A Study on the Activation of Femoral Prostheses:

Focused on the Development of a Decision Tree based Gait Phase

Identification Algorithm

Sun-Jong Na

, Jin-Woo Shin, Su-Hong Eom

and Eung-Hyuk Lee

Department of Electronic Engineering, Korea Polytechnic University, Siheung, Gyeonggi-do, Korea

Keywords: Prothesis, Knee Angle, Decision Tree, Random Forest.

Abstract: This paper aims to classify the phase of gait for passive transfemoral prostheses as a preliminary study for the

development of a knee flexion angle control device in prosthetics by attaching it to the knee joint in order to

produce a walk trajectory like a normal person, while walking on a flat. However, it is not possible to

determine a gait stage according to the inflection point of a knee, since there are few angular changes in the

knee joint in the form of a seat that will support the body. Thus, in previous studies, algorithms were developed

to distinguish between three stages of the stance in the swing phase using a decision tree learning method.

However, the decision-making tree is prone to overfitting. This can be a high level of accuracy for training

data, but it is difficult to generalize when verification data or new data are entered. Therefore, in this paper,

we want to develop an algorithm for preventing the overfitting step-by-step using two different methods.

1 INTRODUCTION

Based on the 2017 report of the WHO (World Health

Organization), an estimated 30 million people with

lower limb amputees are expected to double by 2050

(Ziegler-Graham, 2008). Based on this basis, the

research on prostheses that help to compensate lower

limb amputees is being studied in a variety of ways

for the convenience of the disabled.

The prostheses are divided into passive and active

types according to the way they operate. Passive type

prostheses are able to walk on a level surface through

storing and using the force of the wearer without

power, but it is difficult to implement power for

activities such as stair climbing and running (Yoshida

et al., 2015);(Inoue et al., 2016). However, active type

prostheses are able to make their own strength by

using actuators such as cylinders or motors, so they

can perform various motions. However, they are

expensive (Andrés et al., 2016; Keles et al., 2017).

The active type prostheses can create gait

trajectories similar to normal people through actuator

control even on level walking. However, the passive

type prostheses work only as a supporting stand of the

https://orcid.org/0000-0002-0601-9058

https://orcid.org/0000-0001-8493-1432

body for the next step because they do not have the

power to create gait trajectories like normal people.

Thus, the objective of this study is to classify gait

phases in passive type prostheses as a pilot study for

the development of devices that adjust the flexion

angle of knee joints according to the gait phase by

attaching them to the knee joint of the passive type

prostheses.

Studies to distinguish gait phases have now been

conducted in two different ways. The first method is

to use ground reaction force to separate the point at

which the feet do not touch the ground (Shaikh et al.,

2015). The second method is to distinguish gait

phases according to changes in the knee angle

(Karasawa et al., 2013). Estimates of the gait phase

according to changes in the knee angle are divided by

using the maximum flexion of the knee angle and the

inflection point in the extension trajectory (Lim et al.,

2016).

The passive prostheses, however, have little

changes in the knee angle of the knee joint, and the

gait trajectory varies depending on the length of the

affected area. Also, it is not easy to apply the

inflection point based estimation method of gait

Na, S., Shin, J., Eom, S. and Lee, E.

A Study on the Activation of Femoral Prostheses: Focused on the Development of a Decision Tree based Gait Phase Identiﬁcation Algorithm.

DOI: 10.5220/0007950707750780

In Proceedings of the 16th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2019), pages 775-780

ISBN: 978-989-758-380-3

775

phases as a result of the change in the knee angle of

these passive prosthesis users because the new

practice of gait habits after amputation varies from

person to person.

Therefore, in the previous study, an algorithm was

developed to classify the swing phase through

dividing the stance phase into three stages and

separate the gait phase by using the decision tree

learning method in the form of 'IF ~ THEN' to

distinguish the gait phase of the passive type

prosthesis users (Na et al., 2019).

However, the decision tree learning method

becomes more complex as models become more

complex and may represent overfitting. Although this

may be a high accuracy for training data, it is a

disadvantage that it is difficult to generalize as

verification data or new data are entered.

There are two ways to prevent this overfitting.

The first method is to simplify the decision tree and

control the depth of the tree. The second method is to

use a random forest model which is one of the

machine learning ensemble techniques that results in

more predictive ability and less overfitting in training

data than a single decision tree by categorizing it

through means of multiple independent decision

trees(Rezaei et al., 2018).

In this study, two methods are used to prevent the

overfitting. The first method is to simplify decision

trees, and the second method is the application of a

random forest model. Then, this study compares the

differences between decision trees based on previous

studies.

2 METHOD

In this study, changes in the knee angle are measured

during walking of passive type prosthesis wearers and

identify the limit of the classification of the gait phase

in the stance phase. To solve this issue, the changes

in the hip angles obtained using the acceleration of an

inertia sensor attached to the surface of the prosthetic

adapter and a three-axis gyro are divided into three

stages based on the ground reaction force in order to

produce training data entered as labels. The training

data develop a convergence algorithm by adding both

the decision tree that divides the stance phase into

three stages using the decision tree learning method

and an algorithm that identifies the swing phase using

the inflection point of the knee joint. Also, this study

applies algorithms using the random forest technique

to compensate for the shortcomings of the decision

tree method and compare them with previous research

methods.

2.1 Characteristics of Knee Angle

Changes in Passive Type Prosthesis

Wearers

Figure 1: Gait phases according to changes in the knee

angles, separated by a pressure sensor, of a passive type

prosthesis wearer.

As shown in Figure 1, the knee joint a typical passive

type prosthesis wearer will fold the knee from the

point (④) at which one steps off the ground. Then,

the knee, which had been folded from the point where

it passed through the intermediate swing phase (⑤),

will be stretched out again for the next step.

Therefore, the swing phase (④~⑤) can be identified

by the change of the knee angles.

In the stance phase, however, it is difficult to

distinguish between the initial landing on the ground

(①), the intermediate stance phase (②), and the final

stance phase (③) as there is little change in the knee

angle because the prosthesis acts as a stand to support

the wearer.

In order for the wearers to avoid feeling awkward

in their gaits, it is necessary to create a gait trajectory

of about 18 degrees like the gait by a normal person

at the intermediate stance phase (②). Thus, it was

possible to identify the stance phase as three different

stages using the measured values at each FSR section

through the FSR attached to the toe end and heel and

the ground reaction force. It allows for the creation of

gait trajectories through the device at the point of the

intermediate stance phase (②).

ICINCO 2019 - 16th International Conference on Informatics in Control, Automation and Robotics

776

2.2 Random Forest Method

The random forest technique is one of the ensemble

techniques among the various machine learning

techniques that increases the accuracy of

classification by aggregating the results from several

classification models. It is possible to maintain biased

data and to reduce overfitting by decreasing variance

because the ensemble techniques apply the average

value of the results of multiple classification models.

Figure 2: Structure of the random forest algorithm.

A bootstrap sample must be created to use the

random forest technique. The bootstrap sample can be

created by dividing the original dataset by attributes

and then randomly selecting the attributes for the

original dataset size. As the attributes are extracted

using an iterative extraction method, they can overlap

within a bootstrap and cause missing attributes. Thus,

it can reliably output classification values even when

new data is entered for classification because the

decision tree generated by each bootstrap is

independent of each other.

The equation for determining the probability that

one attribute will be excluded from the sample due to

iterative extraction from one bootstrap sample is as

shown in (1).

lim

→

1−





=



= 0.3678

(1)

The random forest method evaluates the accuracy

of the model using an OOB (Out-Of-Bag) error,

which collects 36.78% of excluded samples and

evaluates performance with verification data

(Breiman, 2001).

3 EXPERIMENT AND RESULTS

3.1 Experimental Procedures

In this study, we intend to simplify decision trees as

the first way of reducing overfitting. The previous

research method was a decision tree that classifies the

total of four categories, three stages in the stance

phase and one swing phase. In this study, however,

we intend to develop a convergence algorithm that

divides only the three stages in the stance phase by

using the decision tree learning method for

simplifying the model, and using the inflection point

of the knee joint in order to identify the swing phase.

In the second option, we intend to verify that this

method reduces overfitting but increases accuracy

over the previous research method by using a random

forest technique, a set of independent trees

Figure 3: Passive type prosthesis adaptor and sensor

attachment locations, experimental setting image and

coordinate system used in this experiment.

In this study, an inertial sensor is described in

relation to the coordinate system shown in Figure 3

by rotating it 90° clockwise by x-axis. A total of five

men in their 20s, two 70kg and three 80kg, walked by

wearing a mechanical prosthesis adapter with a 5

steps at 30cm intervals, and a total 500 steps from 20

times on the ground. The characteristics of the dataset

for training consisted of knee angles, three axis of

acceleration generated during walking, three axis of

gyro, and three axis angles of the hip joint. The gait

phases for each gait were labelled according to the

gait phases using the ground reaction force and FSR.

In the sampling of the obtained dataset, 80% of the

total data was used as training data and 20% of the

data was used as verification data. The inertial sensor

was attached to the point Ⓐ for measuring

acceleration, angular velocity, and angle data in the

hip joint while walking on the surface of the adapter.

The sensor Ⓑ was a variable resistance for measuring

knee angles by converting variations in the resistance

value of walking into angles. The pressure sensors,

Ⓒ, were used to separate the stance phase into three

A Study on the Activation of Femoral Prostheses: Focused on the Development of a Decision Tree based Gait Phase Identiﬁcation Algorithm

777

stages. The specifications for each sensor are shown

in Table 1.

Table 1: Specifications for Each Sensor.

Name Spesification

Inertial Sensor

NGIMU (x-io)

Communication speed: 50 Hz

Communication method: Wifi

Variable Resistance Max. 10 kΩ

Pressure Sensor FSR402 (10N Sensitivity)

3.2 Results

As the depth of the decision tree was adjusted with

the pre-pruning of the decision tree generated by the

previous research method and the convergence

algorithm developed in this study, the results were

obtained as follows.

Table 2: Difference in Accuracy according to the Depth of

Each Algorithm.

Model

Depth_

STANCE 3 STANCE 3 + SWING

Training Test Training Test

0 99.8% 98.4% 100% 99.2%

1 67.9% 69.8% 61.9% 62.5%

2 92.5% 93.5% 79.3 79.3

3 94% 94.4% 92.8 92.9

4 94.3% 95% 94.1 94

5 95.2% 95.4% 95.1 94.4

As the pre-pruning was not applied, the previous

research models represented overfitting. It was

possible to verify that the accuracy of the tree

increased as it became more complex. Both the

developed convergence algorithm and the decision

tree generated by the previous research method with

the same 94% verification data accuracy are as

follows.

Figure 4: Developed convergence algorithm (A) and the

decision tree (B) with the same 94% verification data

accuracy.

In the equation for identifying the swing phase

presented in Figure 4 (A), CKneeAngle is the knee

angle currently being measured, and KneeAnglediff

represents increases or decreases in the knee angle.

SWThreshold is the boundary value for identifying

the swing phase, in this study, the accuracy of 96.87%

was set at about 0.4° considering the amount of

variation in the knee angle generated by the swing

phase. The depths of the decision tree of (A) and (B)

represent 3 and 4 respectively with the same

accuracy, but the decision tree in converged (A) was

found to be simpler than (B).

Figure 5: Changes in OOB error values according to the

number of bootstrap samples.

As shown in Figure 5, 50 bootstrap samples were

98.6% and 50 or more were 98.7%, and the accuracy

of the bootstrap was not significantly different even if

the number of bootstrap increases.

ICINCO 2019 - 16th International Conference on Informatics in Control, Automation and Robotics

778

Figure 6: Feature importance of the single decision tree (A)

and random forest (B).

The feature importance is an indicator of which

properties are used most importantly as a decision

tree is generated. It is determined as a value between

0 and 1, which means that 0 is not used at all, and 1

has all the information for classification. Figure 6 (A)

represents the feature importance of the decision tree

shown in Figure 4 (A) and uses only three features:

Pitch, Roll, and Gyro Z. Figure 6 (B), however, (B) is

the feature importance graph of the random forest that

uses all the attributes in training data. Thus, it reduces

the overfitting and shows easy generalization

compared to (A).

4 CONCLUSIONS

In this study, as the first step in developing a device

for the activation of passive prostheses, the objective

was to identify the gait phase in the walking of

passive prosthesis wearers. Two methods were used

to reduce overfitting. First, a decision tree that

identifies the three stages of the stance phase and a

convergence algorithm that calculates threshold

values for determining the swing phase from the

changes in knee angles were developed. It showed

that it becomes a simple model even though it has the

same accuracy as the previous research method.

Second, it was verified that the accuracy was

improved to 98.6% while reducing the risk of

overfitting in the decision tree through applying the

random forest method.

Future plans will be to develop a machine running

algorithm to identify the gait environment on a level,

slope, and stairs and to automatically change the gait

mode for each environment.

ACKNOWLEDGEMENTS

This research was supported by the Basic Science

Research Program through the National Research

Foundation of Korea (NRF) funded by the Ministry

of Education (NRF2017R1A2B2006958).

This research was supported by the Bio &

Medical Technology Development Program of the

NRF funded by the Korean government, MSIP (NRF-

2017M3A9E2063260).

This research was supported by the Technology

Innovation Program (NO.10082455, Service

Development of Spo-Edutainment School Indoor

Thema park) funded By the Ministry of Trade,

Industry & Energy (MOTIE, Korea).

REFERENCES

K. Ziegler-Graham. (2008). Estimating the prevalence of

limb loss in the United States: 2005 to 2050, Arch.

Phys. Med. Rehabil., vol. 89, no. 3, pages 422 - 429.

Shun Yoshida, Takahiro Wada, Koh Inoue. (2015). A

passive transfemoral prosthesis with movable ankle for

stair ascent, IEEE International Conference on

Rehabilitation Robotics (ICORR), pages 7 - 12.

Koh Inoue, Tomohiro Tanaka, Takahiro Wada, Shin'ichi

Tachiwana. (2016). Development of a passive knee

mechanism that realizes level walk and stair ascent

functions for transfemoral prosthesis, IEEE

International Conference on Biomedical Robotics and

Biomechatronics (BioRob), pages 522 – 527.

Guzhñay Cordero Andrés, Calle Arévalo Luis, Zambrano

Abad Julio. (2015). Walking cycle control for an active

ankle prosthesis with one degree of freedom monitored

from a personal computer, International Conference of

the IEEE Engineering in Medicine and Biology Society

(EMBC), pages 3651 – 3654.

Ahmet Doğukan Keles, Can Yücesoy. (2017).

Development of Artificial Neural Network Based

Active Ankle Prosthesis Algorithm Using Gait

Analysis Data, National Biomedical Engineering

Meeting (BIYOMUT), pages 1 – 4.

Muhammad Faraz Shaikh, Zoran Salcic, Kevin Wang.

(2015). Analysis and selection of the Force Sensitive

Resistors for gait characterization, International

Conference on Automation, Robotics and Applications

(ICARA), pages 370 - 375.

Yuta Karasawa and Yuta Teruyama. (2013). A Trial of

Making Reference Gait Data for Simple Gait

Evaluation System with Wireless Inertial Sensors,

A Study on the Activation of Femoral Prostheses: Focused on the Development of a Decision Tree based Gait Phase Identiﬁcation Algorithm

779

International Conference of the IEEE EMBS, pages

3427-3430.

YoungGwang lim, SeongBong Yoo, SuHong Eom,

EungHyuk Lee. (2016). A Study On Estimation of Gait

Phase Using Peak Point Detector and Finite State

Machine, ICS´, pages 109 - 111.

Sun-Jong Na, Su-Hong Eom, Chol-U Lee, Mun-Seok Jang,

Eung-Hyuk Lee. (2019). A study on the Application of

Decision Tree Algorithm to Differentiate Gait Phases in

the Users of Transfemoral Prostheses, International

Conference on Electronics, Information and

Communication (ICEIC), pages 997 - 1001.

A. Rezaei, A. Ejupi, M. Gholami, A. Ferrone, C. Menon.

(2018). Preliminary Investigation of Textile-Based

Strain Sensors for the Detection of Human Gait

Phases Using Machine Learning, IEEE International

Conference on Biomedical Robotics and

Biomechatronics (Biorob) pages 563 – 568.

L. Breiman, (2001). "Random Forests," Machine Learning,

vol. 45, pages 5 - 32.

ICINCO 2019 - 16th International Conference on Informatics in Control, Automation and Robotics

780