ECE-464 บรรยาย point

7/29/2019 ECE-464 point

1/16

A Study of Fall-Risk Affecting MotionParameters and Develop a Neuron Network-

Based Adaptive Fall-Risk AnalyzerSubmitted for Funding from

Assistive Technology Center/NECTEC

Assoc. Prof. Dr. Wattanapong Kurdthongmee

Piyadhida Kurdthongmee

Taofig Lumsup

Walailak University/Nakorn-Si-Thammarat

Presented to KMUTT Researchers on 27/06/2008

7/29/2019 ECE-464 point

2/16

Abstract

A high percentage of elderly people, the main focus group of this research, hasbeen continually increased in Thailand. This results from the successes of boththe advancements of medical knowledge and the high coverage of medicalservices which altogether make people healthier and live longer.

The current approach to monitor the fall-risk (to be specific for this research)

of elderly people (or at-risk patients) apart from being closely monitored bymedical staffs and their relatives relies on direct attaching medical sensors withtheir body. The common sensors in-use are an EKG (Electrocardiogram)sensor, a pulse sensor and a blood pressure sensor.

These sensors are used with an expectation that if some measuring parametersare changed, they might result in a severe fall and unconsciousness of elderly

people.

In order to function correctly, these sensors must be firmly attached to thebody of wearers by a well-trained medical staff. The sensors may causeunconfortable for wearers and make them lack of self-confidence to live bytheir own.

7/29/2019 ECE-464 point

3/16

Abstract (cont)

The aim of our research is to study the parameters which should be collectedor observed from the movement and motion of wearers.

In addition, the main aim of this research is to propose the algorithm forautomatic classification and making decision of fall-risk parameters.

The proposed algorithm could be later integrated with an embedded systemattaching to wearers for warning and emergency alerting purposes. The keysensors which are going to be used in the course of our research aregyroscopes and accelerometers.

These sensors are now fabricated with a MEMS (MicroElectroMechanicalSystem) technology with donimate points of low power consumption,

acceptable price and size and light in weight.

Above all, the exploitation of these sensors to this application domain wouldresult in more comfortable to wearers. Within the system, the Self OrganizingMap neuron network is going to be used to perform adaptive analyzer andclassification functions which make the system applicable to wearers withdifferent ages and motion conditions.

7/29/2019 ECE-464 point

4/16

Introduction

Falls: Serious or not? most common type of home accidents among elderly people and

major threat to their health and independence (Najafi, 2002).

change elderlys self-confidence and motivation,

affect elderlys ability to function independently

32 percent of a sample of community dwelling persons 75 years and older fellat least once a year.

Among them, 24% sustained serious injuries (Tinetti, 1988).

With the growing proportion of old people in the populations falls willbe one of the major problems.

Most cases of falls sustained by elderly people appear to result from the

cumulative effect of multiple specific disabilities. Among these, balanceand gait disorders play a key role (Tinetti, 1986).

Evaluating the risk of falling is important it enables the provision of adapted assistance and of taking preventive

measures with subjects deemed at risk of falling.

7/29/2019 ECE-464 point

5/16

Introduction

Evaluating the risk of falling: questionnaires with their associated problems of subjectivity and limited

accuracy in recall (Cummings, 1988).

clinical and functional assessment including posture and gait, independence indaily life, cognition, and vision (Tinetti, 1986).

evaluating the characteristics of postural transition (PT) and their correlation

with falling risk (Tinetti, 1988). Wavelet transform was used for dataextraction.

Gyroscopes and accelerometers have long been exploited for monitoring andwarning applications (Patent, 2008) without adaptability features.

Wireless body area network (WBAN) was proposed by (Javanov, 2005) tomonitor wearers activity with a minimal number of accelerometers. Fall risk

was analyzed on the remote server.

(Hwang, 2004) also proposed a novel algorithm and real-time ambulatorymonitoring system for fall detection in elderly people. The system comprisedof accelerometer to measure kinetic force, tilt sensor and gyroscope to estimate

body posture.

7/29/2019 ECE-464 point

6/16

Project Objectives

The aims of our research are to: Investigate the changes of an individuals motion parameters over a

period of time and, possible, age,

Experimentally proof that SOM can be used to learn an individualsmotion parameters and make the decision for an unsafe motion that

could be a fall risk, Study the parameters which should be collected or observed from the

movement and motion of an individual.

Propose the algorithm for automatic classification and making decisionof fall-risk parameters.

The algorithm could be later integrated with an embedded system

attaching to wearers for warning and emergency alerting purposes. The Self Organizing Map (SOM) neuron network is going to be

used to perform adaptive analyzer and classification functionswhich make the system applicable to wearers with different agesand motion conditions.

7/29/2019 ECE-464 point

7/16

Expected Outputs

We expect to present the following outputs:

An algorithm, or software class, for automatic classification and making

decision of fall-risk parameters which have the following features:adaptabilityand open an opportunity to accept other related parameters,

2 international journal publications, 2 national journal publications,

2 final year undergraduate student projects,

Development of 2 researchers in the related field,

7/29/2019 ECE-464 point

8/16

Why Adaptive?

Designing a computer application to whatever field implies solving a numberof problems, mostly deriving from the variability which typically characterizesinstances of the same real-world problem.Whenever the description of a

problem is dimensionally large, having one or more of its attributes out of thenormality range becomes almost inevitable. Real-world applicationstherefore usually have to deal with high-dimensional data, characterized by a

high degree of uncertainty. In response to this, real-world applications need tobe complex enough to be able to deal with large datasets, while also beingrobust enough to deal with data variability.: S. Cagnoni, E. Lutton, and G.Olague, Genetic and Evolutionary Computation for Image Processing and

Analysis

Adaptive behavior is a type ofbehaviorthat is used to adapt to another typeof behavior or situation. This is often characterized by a kind of behavior thatallows an individual to substitute an unconstructive or disruptive behavior tosomething more constructive. These behaviors are most often social or

personal behaviors. For example a constant could be re-focused on somethingthat creates or builds something. In other words the behavior can be adapted tosomething else.

7/29/2019 ECE-464 point

9/16

Self Organizing Map

A 2-D map is defined by klocations orkcells arranged as a 2-D lattice.

Each location contains an n-dimensional model vector which comes toresemble n-dimensional input (teaching) data during the unsupervisedlearning process, the self organization (Joutsiniemi, 1995).

As a result of SOM process, the distribution of the model vectors in the n-

dimensional space will approximate the probability distribution of the inputvectors.

The topographic organization of the map will also approximate the matrixordering relations in the input space. Thus, similar inputs project near eachother onto the map.

The map thus forms in the input space an elastic surface, whichapproximates the probability density function of the input samples.

At the beginning of the self-organization, all model vectors mi may haverandom values.

For each input vectorx(t), a model vectormc(t) with minimum Euclideandistance from the input is searched for: ||x(t) - mc(t)|| = mini{||x(t) - mc(t)||}.

7/29/2019 ECE-464 point

10/16

Self Organizing Map

The best matching model vector mc and the model vector mi in its

neighborhood are then modified toward the value of the input vector:

mi(t+1) = mi(t) + (t)( x(t)mi(t)).

The magnitude of the modification factor (t) decreases monotonically.

Also, the size of the neighborhood ofmc decreases at successive inputs. At

the beginning of self-organization, its neighbourhood on the map is wide,and at the end only the nearest neighbour ofmc(t) are modified.

7/29/2019 ECE-464 point

11/16

Success Stories

We have some success stories relating to SOM: 2 International papers in an image quantization application (hardware based

implementation): W. Kurdthongmee, A Novel Kohonen SOM-Based Image Compression Architecture

Suitable for Moderate Density FPGAs, Journal of Image and Vision Computing,Vol. 26, Issue 8, 1 August 2008, Pages 1094-1105.

W. Kurdthongmee, A Novel Hardware-

Oriented Kohonen SOM Image CompressionAlgorithm and Its FPGA Implementation, Accepted to Publish in Journal of SystemsArchitecture.

1 International paper in agricultural application, W. Kurdthongmee, Colour Classification of Rubberwood Boards for Fingerjoint

Manufacturing Using a SOM Neural Network and Image Processing, Accepted toPublish in Journal of Computers and Electronics in Agriculture.

1 paper in educational application, W. Kurdthongmee, Utilization of a self organizing map as a tool to study and

predict the success of engineering students at Walailak University, Accepted to

Publish in Walailak University Journal of Science and Technology.

1 international conference paper in medical application. W. Kurdthongmee and P. Kurdthongmee, An Exploitation of the Self-Organizing

Map for Human Motion Analysis, Submitted to BioDevices 2008 Conference.

7/29/2019 ECE-464 point

12/16

Feasibility Study

The feasibility study result wasdone and resulted in aninternational conference paper:W. Kurdthongmee and P.Kurdthongmee, An Exploitation

of the Self-Organizing Map forHuman Motion Analysis,Submitted to BioDevices 2008Conference.

A small data logger wasdeveloped to retrieve and store

motion data: dsPIC30F2010 witha pair of AT24C1024s.ADIS16350 was used to measureangular and linear accelerations.

7/29/2019 ECE-464 point

13/16

Feasibility Study

A snapshot of raw data and its

graphs are shown here:

There are some characteristics of

the motion parameters with respect

to the activities (walking, sitting,

walking upstairs, walkingdownstairs, sleep).

Raw Data from Accelerometers at Different Times & Activities

-300

-200

-100

0

100

200

300

400

500

3:43 PM 3:50 PM 3:57 PM 4:04 PM 4:12 PM 4:19 PM

Time Stamp

Accel X

Accel Y

Accel Z

Activity

Raw Data from Gyroscopes at Different Times & Activities

-1500

-1000

-500

0

500

1000

1500

3:43 PM 3:50 PM 3:57 PM 4:04 PM 4:12 PM 4:19 PM

Time Stamp

Gyro X

Gyro Y

Gyro Z

Activity

7/29/2019 ECE-464 point

14/16

Feasibility Study

The resulting map after training process with an individuals input data of 7

dimensions, which are the period of change, the angular accelerations inthe x, y and z axes and the linear accelerations in the x, y, and z axes.

After labelling process

Projection maps

Gx, Gy, Gz

Ax, Ay, Az,, Delta

7/29/2019 ECE-464 point

15/16

Feasibility Study

Consider the leave-one-out testingresults: performed by querying the trained SOM

with an unknown motion parameterswith respect to the SOM (the one whichwas taken out from the training dataset),

the results are presented in the middlecolumn of Table 1 (within group).

The rightmost column of Table 1: thecross-validating test.

Consider the probability of beingmatched to all activities after presentingthe unknown motion parameters to SOM(but known activities to us) in Table 2. Some activities share common motion

parameters: (walking, walking downstairs)

(sitting, walking downstairs),

(jocking, walking downstairs,

sitting).

7/29/2019 ECE-464 point

16/16

Feasibility Study

Conclusions: SOM could be trained with an individuals motion parameters which are the

period of change of a consecutive pair of parameters, the angular accelerationsin (x, y, z) and the linear accelerations in in (x, y, z), resulting in clusteringsimilar motion parameters together.

SOM could match between normal activities and the clusters of motion

parameters on the maps with as high as 73.45 percents of correctness. the matching between abnormal motion parameters that could be a fall risk still

needs some efforts to persue.

According to the experiment results, it can be concluded that different activitiesof individual have different motion parameters (the period of change is alsoincluded in these parameters).

SOM can successfully and correctly cluster these activities in relation to themotion parameters.

In order to classify the activities of a person with a high degree of correctness,SOM needs to be trained with the motion parameters of that person.

With positive experimental results, we expect that SOM can be utilized tomake the decision for an unsafe motion that could be a fall risk in an adaptiveway.