ACCELEROMETER BASED GAIT ANALYSIS

Multi Variate Assessment of Fall Risk with FD-NEAT

Bart Jansen

, Maxine Tan

, Ivan Bautmans

2,3,4,5

, Bart Van Keymolen

2,3

Tony Mets

2,3,4

and Rudi Deklerck

Departement Electronics and Informatic and IBBT, Vrije Universiteit Brussel, Pleinlaan 2, 1050 Brussels, Belgium

Gerontology, Frailty in Ageing Research Department, Vrije Universiteit Brussel

Laarbeeklaan 103, 1090 Brussels, Belgium

Frailty in Ageing Research Department, Vrije Universiteit Brussel, Laarbeeklaan 103, 1090 Brussels, Belgium

Geriatrics, Universitair Ziekenhuis Brussel, Laarbeeklaan 101, 1090 Brussels, Belgium

Stichting Opleiding Musculoskeletale Therapie (SOMT), Softwareweg 5, 3821 BN Amersfoort, The Netherlands

Keywords: Accelerometry, Fall risk, Gait analysis, Step time asymmetry, Classification, Feature selection, FD-NEAT.

Abstract: This paper describes an accelerometer based gait analysis system for the assessment of fall risk. The

assessment is based on 22 different features calculated from the signal. The different features are combined

using machine learning algorithms in order to decide whether the subject has an increased fall risk. Results

from Naive Bayes, Neural Networks, Locally Weighted Learning, Support Vector Machines and C4.5 are

reported and compared. It is argued that the neural networks provide low accuracy results because of the

high dimensionality of the feature space compared to the available data. It is shown that FD-NEAT (a

method from neuro evolution which simultaneously learns the network topology, the network weights and

the relevant features) outperforms the other methods in the given classification task. The system is evaluated

on a database consisting of 40 elderly with known fall risk and 40 healthy elderly controls.

1 INTRODUCTION

The field of accelerometer based fall risk assessment

is characterised by an intense debate on the

relevance of some specific features calculated from

the accelerometer signal in a univariate classification

problem on the distinction between fallers from non

fallers (e.g. Moe-Nilssen and Helbostad, 2005).

Many of these features are known for decades

(e.g. a decrease in step length is related to an

increase in fall risk); others could only be described

since the availability of small, battery powered

accelerometer sensors (e.g. step regularity and

symmetry). Only recently, the validity, reliability

and repeatability of most of these features in the

context of fall risk assessment have been described

in the clinical literature (e.g. Moe-Nilssen, 1998).

For some specific diseases and conditions having

a direct impact on the gait pattern, it is well

described how the disease is affecting the

neurological or muscolatory system and how this

affects the accelerometer based gait features. For

example, in Parkinson patients, the freezing of gait

increases the variability in stride time and the effect

of specific treatment on the freezing of gait can be

evaluated by investigating stride time variability

(Hausdorff et al, 2005).

However, in a majority of the growing

population of elderly, an increase in fall risk cannot

directly be attributed to a specific disease. Rather, a

condition of general frailty, multiple chronic

diseases and a general decrease in mobility all

together contribute to the increased fall risk.

Therefore, it is to be expected that in a general

population of elderly, a less clear relationship

between single accelerometer based gait features and

fall risk can be observed. However, very few

attention was paid so far to the construction of

intelligent multi-variate classifiers for fall risk

assessment.

This paper evaluates the use of the FD-NEAT

algorithm (Tan et al, 2009) for the classification of a

population of eighty elderly into a class of elderly

presenting increased fall-risk and a class of elderly

without an increased risk of falling, based on a wide

variety of accelerometer based gait features.

138

Jansen B., Tan M., Bautmans I., Van Keymolen B., Mets T. and Deklerck R..

ACCELEROMETER BASED GAIT ANALYSIS - Multi Variate Assessment of Fall Risk with FD-NEAT.

DOI: 10.5220/0003132701380143

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2011), pages 138-143

ISBN: 978-989-8425-35-5

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

From a machine learning perspective, the

problem is far from trivial, as there is extremely few

training data available, compared to the large

number of features considered. We will show that

FD-NEAT outperforms traditional neural networks

and other machine learning algorithms, because it

better copes with the dimensionality problem, by

means of intelligent feature selection. Because of the

clinically unclear relationship between many of the

features used in this paper and fall risk in a general

population of elderly, no a priori feature selection

was performed based on the available medical

knowledge.

2 PREVIOUS WORK

2.1 Accelerometer based Fall Risk

Assessment

Most studies in accelerometer based gait analysis for

fall risk assessment are focusing on the repeatability

or validity of single outcomes. Only very few studies

are combining multiple outcomes in intelligent

classifiers. Recently, Marschollek et al. showed how

measures obtained from accelerometry could be

combined with clinical scores, in order to

discriminate between fallers and non-fallers during

an instrumented timed up and go (TUG) test

(Marschollek et al., 2009). Another study,

combining results from accelerometer based TUG

tests, stepping tests and a sit-to-stand tests, is

reported by (Narayanan, 2010).

(Swanenburg et al, 2010) report on a one year

prospective study of fall risk assessment, based on

features calculated from force plates. Intelligent

classifiers based on neural networks were also used

for fall risk assessment from posturography tasks,

instrumented with accelerometer and gyroscopes

(Giansanti et al. 2008).

3 METHODS

3.1 Subjects

Eighty subjects participated in this study. They

consisted of two groups (n=40 each): elderly with

known fall risk (EF) and elderly controls (EC), see

table 1. Each group contained twenty males and

twenty females.

Table 1: General subject information. None of the

parameters was significantly different between the groups

(ANOVA p < 0.05). EF = elderly with increased fall risk,

EC = elderly controls. Standard deviations are shown

between brackets.

Group

EF (n=40) EC (n=40)

Age 80.59 (5.38) 79.03 (4.95)

Weight 66.89 (14.97) 69.74 (11.56)

Length 1.62 (0.12) 1.64 (0.08)

BMI 25.46 (4.29) 25.61 (3.93)

All subjects were older than 70. Known fall risk

was defined as a reported history of falls and/or

Timed-Get-Up-and-Go-Test > 15s and/or Tinetti test

≤ 24/28. The local ethical committee approved this

study and all participants provided their written

informed consent.

3.2 Data Acquisition

The DynaPort Minimod tri-axial accelerometer

(McRoberts BV, The Hague, The Netherlands) was

placed at the sacrum of the subjects. The device

stores the accelerometer signal on a standard SD-

card. Before every walking episode, the SD-card

was emptied, put in the sensor and the sensor was

restarted. After every walking episode, the SD-card

was placed in the laptop and the acceleration data

along the three axes was read out using the Mira

software (same manufacturer) and exported to plain

text files. The plain text files were loaded into our

own gait analysis toolbox, an in-house developed

software package programmed in C#.

3.3 Test Procedure

Subjects were asked to walk a straight line trajectory

of 18 meters, separated by two clear lines on the

floor. Subjects started with both feet in front of the

first line and stopped when the second foot landed

beyond the second line. The distance between the

stop line and the final heel strike was measured and

added to the 18 meter to obtain the total distance

walked. At the beginning of each walk, the observer

placed the sensor at the sacrum and initialized the

sensor as described above. Subjects were always

instructed to walk at preferred speed and no walking

aids were allowed.

3.4 Statistical Analysis

A dataset of eighty elderly consisting of forty elderly

with increased fall risk and forty controls is

available. Compared to similar datasets used in

ACCELEROMETER BASED GAIT ANALYSIS - Multi Variate Assessment of Fall Risk with FD-NEAT

139

accelerometer based gait analysis, this is quite a

large set. However, from a machine learning

perspective given the high amount of included

features, its size is extremely small. Therefore, it

was decided not to partition the dataset into a single

training set and a single test set, but ten-fold cross

validation was used (Duda et al, 2000).

Data analysis was performed using SPSS version

17, WEKA and Excel. For each of the studied

machine learning algorithms averages over the ten

folds of accuracy, true positives, false positives,

precision, recall and Area under the Curve (AUC)

are provided.

Significant differences among outcome measures

are evaluated using a one way ANOVA with

significance set to p<0.05 and with a post-hoc

Bonferroni test to identify two differing measures.

Correlations between different types of features

were assessed by calculating Pearson’s correlation

coefficient.

3.5 Data Analysis

In total, 22 features were calculated from the

accelerometer signal. Features can be divided into

five groups: step count, step time (and derived

statistics), step length (and derived statistics), step

symmetry and step RMS. Each of the five groups is

explained below.

3.5.1 Step Count

The 3D accelerometer signal is rotated to align the Y

axis of the signal to gravity and steps (defined as

initial contacts of the heel (IC)) are identified, based

on the maxima before the zero-crossings in the

forward acceleration signal, after applying a fourth

order zero lag Butterworth low pass filter with a cut-

off frequency of 2 Hz (Zijlstra, 2004).

3.5.2 Step Time

From the IC’s detected from the signals as described

above, the average step time is obtained, as well as a

range of derived statistics including standard

deviation, coefficient of variation, inter quartile

range etc. Also, step frequency, walking speed and

step time asymmetry are available. Step time

asymmetry is the difference of the left step time and

the right step time, scaled by the average and

expressed as a percentage (equation 2).





=200





− 









+ 



(2)

In all features based on step time, the initial two

steps and the final two steps are discarded from the

signal, in order to exclude effects from gait initiation

and gait termination. The study of irregularities in

the gait initiation and termination phases is beyond

the scope of this paper.

3.5.3 Step Length

As the total length of the trajectory and the number

of steps are known, the average step length is

available. Step length can also be calculated without

relying on the measured true trajectory length using

the inverted pendulum model (Zijlstra, 2004). The

inverted pendulum model is a biomechanical model

of human gait, which is relating a vertical movement

of the pelvis during the gait cycle with the step

length, as specified in equation 3:

ℎ=2



2ℎ− ℎ



(3)

where l is the leg length and h is the vertical

displacement of the pelvis, which is obtained from

the double integration of the vertical acceleration

component. The advantage of this method is that the

step length of each gait cycle can be calculated

individually, using only two parameters. The

disadvantage of this method is that the double

integration step in calculating the vertical

displacement is prone to drift. Step length as

calculated from the known trajectory length, step

length according to the inverted pendulum model

and the vertical displacement of the pelvis itself are

included as features of this study.

3.5.4 Step and Stride

Regularity and Symmetry

A whole family of related measures exist which all

capture the regularity of the accelerometer signal

over multiple steps (Moe-Nilssen and Helbostadt,

2004). Suppose y[t] is the auto correlation of the

acceleration signal a[t], then y[t] has maxima

corresponding to a time shift of 1,2,.. k steps. Hence,

the auto correlation at the first maximum expresses

the step regularity, whereas auto-correlation at the

second maximum expresses the stride regularity.

Step symmetry is defined as the step regularity

divided by the stride regularity.

These measures are typically calculated in the

medio-lateral (ML) and cranio-caudal (CC)

direction. Auto correlation could be normalized or

not. Biased and unbiased versions have been

proposed. In this study only the unbiased measures

in the CC orientation were incorporated.

BIOSIGNALS 2011 - International Conference on Bio-inspired Systems and Signal Processing

140

3.5.5 Step and Stride RMS

The root mean squared acceleration per step or per

stride, in the CC and the ML direction were

calculated (equation 4).

=



∑









(4)

is the acceleration in the i-th sample of the

considered step in either the CC or ML orientation

and N is the number of samples in the current step.

RMS values per step are averaged over all steps in

the signal.

3.6 Machine Learning

Algorithms and Classifiers

Standard machine learning methods for pattern

recognition and classification were employed:

Naive Bayes (NB), Multi layered perceptron (MLP),

Support Vector Machines (SVM), Locally Weighted

Learning (LWL) and C45. For more information on

any of these classifiers, the reader is referred to

(Duda et al, 2000). WEKA version 3.4.13 was used

for the classification based on each of these

classifiers. Weka is an open source machine learning

software package, provided by the University of

Waikato. For each of the included algorithms, the

default parameters as proposed by WEKA were

used.

Naive Bayes. This is a simple Bayesian classifier,

assuming conditional independence between the

attributes.

Multi layered Perceptron. We have used a MLP

with one hidden layer, consisting of 22 input units,

12 hidden units, which is (nr inputs + output values)

/ 2, and a single output unit. Learning rate was set to

0.3 and momentum to 0.2. In the hidden units a

sigmoid activation function was used, in the output

unit the identity function was used as an activation

function. Back propagation was used as a training

algorithm, performing 500 iterations.

Support Vector Machines. Sequential Minimal

Optimization is used for fast training (Platt J, 1998).

Locally Weighted Learning. This is a nearest

neighbor classifier, which is considering all

neighbors and attributing a weight to each of the

neighbors. The weight scales linear with the inverse

distance to the query point (Atkeson, Moore and

Schaal, 1996).

C45. is a traditional decision tree learning algorithm

introduced by (Quinlan, 1993).

3.7 FD-NEAT

Training neural networks for classifying gait data

with back propagation as used in the MLP, has

several drawbacks: (1) the user must define the

network topology; (2) the user must carefully select

the relevant features, as (3) a fastly growing number

of training instances is required with the addition of

each new attribute. In the current study, only 80

subjects are available, which might hence lead to

low classification accuracy of neural networks.

Approaches based on genetic algorithms were

proposed in which the network topology and the

weights are learnt simultaneously. An example of

such a system is “neuro evolution of augmenting

topologies” or NEAT (Stanley and Miikkulainen,

2002) in which a population of neural networks

evolves from simple perceptrons into more complex

networks, based on mutation (both weights and

connections evolve) and cross-over. NEAT was

shown to perform superior to classical neural nets in

typical benchmark problems. A straightforward

extension of NEAT is “feature selective” NEAT or

FS-NEAT (Whiteson et al, 2005) which performs

feature selection, topology learning and weight

learning simultaneously. In FS-NEAT the initial

networks in the population only have a single input

neuron, randomly selected from the available

attributes. A mutation operator which can connect

additional input nodes is added. A modification to

FS-NEAT is “feature deselective” NEAT or FD-

NEAT (Tan et al, 2009), which is similar but starts

from networks in which all inputs are connected and

has a mutation operator which drops connections

from input nodes. In most cases, FD-NEAT

outperforms FS-NEAT.

In the experiments we report in this article, the

best result out of ten runs was obtained for each fold.

The populations in the genetic algorithm consisted

of 200 networks that were evolved over 60

generations.

4 RESULTS

4.1 Comparison of Classifiers

The main experiment consists of evaluating the

performance of five different machine learning

algorithms in a binary classification problem based

on 22 features calculated from an accelerometer

signal obtained from a single walk of 18 m. Detailed

results are given in table 3. In summary, the results

show that NB outperforms the other classifiers: it

ACCELEROMETER BASED GAIT ANALYSIS - Multi Variate Assessment of Fall Risk with FD-NEAT

141

has the highest accuracy (0.77), true positive rate (=

recall) (0.75), precision (0.79) and area under the

curve (0.82), for the lowest false positive rate (0.2).

The multi layered perceptron scored very low, on

each of the five performance measures incorporated

in this study. Given the high amount of attributes

(22) compared to the low amount of training

instances (72 out of 80 in each fold), it is to be

expected that inferior results of the MLP are due to a

severe lack of training instances compared to the

complexity of the network. Feature selection is

hence appropriate.

Table 2: Results of the classifiers, 22 features. TP = true

positives, FP = false positives, AUC = area under the

curve, NB = Naive Bayes, MLP=Multi Layered

Perceptron, SVM = Support Vector Machine, LWL =

Locally Weighted Learning. X= AUC for FD-NEAT not

available, see text.

Given the inconclusive results of the debate in

the gait analysis community on the relevance of each

of the individual features, it is not advised to

manually select the relevant features. On the other

hand, one of the main observations in the field of

clinical gait analysis is that almost all features are

somehow influenced by walking speed. In this

study, subjects were asked to walk at normal speed.

Hence, differences in any of the calculated features

between EF and EC may be related to gait speed

differences between both groups.

Using FD-NEAT the accuracy increases up to

82.5%. Also TP (recall) and precision are the highest

of all experiments reported, while FP is the lowest of

all reported experiments. For FD-NEAT ROC

analysis with AUC could not be reported, as it uses a

sigmoid in the activation function, resulting in

nearly binary outputs such that threshold varying is

unfeasible.

Standard deviations of the accuracy over the ten

folds were calculated for MLP-22 and FD-NEAT-22

and are quite high (σ=0.17 and 0.17

respectively). This is caused by the too small fold

size (for N=80, the fold size is 8). At the 0.05 level,

accuracy of FD-NEAT-22 is significantly better than

MLP-22.

5 DISCUSSION

From a clinical perspective, this study confirms that

accelerometer based fall risk assessment is feasible

with high accuracy (82.5%) and with high sensitivity

(80% recall). However, the study population was

recruited based on self reported falls, the timed up

and go test and the Tinetti test. Hence, amongst EF,

a wide variety of conditions and diseases which are

possible related to fall risk are present. In a study

design in which only fallers suffering from a specific

disease or condition are included (e.g. sarcopenia or

Alzheimer’s disease), higher accuracy results could

probably be obtained, using another subset of

features, as each disease results in specific gait

disorders. As this article is focussing on the

screening potential of accelerometer based gait

analysis for fall risk, we have chosen not to restrict

the study population to specific subgroups.

Most measures employed do not show significant

differences between the different classifiers studied.

This is due to the high standard deviations as the

sizes of the folds studied are extremely small.

However, validating over the test set was not

considered a viable approach. Hence, it is to be

advised to repeat the experiment over larger

population sizes in order to reach significance.

Nevertheless, FD-NEAT based on 22 features

significantly outperforms MLP based on 22 features,

confirming our initial hypothesis that FD-NEAT

suffers less from the described dimensionality

problems.

6 CONCLUSIONS

This article evaluated the possibility of fall risk

stratification of elderly based on a single walk of 18

meters, instrumented with an accelerometer.

Opposed to many systems, the system is not limited

to a single feature. We’ve investigated the

performance of five different classifiers using 22

features, commonly used in various gait

experiments. Given the extremely small data set (40

positive and 40 negative cases) compared to the

number of attributes (22), the performance of the

classifiers is suboptimal (60 to 70 % accuracy).

We’ve put forward that FD-NEAT, an evolutionary

approach to perform feature selection, to learn a

accuracy

recision

recall

AUC

NB-22 .77 .75 .2 .79 .75 .82

MLP-22 .61 .6 .38 .62 .6 .72

SVM-22 .69 .6 .23 .73 .6 .69

LWL-22 .69 .58 .2 .74 .58 .74

C45-22 .69 .65 .28 .70 .65 .64

FD-NEAT .82 .8 .15 .84 .8 X

BIOSIGNALS 2011 - International Conference on Bio-inspired Systems and Signal Processing

142

neural network topology and to learn the weights

simultaneously, outperforms the traditional

classifiers (82.5 % accuracy).

From a clinical perspective, this article illustrates

that in a general population of elderly, fall risk is

related to different underlying constructs, with clear

manifestations among different dimensions in the

gait pattern as captured by the accelerometer.

REFERENCES

Atkeson C., Moore A., and Schaal S., 1996. "Locally

weighted learning" AI Reviews.

Duda, R. O., Hart, P. E., & Stork, D. G. (2000). Pattern

Classification. Wiley-Interscience.

Giansanti, D., Maccioni, G., Cesarino, S., Benvenuti, F., &

Maccellari, V. (2008). Assessment of fall-risk by

means of a neural network based on parameters

assessed by a wearable device during posturography.

Medical Engineering and Physics, 30, 367-372.

Hausdorff, J. M., Schaafsma, J. D., Balash, Y., Bartels, a.

L., Gurevich, T., Giladi, N., et al. (2003). Impaired

regulation of stride variability in Parkinson's disease

subjects with freezing of gait. Experimental brain

research. Experimentelle Hirnforschung.

Expérimentation cérébrale, 149(2), 187-94.

Hausdorff J. M., 2005. Gait variability: methods, modeling

and meaning. Journal of NeuroEngineering and

Rehabilitation. 9:1-9.

Marschollek, M., Nemitz, G., Gietzelt, M., Wolf, K. H.,

Meyer Zu Schwabedissen, H., Haux, R., et al., 2009.

Predicting in-patient falls in a geriatric clinic: a

clinical study combining assessment data and simple

sensory gait measurements. Zeitschrift für

Gerontologie und Geriatrie : Organ der Deutschen

Gesellschaft für Gerontologie und Geriatrie, 42(4),

317-21.

Marschollek M., Wolf K., Plischke M., Haux R., 2006.

Classification of activities of daily life from long-term

realistic multi-sensor data. In: Proceedings of IEEE

Health Pervasive Systems HPS06.

Moe-Nilssen, R. (1998). Test-retest reliability of trunk

accelerometry during standing and walking *1, *2.

Archives of Physical Medicine and Rehabilitation,

79(11), 1377-1385.

Moe-Nilssen R., Helbostad J. L., 2004. Estimation of gait

cycle characteristics by trunk accelerometry. Journal

of Biomechanics. 37(1):121-126.

Moe-Nilssen, R., & Helbostad, J. L. (2005). Interstride

trunk acceleration variability but not step width

variability can differentiate between fit and frail older

adults. Gait & posture, 21(2), 164-70.

Narayanan, M. R., Redmond, S. J., Scalzi, M. E., Lord, S.

R., Celler, B. G., Lovell Ast, N. H., et al., 2010.

Longitudinal falls-risk estimation using triaxial

accelerometry. IEEE transactions on bio-medical

engineering, 57(3), 534-41.

Oshima Y., Kawaguchi K., Tanaka S., et al., 2010.

Classifying household and locomotive activities using

a triaxial accelerometer. Gait & posture. 31(3):370-4.

Platt J., 1998. "Fast Training of Support Vector Machines

using Sequential Minimal Optimization". Advances in

Kernel Methods-Support Vector Learning, B. Schoel-

kopf, C. Burges, and A. Smola, eds., MIT Press.

Quinlan R., 1993. "C4.5: Programs for Machine

Learning", Morgan Kaufmann Publishers, San Mateo.

Stanley K. O., Miikkulainen R., 2002. Evolving neural

networks through augmenting topologies. Evol

Comput 10:99–127.

Swanenburg, J., de Bruin, E. D., Uebelhart, D., & Mulder,

T. (2010). Falls prediction in elderly people: a 1-year

prospective study. Gait & posture, 31(3), 317-21.

Tan M., Hartley M., Bister M., Deklerck R., 2009.

Automated feature selection in neuroevolution.

Evolutionary Intelligence. 1(4):271-292.

Whiteson S., Stanley K. O., Miikkulainen R., 2004.

Automatic feature selection in neuroevolution. In:

Proceedings of the genetic and evolutionary

computation conference (GECCO), Seattle,

Washington, USA.

Zijlstra W., 2004. Assessment of spatio-temporal

parameters during unconstrained walking. European

journal of applied physiology

. 92(1-2):39-44.

ACCELEROMETER BASED GAIT ANALYSIS - Multi Variate Assessment of Fall Risk with FD-NEAT

143