Proposal and Evaluation of the Gait Classification Method Using Arm Acceleration Data and Decision Tree

We have been developing a system for falling prevention to classify gait patterns based on stride length and foot clearance by arm accelerations. In this paper, we propose gait classification method using arm acceleration data and decision tree. Also, we evaluate whether decision tree using three-axis accelerations as feature quantities could classify three gait patterns. (Three gait patterns are "Normal", "High step" and "Long step".) The result showed that this method can classify three gait patterns of some subjects.


Introduction
Lately, Japan has a lot of falling accidents of elderly people.Falling accidents often lead to hospitalization on serious injuries.Hospitalization adversely affects exercise capacity and mental state.According to a survey, 86.2% of elderly people who have experienced falling accidents are afraid of falling, and 44% of them have restricted outdoor activities due to the fear of falling [1].From these reports, it can be said that falling accidents of elderly people cause not only damages but also serious effects for daily life after accidents.
It is known that tripping is the most frequent cause of falling [1][2].Due to decreased perceptual ability, elderly people cannot recognize that foot clearance and stride length are insufficient in gait [3].Therefore, we have been developing a system which urges elderly people to improve of gait for falling prevention by classifying gait patterns based on stride length and foot clearance using acceleration data from sensors of a P -104 smartphone [4].In this system, the smartphone is worn on an arm as a part at which it is easy to attach on a daily basis for users, and gait patterns are classified by arm acceleration data.In previous studies, it has been pointed out that movement of the arm is influenced by the gait in terms of dynamic stability [5].
It has been found that time waveforms of arm accelerations change by stride length and foot clearance.However, we also have found that it is difficult to classify gait patterns by only peak value and average value of time waveforms [4].In this paper, we propose the gait patterns classification method using machine learning and arm acceleration data as feature quantities.

Method
In proposed method, gait patterns are classified by decision tree with three-axis acceleration and composite direction acceleration data on an arm as feature quantities.As an advantage of the decision tree, it is known that if-then rules in a tree structure for classification are readable.In addition, some algorithms of the decision tree can use continuous variables.
Gait parameters are often measured in time periods.Because elderly people are unstable in their gait cycle, there is a possibility that astandard deviation may increase in measured data.Therefore, in the proposed method, one set of arm acceleration data is measured per step by detecting the timing at which the forearm crosses the body by a proximity sensor of the smartphone.We conducted the gait experiment on the treadmill by young people as basic examination of the proposed method.In this experiment, we instructed subjects (10 young people) three gait patterns ("Normal", "Long step ", "High step") on the treadmill.As for "Long step", we confirmed that stride lengths based on steps and gait speeds were the largest in three gait patterns during whole analysis.Also, as for "High step", we confirmed that foot clearance was rising higher than 15 cm by infrared sensors.Besides, we set gait speed which is easy to walk for each subject by treadmill.In this experiment, as feature quantities, three-axis arm acceleration data for 60 steps per each gait pattern were measured by the smartphone.In addition, composite direction acceleration data also were calculated from three-axis acceleration data as feature quantities.

Data Analysis
As a basic study of the proposed method, we evaluated whether decision trees using arm acceleration data as feature quantities could classify gait patterns.We used Weka and J48 algorithm for evaluation of classification performance.In the data set used for P -105 classification, arm acceleration data ("x-direction", "ydirection", "z-direction", and "composite direction") measured by the experiment were used as feature quantities, and the information of the gait patterns ("Normal", "Long step", or "High step") was added to each data as a correct answer.We calculated accuracy, precision, recall, and F-measure as evaluation index of classification performance for each subject on 5-fold cross-validation.We also analyzed if-then rules of a tree structure for each subject.

Results and discussions
It was found that gait patterns can be classified with accuracy more than 0.65 in 7 subjects (Table 1).This result shows the possibility that the proposed method is usable for gait classification.Comparing F-measure in each gait pattern, it was confirmed that F-measure of "Normal" was larger than other gait patterns in 6 subjects (Fig. 3).In addition, it was found that precision and recall of "Normal" was larger than other gait patterns too (Fig. 4).As the reason for these, it is assumed that "Normal" was not required an arm swing for dynamic stability, so difference of arm acceleration data appeared between "Normal" and other gait patterns.
Results of the analysis for if-then rules in the tree structures of each subject showed that threshold values of "y-direction" are larger than threshold values of "xdirection" and "z-direction" (Fig. 5).It is conceivable from these results that "y-acceleration" is dominant in "composite direction".
In addition, we examined the root node in the tree structure of each subject, and found that arm acceleration data of "x-direction" and "composite direction" are selected more frequently as the threshold variable for root node in the tree structure of high accuracy subjects (Table 1).In the decision tree, since root node is arranged so that decrease of entropy gets a maximum, root node is considered to be important for gait classification.Therefore, "x-direction" and "composite direction" which are selected more frequently as the threshold variable for root node, and "y-direction" which is dominant in "composite direction" are considered to be important in gait classification.

Conclusions
In this paper, we proposed a method to classify gait patterns by decision tree using arm acceleration data ("x-direction", "y-direction", "z-direction", and "composite direction") as feature quantities.Experimental results revealed that the proposed method could classify three gait patterns ("Normal", "Long step", and "High step") of some subjects.
Moreover, from tree structure of decision tree, we found that arm acceleration data of "x-direction" and "composite direction" were selected more frequently as the threshold variable for root node in the tree structure of high accuracy subjects.In addition, we found that "yacceleration" is dominant in "composite direction".From these results, tree structure of decision tree showed that "x-direction", "composite direction" and "y-direction" are considered to be usable in gait classification.
In order to improve classification performance, it is considered that analysis of "two-axis" acceleration data is needed in further studies.From now on, we will examine the gait classification method using "two-axis" composite direction acceleration data from "x-direction" and "y-direction" which are considered to be important in gait classification.

Table 1 .
Accuracy and threshold variable of root node