Inertial-based Gait Analysis Applied to Patients with Parkinson Disease

Joana Sousa

, Joana Silva

2 a

, Ricardo Leonardo

2 b

, Hugo Gamboa

1,2 c

and Josefa Domingos

3 d

Faculdade de Ci

encias e Tecnologia, FCT, Universidade Nova de Lisboa, 2829-516 Caparica, Portugal

Associac¸

ao Fraunhofer Portugal Research, Rua Alfredo Allen 455/461, 4200-135 Porto, Portugal

Grupo de Patologia M

edica, Nutric¸

ao e Exerc

ıcio Cl

ınico do Centro de Investigac¸

ao Interdisciplinar Egas Moniz (CiiEM),

Caparica, Portugal

Keywords:

Gait Analysis, Parkinson’s Disease, Foot-mounted IMU, Spatio-temporal Gait Metrics, Complementary

Filter, ZUPT.

Abstract:

People with Parkinson’s disease have a high incidence of falls due to motor difﬁculties. Recent studies have

shown that PD patients can receive beneﬁt from motor therapy based on cueing and feedback. This study

describes a system based on a foot-mounted IMU for the calculation of gait parameters applied to different

datasets of healthy elderly people, geriatric patients and patients with PD, in order to integrate it into a real-

time acquisition system with application for tactile cueing. This system is divided into different steps: the

identiﬁcation of gait cycles and their events, the estimation of the path of the foot, which includes the esti-

mation of the orientation of the foot, the application of methods to correct the error derived from the double

integration of acceleration such as ZUPT, and ﬁnally the estimation of the different gait metrics. The results

show that the algorithm developed is an accurate method for stride segmentation and is considered adequate

to assess the gait metrics for gait evaluation of patients with motor difﬁculties.

1 INTRODUCTION

Parkinson’s disease (PD) is the second most common

neurodegenerative disease and it is increasing with the

aging of the world population. However its low inci-

dence, it affects about 1% of the population over 65

years old worldwide (Moore et al., 2005). PD causes

a high incidence of falls due to death dopaminergic

neurons in the substantia nigra that causes movement

disorders (Moore et al., 2005). In fact, ”it is esti-

mated that 60.5% of patients with PD have at least

one fall and that 39% have recurrent falls” (Kalilani

et al., 2016), which consequently contributes to the

increased risk of fractures in patients. The symptoms

of the disease are gradual and include problems with

gait, balance, posture, in addition to tremors, bradyki-

nesia and rigidity. The disease can’t be cured, but it

is usually treated with medication, such as levodopa,

and in some cases with neurosurgery, but the patients

beneﬁt from physiotherapy since diagnose to improve

physical capacity, gait and balance (Tomlinson et al.,

https://orcid.org/0000-0002-6214-5868

https://orcid.org/0000-0003-2695-4462

https://orcid.org/0000-0002-4022-7424

https://orcid.org/0000-0001-9390-6183

2014).

With the increase in the number of Parkinson’s pa-

tients, the high incidence of falls and considering the

need for physical therapy to improve their quality of

life, it is increasingly important to develop solutions

that allow patients to improve their motor skills in

a common environment, outside the physical therapy

environment. In addition, access to physical therapy

and the frequency of sessions can often become in-

creasingly limited, which motivates the development

of a device that assesses the patient’s gait and pro-

vides feedback to correct certain positions, or even

stimulates a more controlled gait.

In fact, motor impairment of PD leads to speciﬁc

gait characteristic in PD, such as reduced step length,

longer step time and consequently reduced gait speed

and less regular gait cycle (Mariani et al., 2013).

Stance duration and double support duration were in-

creased for the Parkinson’s population, whereas sin-

gle support duration, mean cadence, and heel-to-heel

base of support were markedly reduced (Nelson et al.,

2002). Some of these parameters are difﬁcult to as-

sess and vary signiﬁcantly from patient to patient, so

an additional system that can estimate gait metrics is

beneﬁcial in helping doctors conﬁrm their observa-

Sousa, J., Silva, J., Leonardo, R., Gamboa, H. and Domingos, J.

Inertial-based Gait Analysis Applied to Patients with Parkinson Disease.

DOI: 10.5220/0010328403270334

In Proceedings of the 14th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2021) - Volume 4: BIOSIGNALS, pages 327-334

ISBN: 978-989-758-490-9

327

tions and diagnoses during long-term trails.

Gait analysis is usually performed in a laboratory

setting using a motion capture system or a pressure-

sensitive walkway. These systems are considered

gold standard in terms of accuracy. These systems al-

low a high degree of accuracy in measurement, how-

ever they are very complex to use on a daily ba-

sis and are only suitable for laboratory or hospital-

like environments such as physiotherapy clinics, due

to their high monetary value and little ease of use

(Tunca et al., 2017) and adaptation to any environ-

ment. Therefore, a portable system of good precision

that can estimate the parameters of gait is an added

advantage for the evaluation of motor disorders and

for monitoring motion performance in PD.

The evolution of wearable sensor technology en-

abled to obtain an acquisition system with a low cost

and size that allows gait dysfunction and motor symp-

toms to be assessed. The most used wearable sen-

sors are inertial sensors packed into inertial measure-

ment units (IMUs) (Tunca et al., 2020). IMU consists

of a unit that integrates tri-axial accelerometer, gyro-

scope and magnetometer. This unit also allows BLE

or Wireless communication with a device, being able

to acquire synchronized data in real time from several

IMUs, in order to obtain a complete information of

the gait (Tunca et al., 2017).

Currently, there are many studies of inertial sen-

sors in the scope of gait assessment and gait monitor-

ing performance, mainly for lower-body gait analy-

sis with foot-located IMUs. In fact, one of the major

focuses of research in inertial sensors is to develop a

system that extracts gait metrics, both spatial and tem-

poral, over short and long gait paths for accurate as-

sessment of dysfunctions. Methodologies with good

precision have been developed that combine step seg-

mentation methods by identifying different walking

moments and the calculation of spatial metrics (Mar-

iani et al., 2013; Tunca et al., 2017). These studies

vary by the segmentation method of the gait, (Rampp

et al., 2015) use multi-dimensional subsequence dy-

namic time warping (msDTW) for geriatric patients

and some patients with PD, (Ferrari et al., 2016;

Tunca et al., 2017) used algorithms that involve the

identiﬁcation of the angular velocity peaks for Parkin-

son’s patients and healthy.

This study describes a foot-mounted IMU system

for the calculation of gait parameters applied to differ-

ent datasets of healthy people, geriatric patients and

patients with PD, in order to integrate it into a real-

time acquisition system and application of biofeed-

back. This system is divided into different steps: the

identiﬁcation of gait cycles and their events, the esti-

mation of the path of the foot, which includes the esti-

mation of the orientation of the foot, the application of

methods to correct the error derived from the double

integration of acceleration such as zero-velocity up-

dates algorithms (ZUPT), and ﬁnally the estimation

of the different gait metrics.

2 METHODS

2.1 Acquisition Protocol and Datasets

The requirements of development of this system had

the partnership of the Portuguese Parkinson Disease

Patient Association (APDPk). However, due to the

pandemic situation that was established, the coopera-

tion was no longer possible. The pandemic prevented

the involvement of patients in the association, as they

are a group of potential risk. Patient participation

would undermine safety measures and increase risk

of contagion. For this reason, this section describes

the datasets used, which are mostly public.

In that sense, it was used 2 datasets and all have

certain requirements. The datasets include samples of

triaxial inertial data from sensors placed on both feet

of the patient while walking and data from an accu-

rate reference system to validate the methods. In the

initial phase of familiarization with the data and for

the deﬁnition of different gait metrics (such as num-

ber of steps, speed, stride length), the subjects did not

need to have a neurodegenerative disease.

The public eGaIT database (embedded Gait anal-

ysis using Intelligent Technologies) contains two

datasets which was used in this study, the Validation

Stride Segmentation and the Validation of Gait Pa-

rameters. These datasets were provided by Profes-

sor Ph.D Bjoern Eskoﬁer of the Machine Learning

and Data Analytics laboratory at Friedrich-Alexander

University Erlangen-Nuernberg and they are accessi-

ble for collaborative research (FAU, 2015). The iner-

tial data of the both datasets were acquired using the

same material and set-up system (Barth et al., 2015;

Rampp et al., 2015).

The eGaIT database for Validating Stride Segmen-

tation consists of data from 30 individuals. Of the to-

tal participants, 10 are elderly controls, 5 male and 5

female, 10 individuals are patients with Parkinson’s

disease, 5 male and 5 female, and 10 are geriatric pa-

tients, 4 male and 6 female (Barth et al., 2015). The

mean age for the elderly controls is 64.0 ± 8.4 years,

the mean age for patients with PD is 63.8 ± 9.3 years

and for geriatric patients is 81.0 ± 4.1 years (mean ±

standard deviation). The patients with PD were eval-

uated to assess symptoms, severity and degree of dis-

ease. The ground truth of the start and end points of

BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing

328

each stride was provided with the inertial data.

The eGaIT database for Validating Gait Parame-

ters consists of data from 101 inpatient geriatric pa-

tients, where 55 are female and 46 are male, the mean

age is 82.1 ± 6.5 years and the mean height is 164.0 ±

10.0 cm (mean ± standard deviation) (Rampp et al.,

2015). The ground truth of the stride time, stride

length, swing time and stance time was also provided

with the inertial data. This data set does not contain

data from patients with Parkinson’s, but does contain

data from participants with motor difﬁculties. It con-

tains the ground truth of the start and end points of the

stride and the validation of different spatial and tem-

poral metrics (stride length, stride time, swing time

and stance time) from a GAITRite system, which is

useful to validate the algorithm of spatial gait param-

eters (Barth et al., 2015; Rampp et al., 2015).

2.2 Algorithm Architecture

The algorithm architecture is represented in Fig.1 and

is based on two approaches. The ﬁrst is the foot-

mounted complementary ﬁlter (CF) aided IMU ap-

proach for pedestrian tracking in indoor environment

(Fourati, 2015). However, the system proposed in this

study is not for estimating the trajectory, but for esti-

mating gait parameters. The second is the identiﬁca-

tion of gait events to estimate the position for each

stride to calculate gait metrics (Rampp et al., 2015).

Figure 1: System Architecture (Barth et al., 2015).

In this sense, the complementary ﬁlter is used to

estimate the orientation of the foot along the gait, rep-

resented by the quaternion, ˆq

(t). The complemen-

tary ﬁlter has as inputs the raw acceleration ( f ) and

raw angular velocity given by the sensors (w). This

quaternion will be used to transform the acceleration

into the body’s reference, which is the raw accelera-

tion given by the sensor ( f ), in the Earth’s coordinate

system, a =



0 a



a = ˆq

(t) ⊗ f ⊗ ˆq

∗

(t) (1)

Gravitational acceleration G is also calculated

with the use of a quaternion, G

= q

(t) ⊗



0 0 −1 0



⊗q

∗

(t), G =

. Next, it is nec-

essary to remove the contribution of G from the ac-

celeration vector a





, to obtain

the acceleration of movement of the foot in the earth

frame,

A(t) = a

(t) − G (2)

and thus it is possible to integrate the calculated ac-

celeration by obtaining the 3-D vector in the Earth co-

ordinate system. Theoretically, the resulting velocity

vector can be immediately integrated again to obtain

position. However, due to the presence of noise in the

inertial signal, which leads to changes and drifts in the

measured acceleration and errors in the orientation es-

timation by the quaternion, the immediate integration

of the acceleration would result in an estimation of

the wrong position, due to the referred accumulation

of errors. One method to reduce the error is the appli-

cation of the Zero Velocity Update (ZUPT) method,

which will be referred to in section 2.2.2. This method

is based on the correction of linear speed during gait.

Gait is a cyclic movement that results in the transfer

of weight from one foot to another alternately. When

a foot is in contact with the ground, its linear speed is

theoretically zero (Fourati, 2015). Due to the accumu-

lation of error during the referred process, the linear

speed during these support phases may not be zero,

and thus the ZUPT method identiﬁes the moments of

zero speed and corrects the speed initially calculated.

The corrected speed is integrated to obtain position.

Previous studies (Ferrari et al., 2016; Tunca et al.,

2017) showed that this approach was not as accurate

as the Pedestrian Dead-Reckoning (PDR) system us-

ing the Kalman Filter. However, this method will be

used to calculate the position step by step, and in each

step the position will return to zero and a new orienta-

tion will be calculated by not increasing the accumu-

lation of the error throughout the acquisition period.

Thus, through position and speed it is possible to cal-

culate the spatial metrics of the gait. Also, it was used

the CF due to its easy implementation, understanding

and because it requires little computational power.

2.2.1 IC/FO Events Detection

As mentioned in the introduction, there are many lit-

erature that have studied and developed stride seg-

mentation methods and some which are based on the

identiﬁcation of the initial contacts of the foot on the

ﬂoor (IC) and the foot-off (FO), the lifting of the foot

from the ﬂoor (Ferrari et al., 2016; Tunca et al., 2017),

but few are capable of running in closed loop for real

time acquisition or for real time gait events. The iden-

tiﬁcation of these events is important because from

them it is possible to estimate all temporal metrics of

the gait, such as stride time, swing time, stance time,

cadence and double support. The evaluation of the

Inertial-based Gait Analysis Applied to Patients with Parkinson Disease

329

angular velocity signal according to the medio-lateral

axis is a method with good accuracy, ideal for auto-

matic detection of gait events. The method developed

is based in (Ferrari et al., 2016).

Unlike studies of (Ferrari et al., 2016) that cal-

culate in the ﬁrst instance the positive peaks, which

represent the moments that the foot is rotating coun-

terclockwise considering the side view of a person

walking to the right. The developed algorithm ﬁrst

identiﬁes the negative peaks that may correspond to

the FO / IC events, moments of the gait when the foot

is rotating clockwise. Calculating these events ﬁrst,

it is guaranteed that no steps are missed in this pro-

cessing. Next, the positive peaks, which can repre-

sent mid-swing events, are calculated and it is veri-

ﬁed if this positive peak is between the ﬁrst and the

second detected peak. If the previous fact occurs then

the identiﬁed peaks can be considered to correspond

in fact to the events of the gait. If not, that is, if a

ﬁrst FO peak has been identiﬁed and does not follow

a mid swing, the FO is not considered, or if a negative

second peak, a possible IC, has not been identiﬁed af-

ter a positive peak, then it will be looked for a next

negative peak following the mid swing found.

The identiﬁcation of positive and negative peaks is

based on the ﬁnd peaks function of the Python Scipy

Signal library (Community, 2020) and takes into ac-

count limits and conditions that the peaks must meet

in order to be recognized as IC an FO events in all

patients. These limits are based on the algorithm used

by (Ferrari et al., 2016) and adapted so that events are

detected in all datasets for healthy, geriatric and pa-

tients with PD. The developed algorithm is executed

in soft real-time and has a maximum delay of half a

step, since it is necessary to evaluate the gait from step

to step to calculate the metrics. However, this delay is

compatible for mobile applications, such as feedback

tips while walking in loop mode (Ferrari et al., 2016).

The result of the algorithm is represented in Figure 2.

With the identiﬁcation of these events it is pos-

sible to segment the stride, calculate the stride time,

identify moments such as the stance phase and the

swing phase and their respective duration. But for

the calculation of the position of each step is impor-

tant to determine also the mid-stance moments. The

mid-stance events were deﬁned has the middle of the

stance phase, being the stance phase the period be-

tween a IC and a FO moments.

2.2.2 ZUPT Algorithm

The gait can be divided into two important phases,

which can be identiﬁed through the inertial signal of

the IMU ﬁxed on the foot of a person walking. The

ﬁrst is the swing phase which means that the foot and

Figure 2: Detection of IC, FO, mid-stance and mid-swing

events of a Patient with PD from a eGait dataset.

the IMU are in the air. The other is the stance phase,

which is the period that the foot is in contact with

the ﬂoor. According to the studies of (Fourati, 2015),

during the stance phase the angular and linear veloc-

ity must be very close to zero and the accelerometer

should measure only the earth gravitational accelera-

tion, because the foot is on the ground and theoretical

it is not moving. However, this phase includes the IC

event, the foot-ﬂat period in which the foot is com-

pletely resting on the ﬂoor and the FO event, when the

foot is leaving the ground but is still in contact. Con-

sidering that the IMU is ﬁxed only in one part of the

foot, it does not present speed only during the foot-

ﬂat period (Wang et al., 2015). Thus, Multi-condition

ZUPT presented by (Guo et al., 2015) was adapted for

this study, aims to identify these moments, the periods

when the inertial sensor is not in motion, representing

them with a logical function assigning 1 when the lin-

ear velocity should be zero and 0 when the velocity

does not need correction. This algorithm uses data

from the accelerometer and gyroscope and presents

several conditions, as follows:

1. Since the acceleration in periods 1 must be just the

acceleration of gravity, the magnitude of the ac-

celeration (|a

|) must be between two limits. The

limits are deﬁned around the value 9.8 m/s

C1 =



1 8 <

< 11

0 otherwise

(3)

2. The acceleration variance must be above a given

threshold where a

is a mean acceleration value at

time k, and s is the size of the averaging window

(s = 15). The variance is computed by:





2s + 1

k+8

∑

j=k−8





− a



(4)

BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing

330

The second condition is deﬁned by this way:

C2 =



1 σ





< 0.5

0 otherwise

(5)

3. The magnitude of the angular rate (|ω

|) must be

below a given threshold.

C3 =



< 0.8

0 otherwise

(6)

The logical result of the ZUPT algorithm is ob-

tained when the 3 conditions are satisﬁed. The ﬁnal

logical result is ﬁltered out by a median ﬁlter with

neighboring window of 29 samples. The Figure 3

shows de result of this multi-condition algorithm from

data of a patient with parkinson disease.

Figure 3: ZUPT algorithm applied to all data acquisition of

a patient with PD from the eGait dataset.

This method will work for each stride and if the

zero speed zones identiﬁed by the ZUPT are outside

the IC-FO period, the system does not consider the

ZUPT phases and considers the stance phase identi-

ﬁed by the event detection algorithm 2.2.1.

2.2.3 Gait Parameters

Temporal gait parameters were calculated based on

the gait events identiﬁed from each dataset, and were

deﬁned from two consecutive strides. The stance time

is deﬁned in seconds as the duration of stance phase,

starting with initial-contact (IC) or heel strike (HS)

and ending with foot-off (FO) or toe-off (TO) of the

same foot. The swing time is the duration of swing

phase, starting with FO and ending with IC of the

same foot. The stride time is the duration of a gait cy-

cle. Considering the instant when the event occurred

of the stride i as t

(i) and t

(i), it is possible to de-

ﬁne each time parameter (Rampp et al., 2015).

Stride Time (i) = t

(i + 1)−t

(i) (7)

Swing Time (i) = t

(i) − t

(i) (8)

Stance Time (i) = t

(i + 1) −t

(i) (9)

After calculating the 3D position based on the de-

veloped system, it is easy to calculate the spatial met-

rics. The stride length is the distance between two

successive placements of the same foot, deﬁned as the

two norm of a 2-D vector containing the distance in

x- (posterior-anterior axis) and z- (medio-lateral axis)

direction at the end of the stride, since the position in

the start of the stride is zero (Rampp et al., 2015).

Stride length =





(n)





(n)

+ p

(n)

(10)

3 RESULTS

As in this study, different datasets are used with differ-

ent validation metrics. This section is divided into two

parts: stride segmentation results of the two datasets

of the eGait database and gait metrics results of the

second eGait dataset.

3.1 Stride Segmentation Results

The IC / FO event detection algorithm, described in

2.2.1, is the method used in this study to segment the

stride. The accuracy of this algorithm was tested in

the eGait datasets. Validation of this algorithm was

achieved using the annotated strides of the dataset.

The annotation of the samples referring to the begin-

ning and to the end of each stride, for each subject,

was performed by manual identiﬁcation an labeling

of gyroscope peaks. In these studies, the stride was

considered from the moment when the foot leave the

ground (FO) of one cycle to the FO of the next cycle.

In Table 1 the results related to gait segmenta-

tion for the different patients in the Validation of

Stride Segmentation dataset of the eGait database

are provided. The table quantiﬁes the number of

strides, more speciﬁcally, the number of strides an-

notated in the study (Annotated Strides), the number

of strides detected by the developed algorithm (De-

tected Strides) and the number of strides that were

detected by the developed algorithm and that coincide

with the annotated strides (True Detected Strides). In

the study, patients walked in a straight line for 10 me-

ters, four times, and at the end of each straight line

the subject rotates 180 degrees to walk in the opposite

direction. The strides annotated in the dataset do not

refer of the moments when the patient is changing di-

rection. With this information it is possible to verify

that all the annotated strides are correctly identiﬁed

Inertial-based Gait Analysis Applied to Patients with Parkinson Disease

331

Table 1: Results of Stride Segmentation of the eGait Dataset - Validation of Stride Segmentation. Mean of the stride period

annotated, the mean of the stride period calculated, the mean error of the start of each stride and end of each stride are given

as mean ± standard deviation.

Patients Foot

Annotated

Nr. Strides

Detected

Nr. Strides

True Detected

Nr. Strides

Mean Strides

per Patient

Annotated

Stride Period

(sec.)

Detected

Stride Period

(sec.)

Mean Error

Start Stride

(samples)

Mean Error

End Stride

(samples)

Right 248 248 312 24.8 ± 1.8 0.94 ± 0.01 0.94 ± 0.04 -0.98 ± 0.04 0.98 ± 0.04

Left 248 248 313 24.8 ± 1.6 1.04 ± 0.02 1.04 ± 0.02 -1.01 ± 0.02 1.01 ± 0.02

Geriatric

Right 384 384 491 38.4 ±n11.0 1.35 ± 0.09 1.35 ± 0.09 -0.99 ± 0.02 0.99 ± 0.02

Left 381 381 490 38.1 ± 10.9 1.34 ± 0.08 1,34 ± 0,08 -0,99 ± 0.00 0,993 ± 0.00

Control

Right 242 242 306 24.2 ± 2.7 1.09 ± 0.03 1.09 ± 0.03 -0.99 ± 0.01 1.00 ± 0.01

Left 220 220 310 24.4 ± 2.2 1.10 ± 0.03 1.10 ± 0,03 -1.00 ± 0.00 1.00 ± 0.00

by the algorithm for all patients. However, the algo-

rithm has identiﬁed more strides, that are those that

correspond to moments of changing of direction.

The results table 1 also shows the average num-

ber of strides taken by the patients during walking in

straight line, that is, over a distance of 40 meters, the

average period, in seconds, of the stride annotated in

the study, and the average period of the true detected

strides. This period was obtained by dividing the du-

ration of the strides (in samples) by the sampling rate,

which is 102.4 Hz. The average number of steps taken

by geriatric patients is much higher than the average

number of other patients (PD or controls). This group

also has a much higher average age, as can be seen

in 2.1, which implies more limited mobility, leading

to a slower walking cycle (Stride Period) and since

all groups walked exactly 40 meters, it is possible to

deduce that they have a shorter gait length. Patients

with Parkinson have in average the same number of

steps and a relatively longer gait period than the con-

trol group, composed by elderly people. The differ-

ence is also not very clear, since the staging of Hoehn

and Yahr is on average 1.7 ± 0.9, which is a not very

advanced stage in the disease where changes in gait

begin to exist (Keus et al., 2009).

To evaluate the quality of the segmentation, the

average of the differences between the samples at

the beginning of each stride annotated in the dataset

and the ones calculated by the algorithm were calcu-

lated and is presented in the Table 1 (mean error start

stride). The same was retrieved for the end of the

stride, the average of the differences between the sam-

ples referring to the end of each stride annotated and

the ones calculated by the algorithm is also presented

in the Table 1 (mean error end stride). These aver-

age errors are very low and have the same modulus

value for detecting the beginning and the end of each

stride, which means that there is only a lag of about

1 sample leading to an equal stride time between the

annotated values and the ones retrieved by the algo-

rithm developed. These results show a great accuracy

of the algorithm to identify gait events, more specif-

ically FO, compared to other methods of gait events

identiﬁcation (Barth et al., 2015).

Table 2 presents the results of gait segmentation

for the right and left foot of 97 subjects, inpatient geri-

atric patients, from the Validation of Gait Parameters

dataset of the eGait database. Four patients were ex-

cluded from the dataset because there were inconsis-

tencies in the data provided. The validation of the

stride annotations was carried out in the same way as

mentioned for the previous dataset. The table shows

the total number of strides annotated by the dataset in

the 10-meter path, the number of strides detected by

the algorithm, and the average number of strides per

subject during the acquisition period. To check the

accuracy of the algorithm, the average period of the

strides and the average error were also calculated, as

presented in the previous table 1. The results for this

dataset proved the accuracy of stride detection, since

all strides have been correctly identiﬁed with a small

average error leading to a stride period’s mean differ-

ence of milliseconds . In the next section, the results

of the temporal gait metrics and the stride length for

this dataset will be evaluated.

3.2 Gait Parameters Results

This section presents the results for the gait metrics

retrieved for the second dataset. The metrics calcu-

lated were stride length, stride time, stance time and

swing time, as shown in the Table 3. The results were

from 97 subjects of 101 patients of the dataset. The

Table shows the average errors (Mean error), deﬁned

by the average of the differences between the metrics

calculated by the developed system and the metrics

provided in the annotated dataset, the absolute error

(Abs. Error), which is the average of the absolute er-

rors and the correlation that is deﬁned by the Spear-

man correlation coefﬁcient.

Compared to the study of (Rampp et al., 2015),

which uses the same dataset to analyze the same pa-

rameters of the gait, it is possible to verify that the

mean and absolute error of the stride time calculated

in our study for the right foot are smaller (Mean er-

ror=0.002 ± 0.043, Abs.error=0.023 ± 0.044) than

in the previous study (Mean error = 0.002 ± 0.068,

Abs error = 0.029 ± 0.062) and for the left foot

BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing

332

Table 2: Results of Stride Segmentation of the eGait Dataset - Validation of Gait Parameters. mean number of strides per

patient, mean stride period annotated in the dataset, mean stride period calculated by the algorithm, mean error of the initial

of each stride, and mean error of end of each stride are given as mean ± standard deviation.

Foot

Annotated

Nr. Strides

Detected

Nr Strides

Mean Strides

per Patient

Annotated

Stride Period

(samples)

Detected

Stride Period

(sec.)

Mean error

Start Stride

(sec.)

Mean error

End Stride

(Samples)

Left 663 663 6.84 ± 1.98 1.22 ± 0.19 1.22 ± 0.18 -0.09 ± 1.05 0.44 ± 1.81

Right 659 659 6.79 ± 1.96 1.23 ± 0.19 1.23 ± 0.18 -0.15 ± 1.07 0.25 ± 0.17

Table 3: Results of Gait Parameters of the eGait Dataset - Validation of Gait Parameters. Mean error and abs. error are given

as mean ± standard deviation. Time in sec and length in cm.

Foot Parameter

Mean

(Inertial-Sensors)

Mean

(GAITRite)

Mean Error Abs. Error Correlation Total

Right

Stride Time

1.22 ± 0.18 1.22 ± 0.18 0.002 ± 0.043 0.023 ± 0.044 0.97

756 Strides

97 Subj.

Stance Time

0.78 ± 0.15 0.85 ± 0.15 -0.07 ± 0.05 0.07 ± 0.04 0.96

Swing Time

0.44 ± 0.08 0.37 ± 0.08 0.07 ± 0.04 0.07 ± 0.04 0,91

Stride Length

83.0± 23.7 83.0 ± 21.3 -0.90 ± 9.76 7.45 ± 6.03 0.90

Left

Stride Time

1.23 ± 0.19 1.23 ± 0.19 0.003 ± 0.039 0.023 ± 0.045 0.98

759 Strides

97 Subj.

Stance Time

0.78 ± 0.15 0.84 ± 0.16 -0.07 ± 0.04 0.068 ± 0.039 0.96

Swing Time

0.45 ± 0.07 0.38 ± 0.07 -0.068 ± 0.03 0.068 ± 0.038 0.86

Stride Length

82.0 ± 25.4 83.0 ± 22.0 -0.98 ± 9.76 8.61 + 9.71 0.89

besides being higher (Mean error=0.003 ± 0.039,

Abs.error=0.002 ± 0.045), the values are very simi-

lar. The correlation coefﬁcient of the stride time met-

ric is higher in this study, 0.97, since the correlation

achieved in the previous study was 0.95. With this in-

formation it is possible to verify that the segmentation

algorithm developed in our study is quite accurate.

However, for the remaining metrics, stance time,

swing time and stride length, the results are not so

comparable. The average errors obtained in the study

of (Rampp et al., 2015) were -0.008 ± 0.045, 0.009

± 0.069, -0.26 ± 8.37, for the respective metrics. The

average errors calculated by the developed algorithm,

for the time metrics, are higher, with an order of one

decimal place of difference. The reason for the val-

ues being so different is related to the approach used

to calculate the IC and FO events. There are differ-

ences in the literature in the method used to identify

the events at the beginning and at the end of the gait.

According to (Ferrari et al., 2016), the IC and FO

events are considered the negative peaks of angular

velocity in the medio-lateral axis, whereas according

to the study of (Rampp et al., 2015), the FO events

are considered to be the instants when the gyroscope

signal in to the same axis crosses the zero.

The stride length presents a mean error of less than

1 cm, both for the right foot and for the left foot,

which is a very small error considering the use of this

system for rehabilitation purposes. However, the av-

erage and absolute errors are much higher than the

one calculated in the study of (Rampp et al., 2015)

(Mean error = -0.26 ± 8.37, Abs error = 6.26 ±

5.56). The authors calculated the drift of the gyro-

scope that is contained in the linear acceleration, with-

drew its contribution, and after that the integration

for the speed was accomplished. In this study, the

only way of correcting the inertial errors was applying

the ZUPT method, which corrects the linear velocity

during the stride considering that in the moments of

stance the velocity must be zero. This method only

corrects the drift of the result of the integration of ac-

celeration and does not correct the error associated

with the gyroscope. However, in addition to the fact

that this correction does not exist, the results do not

differ signiﬁcantly from the annotated values.

4 CONCLUSIONS

This study describes a foot-mounted IMU system for

the calculation of gait parameters applied to differ-

ent datasets of healthy people, geriatric patients and

patients with PD, in order to integrate it into a real-

time acquisition system and application of biofeed-

back. This system was divided into different steps:

the identiﬁcation of gait cycles and their events, the

estimation of the path of the foot, which includes the

estimation of the orientation of the foot, the applica-

tion of methods to correct the error derived from the

double integration of acceleration such as ZUPT, and

ﬁnally the estimation of the different gait metrics.

There are different walking segmentation methods

(Haji Ghassemi et al., 2018) more robust methods like

msDTW, eDTW, sDTW and hHMM and more con-

ventional methods like peak detection algorithms. A

recent study of (Barth et al., 2015) showed that the

peak detection algorithm used has, in the best of sce-

narios, around 84% accuracy and 90% F-Score, for

geriatric patients and that when using the msDTW

method the accuracy is between 88% and 90% and

the F-measure is between 96% and 98%. Thus, the

algorithm developed in the study, in addition to be-

ing based on conventional peak detection methods,

Inertial-based Gait Analysis Applied to Patients with Parkinson Disease

333

can be compared to robust methods at the level of

precision, given the results obtained in the identiﬁca-

tion of all strides for the datasets used. This method,

also presents an ideal conﬁguration for a possible in-

tegration in a real-time system, which is a prospective

breakthrough in this work.

It was presented a complete algorithm that allows

gait metrics to be calculated using data from iner-

tial sensors in patients with motor difﬁculties, geri-

atric patients. This system presents adequate results

to make the speciﬁc gait evaluation for the right and

left foot. Although this system present less accurate

results than the analogue study (Barth et al., 2015), is

considered adequate for the scope of gait physiother-

apy. In the future, we intend to calculate other gait

metrics such as double support period, stride width,

swing width and gait speed, which allow a more de-

tailed analysis for people with Parkinson’s, and also

integrate it in a real-time system that allows feedback

to patient whenever the algorithm evaluates a risky

gait pattern, based on spatial parameters.

ACKNOWLEDGEMENTS

Supported by project Indoor Activity Notiﬁcation for

Vigilance Services (AAL-2018-5-116), funded under

the AAL JP and co-funded by the European Commis-

sion and the National Funding Authorities of Portu-

gal, Belgium, and Switzerland.

REFERENCES

Barth, J., Oberndorfer, C., Pasluosta, C., Sch

ulein, S.,

Gassner, H., Reinfelder, S., Kugler, P., Schuld-

haus, D., Winkler, J., Klucken, J., and Eskoﬁer,

B. M. (2015). Stride segmentation during free walk

movements using multi-dimensional subsequence dy-

namic time warping on inertial sensor data. Sensors,

15(3):6419–6440.

Community, T. S. (2020). scipy.signal.ﬁnd peaks — SciPy

v1.5.3 Reference Guide.

FAU (2015). Digital Biobank - eGaIT (embedded

Gait analysis using Intelligent Technologies).

https://www.mad.tf.fau.de/research/activitynet/digital-

biobank/.

Ferrari, A., Ginis, P., Hardegger, M., Casamassima, F., Roc-

chi, L., and Chiari, L. (2016). A mobile Kalman-ﬁlter

based solution for the real-time estimation of spatio-

temporal gait parameters. IEEE Trans. Neural Syst.

Rehabil. Eng., 24(7):764–773.

Fourati, H. (2015). Heterogeneous data fusion algorithm for

pedestrian navigation via foot-mounted inertial mea-

surement unit and complementary ﬁlter. IEEE Trans.

Instrum. Meas., 64(1):221–229.

Guo, H., Uradzinski, M., Yin, H., and Yu, M. (2015). Indoor

positioning based on foot-mounted IMU. Bulletin of

the Polish Academy of Sciences: Technical Sciences,

63(3):629–634.

Haji Ghassemi, N., Hannink, J., Martindale, C. F., Gaßner,

H., M

uller, M., Klucken, J., and Eskoﬁer, B. M.

(2018). Segmentation of gait sequences in sensor-

based movement analysis: A comparison of methods

in Parkinson’s disease. Sensors, 18(1):1–15.

Kalilani, L., Asgharnejad, M., Palokangas, T., and Durgin,

T. (2016). Comparing the incidence of falls/fractures

in Parkinson’s disease patients in the US population.

PLoS ONE, 11(9):1–11.

Keus, S. H., Munneke, M., Nijkrake, M. J., Kwakkel, G.,

and Bloem, B. R. (2009). Physical therapy in Parkin-

son’s disease: Evolution and future challenges. Move-

ment Disorders, 24(1):1–14.

Mariani, B., Jim

enez, M. C., Vingerhoets, F. J., and

Aminian, K. (2013). On-shoe wearable sensors for

gait and turning assessment of patients with parkin-

son’s disease. IEEE. Trans. Biomed. Eng., 60(1):155–

158.

Moore, D. J., West, A. B., Dawson, V. L., and Dawson,

T. M. (2005). Molecular Pathophysiology of Parkin-

son ’ s Disease. Annual Review of Neuroscience,

26(1):57–87.

Nelson, A. J., Zwick, D., Brody, S., Doran, C., Pulver,

L., Rooz, G., Sadownick, M., Nelson, R., and Roth-

man, J. (2002). The validity of the GaitRite and the

functional ambulation performance scoring system in

the analysis of Parkinson gait. NeuroRehabilitation,

17(3):255–262.

Rampp, A., Barth, J., Sch

ulein, S., Gaßmann, K. G.,

Klucken, J., and Eskoﬁer, B. M. (2015). Inertial

Sensor-Based Stride Parameter Calculation From Gait

Sequences in Geriatric Patients. IEEE. Trans. Biomed.

Eng., 62(4):1089–1097.

Tomlinson, C. L., Patel, S., Meek, C., Herd, C. P., Clarke,

C. E., Stowe, R., Shah, L., Sackley, C. M., Deane,

K. H., Wheatley, K., and Ives, N. (2014). Physio-

therapy for Parkinson’s disease: a comparison of tech-

niques. CDSR, pages 1–108.

Tunca, C., Pehlivan, N., Ak, N., Arnrich, B., Salur, G.,

and Ersoy, C. (2017). Inertial sensor-based robust gait

analysis in non-hospital settings for neurological dis-

orders. Sensors, 17(4):1–29.

Tunca, C., Salur, G., and Ersoy, C. (2020). Deep Learn-

ing for Fall Risk Assessment with Inertial Sensors:

Utilizing Domain Knowledge in Spatio-Temporal

Gait Parameters. IEEE J. Biomed. Health Inform.,

24(7):1994–2005.

Wang, Z., Zhao, H., Qiu, S., and Gao, Q. (2015). Stance-

Phase Detection for ZUPT-Aided Foot-Mounted

Pedestrian Navigation System. IEEE/ASME Trans.

Mechatronics, 20(6):3170–3181.

BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing

334