Smart Sensing System for the Detection of Specific Human Motion

Symptoms of the Parkinson’s Disease

A. Kita

, P. Lorenzi

, G. Romano

, R. Rao

, R. Parisi

, A. Suppa

, M. Bologna

, A. Berardelli

and F. Irrera

Department of Information Engineering, Electronics and Communications, Sapienza University of Rome,

Via Eudossiana 18, Rome, Italy

Department of Neurology and Psychiatry, Sapienza University of Rome, Rome, Italy

Keywords: Wearable Wireless Inertial Sensors, Motion Features, Freezing of Gait, Neural Network Algorithm, Time

based Analysis, Parkinson’s Disease, Rhythmic Auditory Stimulation.

Abstract: We propose two different wearable wireless sensing systems based on Inertial Measurement Units for the

home monitoring of specific symptoms of the Parkinson’s disease. In one configuration just one sensor is

inserted in a headset, in the other configuration two sensors are positioned on the patient’s shins. They

recognize and classify noticeable motion disorders potentially dangerous for patients and give an audio

feedback. The systems use dedicated algorithms for real time processing of the raw signals from

accelerometers and gyroscopes, one of which is based on an artificial neural network and another on a time-

based analysis. The headset system detects satisfactorily a wide class of motion irregularities including the

trunk disorders, but is poorly reliable on Parkinson’s patients. The other system with sensors on the shins

provides an early detection of the freezing of gait with excellent performance in terms of sensitivity and

precision, and timely provides a rhythmic auditory stimulation to the patient for releasing the involuntary

block state.

1 INTRODUCTION

A wide variety of movement disorders and gait

irregularities are typical symptoms of the Parkinson

Disease (PD) (Nieuwboer et al., 2001). Among

others, the freezing of gait (FOG) is a really disabling

one. FOG is paroxysmal block of movements, which

takes place in an advanced stage of the PD if the

patient is not properly covered by the therapy. During

the FOG, patients refer that their feet are “stuck to the

ground” (Spildooren et al, 2010). In this situation, the

patients make attempts to make a step, oscillating and

thrusting forward the trunk, which can cause

catastrophic events as falls (Bloem et al, 2004). Often,

the FOG is anticipated by a progressive step

shortening (pre-freezing state) (J. Spildooren et al.,

2010), after which the patient stops completely. It has

been shown that a rhythmic auditory stimulation

(RAS) can lead the patients out of the FOG state (P.

Arias and Cudeiro, 2010). The possibility to provide

a RAS timely at the onset of the symptom or in the

pre-freezing state would avoid the undesired

consequences of the block. During the last few years,

several different systems for the automatic detection

of the FOG have been proposed. These are based on

the classification of electrical signals coming from

inertial sensors properly positioned on the patient

body (Lorenzi et al., 2015), (Mazilu et al., 2014), (

Bachlin et al., 2010), (Moore et al., 2013), (B.

Sijobert et al., 2014), (Mazilu et al., 2013) (Cola et

al., 2015), (Atallah et al., 2014). In our work, we

propose the realization of two types of wearable

wireless sensing systems based on MEMS

accelerometers and gyroscopes, able to recognize in

real time specific kinetic features associated to

motion disorders typical of (but not limited to) the PD

and eventually give an auditory stimulation to the

patient to release the involuntary block. They have

been designed to be used at home or outdoor, during

the daily patient life. One system has the sensor in a

headset and uses an artificial neural network (ANN)

for the recognition of the motion features as regular

steps, short steps, gait blocks, trunk oscillations.

Another headset system recently proposed in

152

Kita, A., Lorenzi, P., Romano, G., Rao, R., Parisi, R., Suppa, A., Bologna, M., Berardelli, A. and Irrera, F.

Smart Sensing System for the Detection of Speciﬁc Human Motion Symptoms of the Parkinson’s Disease.

DOI: 10.5220/0005666801520159

In Proceedings of the 9th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2016) - Volume 1: BIODEVICES, pages 152-159

ISBN: 978-989-758-170-0

literature (L. Atallah et al., 2014) uses only

accelerometers just to detect the gait asymmetries

without making any recognition of specific gait

features (which is the topic of our system). The other

system proposed here has two sensors on the shins

and uses a time-based algorithm for the recognition.

Compared to other systems, the headset has the

advantage that it is composed by a single sensor

integrated in the headphone. This makes the system

compact and energy efficient since no wired/wireless

connection is required to give the audio-feedback. On

the other hand, the headset has the disadvantage that

the neck joint mixes signals from the amount of

postural problems and irregular movements typical of

the Parkinson disease, which makes the detected

traces extremely "noisy" and confused (as

experimentally proved).

The second system proposed here requires an

additional device for the audio feedback, but the two

sensors on the shins guarantee the best performance

presented in literature to date in terms of sensitivity,

specificity, precision and accuracy. It has been tested

on a population of PD patients with excellent results.

The board used in the two systems is a prototype

called neMEMSi (D.Comotti et al., 2014) whose size

is 25x30x4 mm3 (with battery, see Fig. 1a).

Figure 1: a) A picture of the NeMEMSi board. b) sketch of

the reference framework of the headset sensor.

The sensor unit LSM9DS0 integrates a ±16 g (g-

force) 3D accelerometer, a ±12 Gauss 3D

magnetometer and a ±2000 dps 3D gyroscope.

Bluetooth communication is supported. The board

integrates an ultralow-power 32 bit microcontroller

(MCU) by STMicroelectronics (STM32L1) with 33.3

DMIPS peak computation capability and very low

power consumption (down to 233 uA/MHz), Flash

memory 256 KB, SRAM 16 KB, EEPROM 4 KB.

Thanks to the Cortex™ M3 architecture and the 32

MHz clock frequency, this MCU is optimized for

advanced and low-power embedded computations.

Actually, until now we performed the measurements

on patients using an external station (a pc) for the

calculations, since the porting on board requires

disclosure of the MCU firmware. However, we are

fully confident that the excellent capabilities of the

MCU guarantee the same system performance since

they are redundant respect to the system

requirements. In fact, the same algorithms have been

already implemented in an Arduino platform (16 bit

MCU ATmega 328P, Flash memory 32KB, SRAM

2KB, EEPROM 1KB, Clock Speed 16MHz, MIPS

16) which is largely less performing of the STM32L1.

2 THE HEADSET SENSING

SYSTEM

2.1 The Soft Operation with an

Artificial Neural Network

This system is composed of a single sensor inserted

in a headset. The reference framework is depicted in

Fig. 1b. The y-axis represents the vertical direction,

the x-axis represents the direction of the walk. The

acceleration along the x direction (Ax, blue) and

along the y direction (Ay, red) in the two states are

drawn in Fig.2. During the walk, the two

accelerations have an oscillatory behaviour, in the

stop state Ay is around 10 m/s

and Ax is around 0.

Figure 2: Typical curves of raw data of Ay (upper curve)

and Ax (lower curve) during a regular walk and in the stop

state.

In the case of Fig.2, the person was first in a stop

state, then he started walking and made 10 steps. In

the walk state, 10 peaks of acceleration can be clearly

distinguished. We need to implement an algorithm

able to recognize the movement disorders typical of

PD: block, the regular steps, the irregular and short

steps, the trunk oscillations. Hereafter, the results

obtained with an artificial neural network (ANN) will

be discussed, since other algorithms revealed less

satisfactorily. We used an ANN with two layers (the

hidden and the output layer). The network consists of

10 neurons, with a sigmoid weight function,

connected in a feedforward topology. We used the

80% of the data for the training with a scaled

conjugate gradient backpropagation algorithm

Smart Sensing System for the Detection of Speciﬁc Human Motion Symptoms of the Parkinson’s Disease

153

already implemented in Matlab (C.M. Bishop et al.,

1995), (T. Chau, 2001). The remaining 20% of the

data was used to validate the algorithm. The cross

entropy is chosen as performance function (D. Kline

and V. Berardi, 2005). Ten epochs are sufficient to

train the ANN in any studied case (discussed in the

following), which indicates that the algorithm is very

light and fast.

Figure 3: Flow diagram of the DTW-ANN training

procedure.

Training: The flow diagram of training is

reported in Fig.3. First of all, we choose the Ay signal

containing a known number of reference patterns with

a known size relative to steps, and we pick out a

reference pattern from it. An example of a step

reference pattern is shown in Fig.4a (selected in

region II of Fig.2).

Figure 4: (a) Reference pattern associated to a regular step.

(b) Reference pattern associated to a short step.

Apart from the amplitude, the reference pattern is

characterized by the size (number of frames) related

to the step time. The known signal is partitioned in

sub-sequences having the same size of the reference

pattern and the reference pattern is compared with the

sub-sequences. To improve flexibility, the Dynamic

Time Warping (DTW) technique is used, since it

allows comparing similar patterns rather than just one

specific pattern in the time subsequence (K.Wang et

al., 1997). DTW is a nonlinear time normalization

technique based on dynamic programming. Given

two time series of different duration, a cost function

can be calculated (E. Keogh and C. A.

Ratanamahatana, 2005). A threshold of the cost

function is set, which determines the degree of

similarity between the reference signal and the

specific subsequence. An example of the cost

function of the DTW is shown in Fig.5. When the

DTW recognizes the reference pattern in a sub-

sequence then the corresponding ANN input is

positive. On the contrary, if the known steps are not

all recognized, the size of the reference pattern and/or

the threshold of the cost function are changed (DTW

optimization) and the DTW is run again.

Figure 5: Cost function of the DTW and optimized

threshold.

2.2 Experimental Results

The ANN is now tested using unknown signals. First,

we monitored four young persons (all male) with

temporary orthopedics problems in deambulation

(defined “healthy”, in comparison with PD patients)

who made the following exercise: stop, walk a few

steps, turning, walk back, stop. The tests regarded the

detection of regular steps, the irregular gait with step

shortening (during turning), the trunk fluctuations. At

a second stage, we monitored PD patients who made

exactly the same exercise.

2.2.1 First Test on Healthy Persons: Regular

Steps and Block

The raw Ay signal of an unknown walk is plotted in

Fig.6 (upper red curve). Four intervals can be

distinguished: interval I is intuitively associated to a

BIODEVICES 2016 - 9th International Conference on Biomedical Electronics and Devices

154

stop state (the Ay value keeps constant at 1000 mg),

intervals II and IV are clearly associated to a periodic

movement, interval III refers to an irregular gait with

short steps (while turning). The ANN was trained to

recognize the stop state. The reference signal in this

case was selected in interval I (it was an almost

straight line, not shown for brevity) and the result is

the lower dotted (ciano) curve. As expected, the ANN

output is 1 in the I interval, is 0 during intervals II and

III and assumes values between 0 and 1in the IV state,

as whether short steps were present. The presence of

regular steps was investigated using the reference

pattern of Fig.4a. The outputs of the ANN in this case

are shown in Fig.6 with the lower (blue) curve (ANN

out-puts close to 1). As one can see, nine steps were

recognized in region II and nine steps in region IV.

No regular steps were identified in region III. We can

conclude that interval III was recognized as a not

walk state and a not stop state. The irregular steps

need further investigation, and are the next focus.

Figure 6: Raw Ay signal of the first unknown test signal

(upper red curve) composed by stop state and walk state.

ANN output associated to stop (ciano) and to walk (blue).

2.2.2 Second Test on Healthy Persons: Short

Steps

The second unknown Ay signal is shown in Fig.7

(upper red curve). In this case, the exercise was

focused on the step shortening. Here, the ANN had to

recognize short steps and distinguish them from

regular ones. Fig.4a, outlines different shapes,

amplitudes, sizes). Therefore, in this experiment the

ANN was trained using a reference pattern selected in

region III. The new reference pattern is displayed in

Fig.4b with arbitrary origin. The ANN outputs are

shown in Fig.7. Although steps were irregular and

featured variable length, the ANN recognized the

short steps in interval III, where just one of the fifteen

was regarded as uncertain (step # 10). Furthermore, a

couple of irregular steps were also detected, when

passing from region I to region II and from region II

to region III.

Figure 7: Raw Ay signal of the second unknown test signal

(upper curve) composed by stop, walk state and irregular

short steps. ANN output associated to the irregular short

steps (lower curve).

2.2.3 Third Test on Healthy Persons: Trunk

Oscillations

In this test, the ANN had to recognize trunk

fluctuations in the x-y plane (referring to Fig.1b). In

this experiment, legs were motionless and only the

trunk oscillated pivoting on the pelvis. This situations

is of particular interest because during a freezing of

gait PD patients feel that their feet are stuck to the

ground and they try repeatedly to make a step

thrusting out and overbalancing. This is clearly

associated to an increased risk of fall. In this case, the

fact that the sensor is positioned on the head

guarantees the maxi-mum sensitivity to the

movement. In this experiment, the angle respect to the

vertical axis varied in the range ±20 degrees. Again,

the Ay raw signal was analysed and the curve is

shown in Fig.8 (upper red curve). As expected, the

trunk oscillations are very well characterized (region

III). Regions I and II are associated, respectively, to a

stop state and a walk state. The ANN was trained to

recognize trunk fluctuations using a reference pattern

selected in region III. It is shown in the inset. The

ANN outputs are displayed in Fig.8 (lower blue

curve). Recognition was excellent and all the trunk

oscillations yielded ANN = 1.

Figure 8: Raw Ay signal of the third unknown test signal

(upper curve) composed by stop state, walk state and trunk

oscillations. ANN output associated to the trunk oscillations

(lower curve).

Smart Sensing System for the Detection of Speciﬁc Human Motion Symptoms of the Parkinson’s Disease

155

2.2.4 Test on PD Patients

All the results of the tests discussed above revealed

that the headset system recognizes successfully the

gait features and the trunk (and head) movements of

healthy persons with temporary orthopedic problems.

Then, we started monitoring PD patients, but we

limited to just a few (male, over 70) since results were

not satisfactorily. Patients were asked to make the

exercise described before. Tests were registered by a

camera and supervised by doctors in order to establish

the exact starting and ending time of the eventual

FOG event. As an example, Fig.9 reports results

relative to one of the PD patients experiencing a FOG

during the test. The patient made four regular steps

and a fifth short step, very close to the previous one.

This fifth step defines an incipient FOG (pre-freezing

state). Then the FOG event occurred, during which

the patient made some irregular movements of the

whole body, without making steps. The five steps are

outlined with arrows on the Ay curve sketched in the

figure. Looking at the ANN output, the first three

steps are correctly detected, step 4 and step 5 are false

negative, whereas a false positive is present during

the FOG state, probably confusing a large movement

of the body with a step. We can conclude that the

headset is not similarly effective on PD patients as on

healthy persons. This is due to the fact that PD

patients feature a great variety of postural problems

and irregular movements in many sections of the

body, all mixed together. The headset device suffers

from the presence of a joint (the neck) which can mix

(or hide part of) the signals, making the information

extremely "noisy" and confused, thus introducing

false positive and false negative outputs.

Figure 9: Raw Ay signal of the test signal taken on a PD

patient (upper curve) composed by four regular steps, one

short step close to the forth one, a FOG during which the

body makes oscillations. The ANN reveals false negative

and false positive outputs.

3 THE SENSING SYSTEM ON

THE SHINS

In order to monitor movement disorders specifically

in PD patients we designed another system with

sensors positioned on the shins. In this case, the

recognition algorithm is based on a time domain

analysis of the sensor signals. The raw signals of

accelerometers and gyroscopes are fused together

through the attitude and heading reference system

(AHRS) and the Madgwick's algorithm (β=0.15)

(Madgwick et al., 2011). The data reading frequency

from the sensor is 60Hz, which allows a correct

sampling of the signal during FOG events since the

relevant spectrum of FOG is 3 – 10 Hz. A quaternion

based representation of the limb orientation and

position is calculated. The angles α

right

and α

left

between the vertical axis and the right/left shin are

sketched in Fig.10.

Figure 10: Angle between the vertical axis and the shin.

The angular velocities ω

right

, ω

left

obtained after

angle derivation are used as the input for the FOG

detection. A new algorithm was developed which

calculates the low-pass of the angular velocities:

= lowpass(|ω

|) (1)

Lef

= lowpass(|ω

Lef

|) (2)

and introduces an index K = k

right

+ k

left

. The

algorithm in eqs.(1) and (2) is an improvement of a

very recent algorithm proposed in literature (Y. Kwon

et al., 2014), which uses the root mean square of the

accelerometer signal of a single sensor and does not

perform fusion with the gyroscope signal. Actually,

thanks to the fusion of gyroscope and accelerometer

signals, our algorithm allows to achieve a higher

precision. This is paid in terms of the number of

calculations, but the low pass filtering in the

equations above needs a lower number and a lower

rate of accesses to the microcontroller memory,

respect to the root mean square method. As a matter

of fact, we tested both algorithms and the proposed

one exhibited better performance in terms of

precision, with comparable calculation time.

A population of sixteen patients of different age

and sex, at different stages of the disease was asked

BIODEVICES 2016 - 9th International Conference on Biomedical Electronics and Devices

156

to wear the two sensors and make an exercise several

times. The population is described in Table I. The

exercise was: walking some steps, passing through an

open door, turning and going back.

Table 1: Sex, age, disease stage of monitored patients.

male/female

under/over 65 early/advanced

9/7 5/11 6/10

FOG events occurred frequently during the

exercises, especially when passing through the open

door and during turning. In order to classify properly

the states, a preliminary calibration of the system was

performed. To this aim, the whole exercise was

filmed with a camera and the sensor signals were

recorded. The films were studied by doctors, who

indicated the exact timing of the freezing events.

Then, the calculated K curves were compared with

the clinical observations by the doctors. This allowed

defining three threshold values of K (T1-3) which

classify the four states: regular gait (K>T3), pre-post

freezing-state (T3>K >T2), involuntary freezing state

(T2>K>T1) and voluntary rest state (K<T1). It is

worth noticing that the values of T1-3 are the same

for all the patients. From clinical side, distinguishing

the involuntary freezing state from the rest state is

crucial, and, fortunately, it relatively simple using

inertial sensors since in the involuntary freezing state

the muscle activity is always present and gives rise to

lots of small movements which are clearly detected

by the sensors. An example of α, ω and K is shown

in the diagrams of Fig.11. Clinical report by the

doctor about the exact FOG timing is sketched in the

bottom diagram. The comparison between the K

curve and the clinical reports allowed defining the T

thresholds and the four classified states. In the

example of Fig.11, a few FOG and pre-FOG events

were identified by both doctors and the system. In one

case (time=23-28s), the system distinguished

between pre-FOG and FOG states, whereas doctors

reported just a FOG in the whole time interval. Values

of T

1-3

remained the same along all the measurements.

Subsequent cross checks outlined an excellent

agreement between the doctors reports and the

automatic recognition of FOG performed by our

system. An extremely low number of errors (false

positive or false negative) were found. The particular

algorithm implemented allowed to get the best

performance published to date in terms of sensitivity,

precision, accuracy and specificity. The average

results on about two hours recording time and sixteen

patients are shown in Table 2. As a comparison with

the state of art, another system using inertial sensors

positioned on the ankles featured a sensitivity of 77

% and a specificity of 86.5 % (S. Mazilu et al., 2013).

This result was obtained on a population of fourteen

PD patients.

Figure 11: An example of angle and angular velocity

measured by the sensor on the shin during the exercise. The

calculate K index is also displayed. The bottom diagram

reports the clinical observation of the FOG events timing.

Table 2: Performance of the system.

Sensitivity Specificity Precision

Accuracy

94.5% 96.7% 93.8% 95.6%

4 CONCLUSIONS

In this paper we proposed the realization of two

wearable wireless sensing systems based on silicon

integrated micro-electro-mechanical inertial sensors

able to recognize in real time specific kinetic features

associated to human motion disorders. The system is

designed specifically for the Parkinson’s disease and

gives an auditory stimulation to the patient to release

block states in the freezing of gait. One system has

the sensors in a headset, while the other one has

sensors on the shins. They can be used at home or

outdoor, during the daily activity of the patient. The

hardware used for the two solutions is the same and

uses the same integrated sensors. On the contrary,

different algorithms were implemented in the two

cases, accounting for the distinct peculiarities of the

Smart Sensing System for the Detection of Speciﬁc Human Motion Symptoms of the Parkinson’s Disease

157

two solutions. In the case of the headset, a number of

different algorithms were used in order to improve

recognition of the gait features. In the paper, we

discussed results regarding the recognition of the

block state, regular and irregular steps, trunk

oscillations obtained with an artificial neural network.

For the system with sensors on the shins we used an

algorithm filtering and processing in real time the

angular velocities of the two legs. This gives excellent

recognition of the irregularities of each step and

detects even barely perceptible tremors in all the

monitored PD patients, allowing distinguishing

doubtless between the voluntary stop state and the

involuntary block due to the FOG. Further

optimization and simplification of the detection

algorithm can be achieved by better manipulating the

quaternions representation of the limbs.

The headset has advantages in terms emphasized

sensitivity to trunk oscillations, easy wearability and

direct auditory feedback. This implies an excellent

detection of specific typologies of motion disorders,

and makes the system compact and energy efficient

since gives the audio-feedback without any

wired/wireless connection. Unfortunately, PD

patients feature a great variety of postural problems

and irregular movements in all the sections of the

body, and the system suffers from the presence of a

joint (the neck) which can mix (or hide part of) the

signals, making the information extremely "noisy"

and confused, thus introducing false positive and

false negative outputs. For this reason, the headset

can be better employed for other types of motion

disorders, as in the case of temporary orthopedics

ones.

The other device requires an additional device in

the ear for the audio-feedback, but guarantees the best

performances presented in literature to date in terms

of sensitivity, specificity, precision and accuracy in

the detection of the FOG events. The system was

validated on a population of sixteen patients of

different age, sex and stage of the disease.

ACKNOWLEDGEMENTS

Authors thank the patients for credit and forbearance,

and STMicroelectronics for providing the NeMEMSi

boards.

REFERENCES

Arias, P. and Cudeiro, J. "Effect of Rhythmic Auditory

Stimulation on Gait in Parkinsonian Patients with and

without Freezing of Gait" PLoS ONE. Vol.5 (2010).

Atallah, L., et al. "Gait asymmetry detection in older adults

using a light ear-worn sensor." Physiological

measurement 35.5 (2014): N29.

Bachlin, M. et al. "A wearable system to assist walking of

Parkinson s disease patients" Methods Inf Med, vol.49,

pp.88–95, 2010.

Bishop, C.M. et al. “Neural networks for pattern

recognition. Clarendon, Oxford (1995).

Bloem, B.R. et al. "Falls and freezing of gait in Parkinson's

disease: a review of two interconnected, episodic

phenomena." Movement Disorders 19.8 (2004): 871-

884.

Chau T. “A review of analytical techniques for gait data”

Part 2: neural network and wavelet methods. Gait &

Posture. Vol.13 (2001) 102–120.

Cola, G. et al. "An On-Node Processing Approach for

Anomaly Detection in Gait." Sensors Journal, IEEE ,

vol.15, no.11, pp.6640-6649, Nov. (2015).

Comotti, D. et al. "neMEMSi: One step forward in wireless

attitude and heading reference systems" Inertial Sensors

and Systems (ISISS), Intern. Symposium on 1–4

(2014).

Keogh E. and Ratanamahatana C. A.: Exact indexing of

dynamic time warping. Knowledge and information

systems. Vol.7 (2005) 358–386.

Kline D. M. and Berardi V. L.: Revisiting squared-error and

cross-entropy functions for training neural network

classifiers. Neural Computing & Applications. Vol.14

(2005) 310–318.

Kwon, Yuri, et al. "A practical method for the detection of

freezing of gait in patients with Parkinson’s disease."

Clinical interventions in aging 9 (2014): 1709.

Lorenzi, P. et al. “Wearable Wireless Inertial Sensors for

Long-Time Monitoring of Specific Motor Symptoms in

Parkinson’s Disease” BIODEVICES 2015 - 8th Int.

Conference on Biomedical Electronics and Devices,

Proceedings; Part of 7th International Joint Conference

on Biomedical Engineering Systems and Technologies,

BIOSTEC 2015, Lisbon (P) pp.168-173.

Madgwick, S.O.H et al. "Estimation of IMU and MARG

orientation using a gradient descent algorithm", IEEE

International Conference on Rehabilitation Robotics ,

Rehab Week Zurich, ETH Zurich Science City,

Switzerland, 2011.

Mazilu, S. et al. "Feature learning for detection and

prediction of freezing of gait in Parkinson’s disease."

Machine Learning and Data Mining in Pattern

Recognition. Springer Berlin Heidelberg, 2013. 144-

158.

Mazilu, S. et al. "GaitAssist: A Daily-life Support and

Training System for Parkinson's Disease Patients with

Freezing of Gait", Proceedings of the SIGCHI

Conference on Human Factors in Computing Systems

2531–2540 (2014).

Moore, S. et al. "Autonomous identification of freezing of

gait in Parkinson's disease from lower-body segmental

accelerometry" Journal of NeuroEngineering and

Rehabilitation, vol.10, pp.19, 2013.

BIODEVICES 2016 - 9th International Conference on Biomedical Electronics and Devices

158

Nieuwboer, A. et al. "Abnormalities of the spatiotemporal

characteristics of gait at the onset of freezing in

Parkinson's disease." Movement Disorders 16.6 (2001):

1066-1075.

Sijobert, B. et al. "IMU Based Detection of Freezing of Gait

and Festination in Parkinson's Disease" IFESS

MALAYSIA, 2014.

Spildooren, J. et al "Freezing of gait in Parkinson's disease:

the impact of dualtasking and turning" Movement

Disorders25.15 (2010)2563-70.

Wang, Kongming, et al. Alignment of curves by dynamic

time warping. The Annals of Statistics, (1997), 25.3:

1251-1276.

Smart Sensing System for the Detection of Speciﬁc Human Motion Symptoms of the Parkinson’s Disease

159