Of

these users, only 21 did the second visit, and of these

21 only 7 did the third one.

Figure 4: Schema of the recorded data.

As an addition, two people with an ankle sprain

(one on the right and the other one in the left) were

recruited for the experiment. Each user did 10 walks

walking freely. These people will be used at the end to

test if the system is capable of identifying real pathol-

ogy having been trained with fake pathology.

Table 1: Number of users, walks, and samples per visit.

Visit Users Walks Samples

1st 31 496 1984

2nd 21 336 1344

3rd 7 112 448

Total 59 994 3776

As we can see in 1 there are 3776 samples, 1888

are from healthy walks, and 944 from each left and

right limp walk. The ranged age of de database is

from 18 to 77 years old, and the gender distribution is

47 % female and 53%, male.

4 SYSTEM

The aim of the system is to distinguish between right-

limp walks, left-limp walks and no limp walks. The

system is divided into three steps. The ﬁrst step is the

pre-processing, where the signals are ﬁltered, once

the signal is ﬁltered the signal is divided into frag-

ments. Lastly these fragments are classiﬁed using a

neural network. In Figure 5 the workﬂow of the whole

system is presented.

Figure 5: System workﬂow.

4.1 Pre-processing

Due to the nature of the gait signals, the bigger part

of the information is low frequency. So, to remove

unnecessary noise the signals were ﬁltered by a But-

terworth low pass ﬁlter. The Butterworth ﬁlter is used

due to its simplicity and the -3 dB gain at the cut-

off frequency. We set the cut-off frequency, approx-

imately, in the frequency where the power spectrum

falls down under -40 dB. To ﬁnd this frequency all

the signals were studied. Heuristically a group of dif-

ferent cut-off frequencies were obtained.

Table 2: Cut-off frequencies of the different signals.

Position Signal

Frequency

(Hz)

All Angle 20

Lumbar and thigh

Accelerometer 20

Gyroscope 10

Magnetometer 10

Shin and feet

Accelerometer 40

Gyroscope 20

Magnetometer 20

Once the signals are ﬁltered the next step is to ex-

tract the data.

BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing

4.2 Data Extraction

Walking is a pseudo-periodic movement, this means

that it is almost the same movement every time: right

step, left step and repeat. These pseudo-periodic

movements are called gait cycles. In some studies,

these gait cycles are used as a basic unit for biometric

recognition (Fernandez-Lopez et al., 2018).

In this study, instead of gait cycles, a bigger part

of the data is used as a basic unit. By using fragments

larger than the gait cycles the evolution from the right

to the left and then from the left to the right step can

be observed, having much more information to locate

the pathologies.

• To avoid non-stable signal there must be at least

three seconds left before the start point. There

also must be three seconds left between the end-

point and the end of the signal;

• All the 81 signals of a walk must be trimmed using

the same points;

• The start and endpoint must be the same moment

in the cycle.

Once the constraints are ﬁxed, the next step is to

perform the frame extraction. Since all the signals in

a walk are synchronized, the timestamps of all signals

are the same. This is why once we ﬁnd the beginning

and end of a fragment in one signal the same times-

tamps are used to fragment the remaining 80 signals.

To choose the model signal, all the signals were

studied. The signals with the clearest cycles are from

the hip and the knee. There are not noteworthy differ-

ences between the signals from the right and the left,

so the right signals were selected. Once the lateral-

ity of the signals was selected, the signals of the hip

and the knee were inspected. The one with clearer re-

sults is the knee signal, this signal corresponds with

the ﬂexion/extension movement. Thus, the right knee

signal is used as a model signal for the data extraction

process.

The ﬁrst step of the extraction process is to seg-

ment the gait cycles. In this study, the maximum point

is used as the beginning of the cycle. This point cor-

responds to the position where the right leg is fully

stretched forward, that is, the knee has the greatest

extension angle. To ﬁnd the proper maximum values,

and avoid the local maximum data, the peaks must

accomplish that:

• The distance between peaks must be greater than

900 ms (Fernandez-Lopez et al., 2017).

• The peak must be placed in the upper third.

Once the edges of the fragment are found, all the

signals in a walk are trimmed off in segments. After

Figure 6: Right knee signals. Different colours represent

different ﬁnal fragments. Black and doted signal are dis-

carded.

the fragments have been obtained, the problem arises

from the fragments of each user and walk have a dif-

ferent duration. To solve this problem, all the signals

are trimmed to the length of the shorter one. Since the

signals still long enough, all of them are divided into

four segments.

In order to prepare the data to feed the neural net-

work, it has to be organised in samples. Each sample

is a matrix of N x 81 where N represents the number

of samples of the fragment. Each column represents

the different axis (x, y and z) and sensors (accelerom-

eter, gyroscope, magnetometer and body angles). Fol-

lowing is presented the organization of the data inside

the matrix.







Acc

(1) Acc

(1) · · · LAnkle

(1) LAnkle

(1)

Acc

(2) Acc

(2) · · · LAnkle

(2) LAnkle

(2)

Acc

(3) Acc

(3) · · · LAnkle

(3) LAnkle

(3)

· · · · ·· · · ·

Acc

(N) Acc

(N) · · · LAnkle

(N) LAnkle

(N)







4.3 Neural Network

With the aim of develop a neural network (NN), the

python deep learning library Keras (Chollet et al.,

2015) is utilised.

To feed the neural network (NN) the data is orga-

nized in a matrix. The output data corresponds to one

of the three cases of walk: no limp, left limp or right

limp.

A NN is a group of algorithms that attempt to rec-

ognize the relationship of a group of data through a

process that mimics the way the human brain oper-

ates. Depending on the data there are some NN that ﬁt

better. In these algorithms, there are multiple options

for conﬁguration. In order to obtain the better results,

the number of layers, the neurons of each layer, num-

ber of ﬁlters, the activation algorithm, ratio of train-

Recurrent Neural Network for Gait Pathology Detection

ing/testing, etc. can be set up. These adjustable pa-

rameters are called hyperparameters.

There is not a ﬁxed rule about how many layers

should be used. Three layers are the minimum num-

ber of layers we can have. More layers can give better

results, but it will be harder to train. This study com-

pares the results and the training time of two neural

networks: the ﬁrst one with three layers and the sec-

ond one with four layers. The number of neurons per

layer is also studied. To do this the relationship be-

tween the layers is ﬁxed. The input layer has 2

neu-

rons. The hidden layers have 2

n−1

neurons. The out-

put layer never has 3 neurons, since it is the number

of classes to predict. The value of n starts in 2. The

value is increased by 1 until the results stop improv-

ing. Thus, the results of how changes the accuracy

with the complexity of the NN can be observed.

Figure 7: Structure of the neural networks.

Due to gait signals being temporal, a recurrent

neural network (RNN) (Hochreiter and Schmidhuber,

1997) is used. The structure of the three-layer NN is

shown in Figure 7. The input layer is an RNN with 2

Long Short-Term Memory (LSTM). The hidden layer

is a Dense layer and the output layer is a three neurons

layer. To avoid overﬁtting and improve the accuracy

of the system a dropout of 0.2 is performed after the

RNN layer. The structure of the four-layer NN is the

same, but it includes another dense layer between the

hidden layer and the output layer.

After each layer we use an activation function. Af-

ter the RNN and the hidden layers, we use relu, as it

maintains linearity. However, for the output layer, we

use a sigmoid function to get a probabilistic output.

In this study, to obtain the best results of the NN

the ratio of training/testing data and the number of

neurons in each layer have been analysed. The ratio

of train/test data goes from 10/90 % to 90/10 %.

5 RESULTS

The results of the different conﬁguration of the sys-

tem are presented below. First case under study is

the effects of changing the ratin of raining/testing data

and the numbers of neurons per layer. Following this

the outcomes of adding an extra layer to the RNN are

presented. Two experiments have been performed: to

classify two real pathologies with the RNN and to

check if the system classiﬁes equally right and left

walks. The results of these experiments are presented

at the end of the section.

The ﬁrst parameter to study is the training-testing

ratio. In the state of the art, there is not a ﬁxed ratio of

training-testing for the data, even though it is recom-

mended 20-40% for testing and 80-60% for training.

To solve this problem and to give a clearer solution,

we trained the network with different percentage of

data between 10 and 90%. For each ratio of data, the

algorithm has been trained and tested 10 times and the

ﬁnal accuracy is the mean value of all of them.

Figure 8: Results of 3 layers network. Different colors rep-

resent different neurons per layer.

In Figure 8 and 9, we can see the accuracy of the

algorithm against the percentage of training data. In

general, a bigger training dataset provides better re-

sults. However, we can notice that there is a point

from which the increase of the training data does not

provide signiﬁcant improvement. This point is lower

when the ratio of training data is bigger. For exam-

ple, in the network with n = 7 (128/64/64 neurons)

training with only 50% of the data, the 91% of the

frames are correctly classiﬁed,. However, if the train-

BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing

ing data is increased until 90% the accuracy only im-

proves 6%.

Figure 9: Results of 4 layers network. Different colors rep-

resent different neurons per layer.

A different number of neurons per layers has been

used in order to ﬁnd the best conﬁguration. The num-

ber of neurons per layer has been increased until the

improvement stops. As we can see in Figure 8 and 9

the accuracy grows with the number of neurons. The

more neurons the layers have, better the results. This

can be appreciated especially in the networks with

16/8, 32/16, 64/32 and 128/64. The optimal conﬁg-

uration occurs when n = 8 and it produces a network

with 256 neurons in the ﬁrst layer and 128 in the sec-

ond and the third. The accuracy of this network does

not ﬁt into any pattern. In Figure 8, the accuracy of

the network with a ratio of 50/50 of the data is 88%,

but it suddenly drops to 60% with a bigger training

dataset. This behaviour is due to the complexity of

the network as it is too high. On the other hand, the

networks with n = 2 and n = 3 have similar conduct,

but in this case, is due to not enough complexity.

Up to now, the best conﬁguration is the one with

n = 7 and 90% of training data. The last hyperparam-

eter under study is the number of layers of the NN.

Two different NN have been created to determine the

inﬂuence of adding a layer: the ﬁrst one with 3 layers

and the other one with 4. Figure 8 shows the results of

the 3-layer NN. The outcomes of the 4-layer NN are

presented in Figure 9. We can verify that the results

for both are similar.

Figure 10 presents the comparison of the best con-

ﬁguration of NN with 3 and 4 layers and the training

time of each. As we can see the improvement in the

best case is lower than 3%. So, there is no signiﬁcant

improvement adding an extra layer to the NN.

For all the possible conﬁguration of the RNN

the one that offers the best results is the one with

128/64/64 neurons per layer and with 4 layers. In or-

der to test the RNN two experiments have been per-

formed: classifying walks of two users with sprain,

Figure 10: Comparison of the results with 3 and 4 layers.

The dotted line shows time. The solid line shows the accu-

racy.

and to check if the system classiﬁes equally all the

cases.

For the ﬁrst experiment two users with real ankle

sprain have been recruited. Each of these users has

done 10 walks, so 80 frames have been classiﬁed. The

results of classifying the frames are presented in table

Table 3: Accuracy of classify sprain fragments.

Percentage of training data Accuracy (%)

50 % 45

60 % 50

70 % 66

80 % 71

90 % 83

The classiﬁcation of the frames have been done

using the NN with 4 layers and 128/64/64 neurons

per layer. The network have been trained using the

fake pathology data, this network shows a result of

83% of accuracy classifying the samples. The drop

in the accuracy may be due to there are not sprains in

the training dataset. Even though it is not a negligible

result, it cannot be considered since the data used is

not enough.

The last experiment lies in to check if there is any

difference in the accuracy when classifying the differ-

ent walks separately. The network has been trained

with all data and tested only with the corresponding

cases. Seven cases have been studied: healthy walk

and left limp, healthy walk and right limp, either right

or left limp, only left, only right, only healthy and

all the walks. The accuracy against the ratio of train-

ing/testing data is presented in ﬁgure 11. The results

show that there is no signiﬁcant difference when clas-

sifying the different cases.

Recurrent Neural Network for Gait Pathology Detection

Figure 11: Comparison of classiﬁcation the walks sepa-

rately.

6 CONCLUSIONS AND FUTURE

WORK

The aim of this paper is to present a system capable of

classifying pathology walks using RNN. The signals

of the lower train are recorded using a commercial

device, that records signals from accelerometer, gyro-

scope, magnetometer and orientation data. The sig-

nals are ﬁltered and processed in order to extract the

information required to feed the neural network. With

the aim to obtain the optimal conﬁguration of the NN

hand-tunning of the hyperparameters has been done,

among them the ratio of training/testing data, number

of neurons per layer and number of layers have been

studied.

To the light of the outcomes, we can see that, in

general cases, the accuracy of the system rises with

the training ratio. There are two cases where it is not

precise. This happens when the complexity of the NN

is too elevate, and the results do not converge at any

point. Adding an extra layer to the NN improves less

than 3% the accuracy of the system, however, it does

not increase signiﬁcatively the effort.

The experiment of classifying the real cases of

pathology walks give proper results. However, the ex-

periment does not have a signiﬁcant impact due to the

testing does not hold abundant data. Additionally, it

is able to classify equally the walks from the right and

left and the health walks.

Even though this system works and classiﬁes

properly the walks, there are some improvements that

would be interesting to perform. The ﬁrst one is to

study the inﬂuence of the origin (e.g. the beginning

or the middle of the walk) and the amount of data of

the fragments. In this way, if the amount of data can

be reduced without remarkable changes in the accu-

racy the algorithm can be optimized.

Once a system capable of distinguishing walk

limp and the laterality of them, the next step is to try

the system with real pathologies. So, the next step is

the acquisition of a new database of users with dif-

ferent problems in the lower train. In this case, the

algorithm should be reﬁned to get as a result, not only

the laterality of the limp but also the joint where it

occurs.

ACKNOWLEDGEMENTS

This work was partially supported by the Spanish

National Cybersecurity Institute (INCIBE) under the

Grants Program “Excellence of Advanced Cybersecu-

rity Research Teams.”

The work of this paper has been partly funded by

PREVIEW project, granted by the Spanish Min-

istry of Economy and Competence, with the code

TEC2015-68784-R (MINECO/FEDER)

REFERENCES

Banala SK, Agrawal SK, S. S. (2007). Banala SK, Agrawal

SK, Scholz SP. Active leg exoskeleton (ALEX) for

gait rehabilitation of motor-impaired patients. IEEE

Int Conf Rehabil Robot 2007, 401–7. Rehabilitation

Robotics, . . . , 00(c).

Chollet, F. et al. (2015). Keras. https://keras.io.

Del Din, S., Godfrey, A., Galna, B., Lord, S., and Rochester,

L. (2016). Free-living gait characteristics in ageing

and Parkinson’s disease: Impact of environment and

ambulatory bout length. Journal of NeuroEngineering

and Rehabilitation, 13(1):1–12.

Dickstein, R. (2008). Review article: Rehabilitation of gait

speed after stroke: A critical review of intervention

approaches. Neurorehabilitation and Neural Repair,

22(6):649–660.

Fear, E. C., Li, X., Hagness, S. C., and Stuchly, M. A.

(2002). Confocal microwave imaging for breast can-

cer detection: Localization of tumors in three dimen-

sions. IEEE Transactions on Biomedical Engineering,

49(8):812–822.

Fernandez-Lopez, P., Sanchez-Casanova, J., Liu-Jimenez,

J., and Morcillo-Marin, C. (2017). Inﬂuence of walk-

ing in groups in gait recognition. Proceedings - Inter-

national Carnahan Conference on Security Technol-

ogy, 2017-Octob:1–6.

Fernandez-Lopez, P., Sanchez-Casanova, J., Tirado-Martin,

P., and Liu-Jimenez, J. (2018). Optimizing resources

on smartphone gait recognition. IEEE International

Joint Conference on Biometrics, IJCB 2017, 2018-

Janua:1–6.

Go, M., Iacovelli, C., Russo, E., Pournajaf, S., Blasi, C. D.,

and Franceschini, M. (2019). Stroke Gait Rehabilita-

tion : A Comparison of End-E ff ector , Overground

Exoskeleton , and. Applied Sciences.

BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing

Hausdorff, J. M., Rios, D. A., and Edelberg, H. K. (2001).

Gait variability and fall risk in community-living older

adults: A 1-year prospective study. Archives of Phys-

ical Medicine and Rehabilitation, 82(8):1050–1056.

Henriksen, E., Carlsen, J., Vejborg, I., Nielsen, M. B., and

Lauridsen, C. (2018). The efﬁcacy of using computer-

aided detection (cad) for detection of breast cancer in

mammography screening: a systematic review. Acta

Radiologica, 60:028418511877091.

Hochreiter, S. and Schmidhuber, J. (1997). Long short-term

memory. Neural Computation, 8(8):1735–1780.

Kainz, H., Graham, D., Edwards, J., Walsh, H. P., Maine, S.,

Boyd, R. N., Lloyd, D. G., Modenese, L., and Carty,

C. P. (2017). Reliability of four models for clinical

gait analysis. Gait and Posture, 54(March):325–331.

Khouri, N. and Desailly, E. (2013). Rectus femoris trans-

fer in multilevel surgery: Technical details and gait

outcome assessment in cerebral palsy patients. Or-

thopaedics and Traumatology: Surgery and Research,

99(3):333–340.

M.H., T., G.C., M., and R.R., R. (1997). Rhythmic fa-

cilitation of gait training in hemiparetic stroke re-

habilitation. Journal of the Neurological Sciences,

151(2):207–212.

Moreau, N. G., Bodkin, A. W., Bjornson, K., Hobbs, A.,

Soileau, M., and Lahasky, K. (2016). Effectiveness of

Rehabilitation Interventions to Improve Gait Speed in

Children With Cerebral Palsy: Systematic Review and

Meta-analysis. Physical Therapy, 96(12):1938–1954.

Raveh, E., Schwartz, I., Karniel, N., and Portnoy, S. (2019).

Evaluation of the effectiveness of a novel gait trainer

in increasing the functionality of individuals with mo-

tor impairment: A case series. Assistive Technology,

31(2):106–111. PMID: 29035638.

Rutovi

c, S., Glavi

c, J., and Cvitanovi

c, N. K. (2019). The

effects of robotic gait neurorehabilitation and focal

vibration combined treatment in adult cerebral palsy.

Neurological Sciences, 40(12):2633–2634.

Schaafsma, J. D., Giladi, N., Balash, Y., Bartels, A. L.,

Gurevich, T., and Hausdorff, J. M. (2003). Gait dy-

namics in Parkinson’s disease: relationship to Parkin-

sonian features, falls and response to levodopa. Jour-

nal of the neurological sciences, 212(1-2):47–53.

Winter, D. A. (2010). The Biomechanics and Motor Control

of Human Gait. 74(2):94.

Yu, H., Riskowski, J., Brower, R., and Sarkodie-Gyan, T.

(2009). Gait variability while walking with three dif-

ferent speeds. 2009 IEEE International Conference

on Rehabilitation Robotics, ICORR 2009, pages 823–

827.

Recurrent Neural Network for Gait Pathology Detection