Automatic Identification of Motor Patterns Leading to Freezing of

Gait in Parkinson’s Disease

An Exploratory Study

Luca Palmerini

, Laura Rocchi

, Jeffrey M. Hausdorff

and Lorenzo Chiari

Department of Electrical, Electronic and Information Engineering “Guglielmo Marconi”, University of Bologna,

Bologna, Italy

Laboratory for Gait & Neurodynamics, Movement Disorders Unit, Department of Neurology,

Tel-Aviv Sourasky Medical Center, Tel-Aviv, Israel

Keywords: Freezing, Parkinson’s Disease, Symbolic Aggregate Approximation, Acceleration, Wearable Sensors.

Abstract: Freezing of gait (FOG) is a common and disabling gait disturbance among patients with advanced

Parkinson’s Disease (PD). FOG episodes are often overcome using attention or cues from the environment.

Hence, identification of events prior to FOG may be very effective to improve mobility in PD patients.

Previous work has suggested that there are changes in the gait pattern just prior to freezing. Nonetheless,

little work has been done to explore the possibility of identifying motor patterns that are characteristic of the

pre-FOG phase (few seconds before the FOG). We analysed the acceleration signals from sensors worn on

the ankle, thigh, and trunk of eight patients with PD who experienced freezing. We translated windows of

the raw signals in symbols by using Symbolic Aggregate approXimation. The aim was to discriminate the

patterns of symbols characterizing pre-FOG from the ones characterizing normal activity (standing and

walking with no FOG). Sensitivity over 50% and Specificity over 70% were obtained by using a classifier

on symbolic data, with different combinations of sensor position/sampling/windows duration. These

preliminary findings demonstrate that it is possible to automatically identify (some of) the motor patterns

that eventually lead to FOG events before they occur by using wearable sensors.

1 INTRODUCTION

Freezing of gait (FOG) is a disabling gait

disturbance that is common among patients with

advanced Parkinson’s Disease (PD). FOG can

manifest as an impairment of the initiation and

termination of gait and as a sudden interruption of

walking (Nieuwboer, 2004). During the FOG

episode the patients feel as if their feet are glued to

the ground and cannot resume walking. Recent work

has focused on using wearable motion sensors to

detect freezing of gait (FOG) as soon as it starts

(Bächlin, 2010); (Moore, 2013); (Mazilu, 2013),

obtaining satisfactory accuracies. Although previous

work has suggested that there are changes in the gait

pattern just prior to freezing (Nieuwboer, 2004),

only one work has recently explored the possibility

of identifying motor patterns that are characteristic

of the pre-FOG phase (i.e., a few seconds before the

FOG happens) (Mazilu, 2013). FOG episodes are

often overcome using attention or cues; hence,

identification of events prior to FOG may be very

effective to improve mobility in PD patients by

producing an auditory cue just before the FOG

starts. This is why in the current study we focused

on identification of the pre-FOG phase: we analyzed

the acceleration signals from sensors worn on the

ankle, thigh, and trunk of eight patients with

Parkinson’s disease who experienced freezing

(Bächlin, 2010). We translated windows of the raw

signals in sequences of symbols by using SAX

(Symbolic Aggregate approXimation) (Lin, 2002;

2003). A previous work applied this technique to

study gait symmetry in patients with PD (Sant’Anna,

2011). The aim of the current study was to

discriminate the patterns of symbols characterizing

pre-FOG from the ones characterizing normal

activity (i.e., standing and walking with no FOG).

730

Palmerini L., Rocchi L., M. Hausdorff J. and Chiari L..

Automatic Identiﬁcation of Motor Patterns Leading to Freezing of Gait in Parkinson’s Disease - An Exploratory Study.

DOI: 10.5220/0004912107300734

In Proceedings of the 3rd International Conference on Pattern Recognition Applications and Methods (ICPRAM-2014), pages 730-734

ISBN: 978-989-758-018-5

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

Figure 1: Acceleration signal converted in symbols.

2 METHODS

Symbolic data analysis was applied on the Daphnet

Database, which was described in (Bächlin, 2010).,

and which stores acceleration signals from

acceleration sensors positioned monolaterally on the

ankle, thigh and trunk of 8 patients with PD who

experienced freezing (the two patients who did not

experience freezing were not considered in this

study). The acceleration signals available for each

sensor were: Antero-Posterior (forward), Medio-

Lateral (horizontal), and Vertical. We chose to

consider the norm of the three acceleration signals

for the following analysis.

Parts of signal which did not correspond to the

experiment part (as explained in the Daphnet

documentation) were deleted. Resulting signals were

normalized (z-score).

As exemplified in Fig. 1, we divided the signals

in a No-Event part (i.e., standing and walking with

no FOG), followed by a pre-FOG window (i.e., few

seconds before the FOG happened), followed by a

FOG part (i.e. during FOG).

The FOG part was not considered in this

exploratory analysis, since we were interested in

identifying the difference between normal activity

and pre-FOG.

The pre-FOG window was defined as a period of

1, 2, or 3 seconds before the FOG. Correspondingly,

the “No-Event” part of the signal was divided in

consecutive No-Event windows of the same duration

as the pre-FOG windows.

Since the No-Event part of the signal was

generally much longer than the pre-FOG window,

the No-Event windows were much more than the

pre-FOG windows, resulting in an unbalanced

dataset (Table 1): later in this section, this issue will

be dealt with.

The different windows duration were considered

in order to see if there was a duration which could

allow a better identification of the pre-FOG patterns.

When the time between two following FOG

events was less than 3 seconds, no pre-FOG window

was considered, for any window durations.

This was done for two reasons:

- In order to obtain a fair comparison between

different window durations: in fact in this way

the number of pre-FOG windows is the same for

all the different durations (see Table 1).

- In order to avoid that the pre-FOG window

would partly capture patterns of the previous

FOG.

Table 1: Mean and std values of the number of pre-FOG

and No-Event windows across all subjects.

Win Duration

num. pre-FOG

windows

num. No-Event

windows

1 25.75±14.2 1427±409

2 25.75±14.2 693±203

3 25.75±14.2 448±134

AutomaticIdentificationofMotorPatternsLeadingtoFreezingofGaitinParkinson'sDisease-AnExploratoryStudy

731

Figure 2: Results of sensitivity (± STD) and specificity (± STD) as function of different combinations of observation

window duration, symbolic frequency, and sensor position.

The obtained windows were translated into

sequences of symbols by using the SAX algorithm

(Lin, 2003). Symbols can be considered as letters

from an alphabet (we arbitrarily considered a 10

symbols alphabet) that represent the considered

window instead of the raw acceleration values (as

shown in Fig. 1).

In SAX, a single symbol represents consecutive

samples of the raw signals in the considered

window, thus automatically performing

dimensionality reduction. Therefore the symbolic

data will have a new (lower) sampling frequency.

The original sampling frequency of acceleration

signals was 64 Hz: in order to choose the optimal

symbolic sampling frequency (how many symbols in

one second), we considered different options: 8, 16,

and 32 Hz, corresponding to translate 8, 4, 2 original

samples in one symbol respectively.

In order to find patterns of symbols which are

characteristic of the pre-FOG window (Fig. 1), and

to discriminate them from the No-Event windows,

we used the K-nearest neighbour’s classifier (with

k=1). Instead of the Euclidean distance, we used the

symbolic distance between sequences of symbols,

which is defined in (Lin, 2002; 2003).

In order to obtain a method that would be as

generalizable as possible and that would perform

well regardless of the different patients considered,

we used a leave-one-subject-out cross validation to

determine the accuracy of the proposed approach. In

the leave-one-subject-out cross validation the data of

one patient (all his/her signal segments) is classified

by using the data captured from the rest of patients.

The results will be presented in terms of

sensitivity (proportion of pre-FOG windows which

are correctly identified) and specificity (proportion

of No-Event windows which are correctly

identified).

Since the dataset is highly unbalanced between

the two classes (see Table 1), a random under-

ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods

732

sampling of the majority class (No-Event) was

performed in the training phase of the classifier. This

was done in order to have the same number of Pre-

FOG and No Event windows to train the classifier

with. If no under-sampling had been performed, the

classifier would have “learned” mostly No-Event

patterns thus leading to high specificity but very

poor sensitivity.

In order to test the significance of results, a

random classifier was made, which randomly

assigned “No-Event” or “pre-FOG” classes based on

the proportion of classes in the under-sampled

training set. One would expect such a classifier to

perform with sensitivity and specificity around 50%.

3 RESULTS AND DISCUSSION

Results are reported in Fig. 2.

The best obtained result, in terms of trade-off

between sensitivity and specificity (arithmetic

mean), was

- Sensitivity: 66.5 %

- Specificity: 73.9 %,

In the following, the details of all the parameters of

the data analysis corresponding to this result are

listed:

- Thigh sensor

- Norm of the signal

- Symbolic Frequency of 16 Hz

- Duration of the windows of 2 seconds

- Alphabet size of 10

Both sensitivity and specificity of this combination

resulted significantly better than the ones of the

random classifier (Fig. 3), which performed, as

expected, with sensitivity and specificity around

50%.

Figure 3: Comparison between the performance of the best

classifier and of the random classifier.

From results in Fig. 2 it can be noted that

different combinations of sensors/signals/ frequency

can lead to higher specificity or higher sensitivity

(but not to both).

From results in Fig. 2 it can also be noted that

thigh sensor seems to perform generally better than

ankle and trunk sensors in sensitivity, and

comparably in specificity.

Also, sensitivity estimates tend to be less

consistent (higher variability of performance across

subjects) than specificity estimates.

Although the best result is obtained with a 2-

seconds window, it seems that there is not a clear

difference or pattern in considering windows of

different durations.

Finally, considering different symbolic

frequencies leads to different combinations of

sensitivity and specificity but no consistent pattern

can be observed (e.g. higher symbolic frequency

always leads to better sensitivity/specificity).

Interestingly, the sensitivity in discriminating

between pre-FOG patterns and normal activity is

comparable to the sensitivity in discriminating

between FOG patterns and normal activities

obtained by previous studies (73.1% in Bächlin

2010, 66.3% in Mazilu 2012, 68.5% in Mazilu

2013).

On the other hand, specificity is lower than the

ones obtained in those studies (81.6% in Bächlin

2010, 95.4% in Mazilu 2012, 86.8% in Mazilu

2013).

However, an overall lower performance was

expected because the task of discriminating the

patterns before the event occurs is generally more

complex than detecting the event after it has

happened.

These preliminary findings demonstrate that it is

possible to identify (some of) the motor patterns that

eventually lead to FOG events before they occur,

support the idea that the gait pattern changes prior to

freezing, and suggest that this pre-event period can

be automatically identified by using wearable

sensors.

As a limitation of this study, the algorithm

presented in this study was not optimized for speed;

in future work, a real-time implementation should be

done.

Moreover, the use of different classifiers and the

fusion of decisions made from different

combinations of sensors, time windows and

frequencies, could possibly permit to improve the

obtained results.

AutomaticIdentificationofMotorPatternsLeadingtoFreezingofGaitinParkinson'sDisease-AnExploratoryStudy

733

ACKNOWLEDGEMENTS

The research leading to these results has received

funding from the European Union - Seventh

Framework Programme (FP7/2007-2013) under

grant agreement n°288516 (CuPiD project).

REFERENCES

Bächlin M., Plotnik M., Roggen D., Maidan I., Hausdorff

J. M., Giladi N., Tröster G., Wearable Assistant for

Parkinson’s Disease Patients With the Freezing of Gait

Symptom, IEEE Trans on Information Technology in

Biomedicine, 14(2), March 2010, pages 436-446.

Lin J., Keogh E., Patel P., Lonardi S., Finding Motifs in

Time Series, proceedings of the 2nd Workshop on

Temporal Data Mining, 8th ACM SIGKDD

International Conference on Knowledge Discovery

and Data Mining. Edmonton, Alberta, Canada. July

23-26, 2002

Lin J., Keogh E., Lonardi S., Chiu B., A Symbolic

Representation of Time Series, with Implications for

Streaming Algorithms, proceedings of the 8th ACM

SIGMOD Workshop on Research Issues in Data

Mining and Knowledge Discovery. San Diego, CA.

June 13, 2003.

Mazilu S., Hardegger M., Zhu Z., Roggen D., Tröster G.,

Plotnik M. and Hausdorff J. M., Online Detection of

Freezing of Gait with Smartphones and Machine

Learning Techniques, 6th International Conference on

Pervasive Computing Technologies for Healthcare

(PervasiveHealth), 2012

Mazilu S., Calatroni A., Gazit E., Roggen D., Hausdorff J.

M., and Tröster G., Feature Learning for Detection and

Prediction of Freezing of Gait in Parkinson’s Disease,

MLDM, Lecture Notes in Computer Science (LNCS),

Springer 2013.

Moore S. T., Yungher D. A., Morris T. R., Dilda V.,

MacDougall H. G., Shine J. M., Naismith S. L., Lewis

S. J., Autonomous identification of freezing of gait in

Parkinson’s disease from lower-body segmental

accelerometry. Journal of neuroengineering and

rehabilitation, 10(1), 19. 2013

Nieuwboer A., Dom R., De Weerdt W., Desloovere K.,

Janssens L., Stijn V., Electromyographic profiles of

gait prior to onset of freezing episodes in patients with

Parkinson's disease. Brain. 2004 Jul;127(Pt 7):1650-

60.

Sant’Anna et al., “A new measure of movement symmetry

in early Parkinson’s disease patients using symbolic

processing of inertial sensor data.,” IEEE trans on bio-

medical engineering, vol. 58, no. 7, pp. 2127–35, Jul.

2011.

ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods

734