The 7th Asian Conference on Computer Aided Surgery (ACCAS 2011)

Improving of sEMG Signal Feature Combination for Gait Phase Detection

Jirapong Manit King Mongkut's University of Technology, Thonburi 52432302@study.fibo.kmutt.ac.th 1. Introduction
Surface electromyography (sEMG) signal is widely used in many studies as the intelligent prosthetic control input. While using the raw sEMG signal is difficult to classify subject movements or activities, various type of domains extracted from raw signal such as time domain, frequency domain and time-scale domain were introduced to make the sEMG pattern classification more reliable [1, 2]. But the only one domain feature able to show limited signal information [3]. {Explain some more} In this study, we present the combination of multiple domain feature set for improving gait phase recognition performance which is used in the variable damper above-knee prostheses []. {Explain some more}

Prakarnkiat Youngkong King Mongkut's University of Technology, Thonburi youngkong@fibo.kmutt.ac.th

domain features include Mean Absolute Value (MAV), Variance (VAR), Waveform Length (WL), log-Detector (logDet) and Willison Amplitude (wAmp). And, frequency domain features consist of Mean Frequency (MNF) and Median Frequency (MDF). Data from each channel was first segmented into a 256-sample window, extracted to seven features from the window data. Next, the window was shifted by 64 samples until reaching the end of data. The combinations of two different domain features were made by selectively choosing one feature from time domain and another one from frequency domain. Therefore, a total number of features for testing is fifteen. C. Dimensional Reduction This section is very important for sEMG pattern classification to obtain an optimal computational time. Hence, the dimensional reduction technique helps making the input data size smaller. Principle component analysis (PCA) was proved to be as a good dimensional reduction technique for applying with sEMG signal data [ ]. It is a simple linear projection technique that is fast and easy to implement. So, it was implemented in this study. At the end of the part, the dimension of the input feature was reduced from three to two. 2.4 Pattern Classification Support Vector Machine (SVM) had been shown to be the best classifier. Fast computational time and ability to handle with non-linearity of SVM are advantages which are suitable for sEMG signal classification [ ]. Radial basis function (RBF) was used as the kernel function in SVM in this study. The feature with reduced dimension was divided into two sets, equally. The first set was used for training the classifier model and the second set was used for testing. Output of the classifier is the number that associates to the gait phase.

2. Methods
A. Data Acquisition The data was collected with sampling rate of 2000 Hz from three muscles; Rectus femoris, Vastus medialis and Biceps femoris, on both legs. ZeroWire electrodes were attached on a single leg as shown in figure 1. Five healthy males subjects between 22 to 34 years old were instructed to walk in a comfortable pace on a 10 meters length hard floor repeatedly 8 times. Three force plates were placed on the floor, and used to identify the interval of swing phase and stance phase. All collected data was labeled its class number manually; 0 for stance phase and 1 for swing phase.

3. Results
Accuracy is used to show the performance, e.g. the best feature set shall give the highest accuracy. Table 1 to 4 display accuracies in percent of gait phase detection for each feature. The average values of accuracy in the time domain features and in the combination between time domain and frequency domain are 89.97% and 96.15%, respectively. The result emphasizes that the combination between two domains feature give higher detection accuracy. give detection accuracy higher than a single time domain feature.

Figure 1. sEMG electrode location : (1) Rectus femoris (2) Vastus medialis and (3) Biceps femoris B. Feature Extraction Two kinds of feature domain, time and frequency domain, were applied to raw sEMG signal for creating single and multiple domain feature combination. Time

The 7th Asian Conference on Computer Aided Surgery (ACCAS 2011)

Table 1 Detection accuracy of time domain features. Feature % Acc Table 2 Detection accuracy of frequency domain features. Feature % Acc Table 3 Detection accuracy of combination between time domain features and Median Frequency. Feature % Acc Table 4 Detection accuracy of combination between time domain features and Mean Frequency. Feature % Acc MAV+ VAR+ WL+ wAmp+ logDet MNF MNF MNF MNF +MNF MAV+ VAR+ WL+ wAmp+ MDF MDF MDF MDF logDet +MDF MDF MND MAV VAR WL wAmp logDet

4. Conclusions
As shown in the result, the combinations between time domain feature and frequency domain feature improve the accuracy of gait phase classification and the best combination feature is {wait'n for the result} which shall be applied in the variable damper above-knee prostheses.

Acknowledgements
This research was financially supported by National Science and Technology Development Agency (NSTDA), the Higher Education Research Promotion and National Research University Project of Thailand, Office of the Higher Education Commission and King Mongkuts University of Technology Thonburi, Thailand, and medically supported by Sirindhorn National Medical Rehabilitation Centre (SNMRC).

References
[1] Dennis T, He H, Tood A K. Journal of NueroEngineering and Rehabilitation, 2010, 7:21 [2] Zeeshan O K, Zhen G X, Carlo Menon. BioMedical Engineering Online, 2010, 9:41 [3] Mahammadreza A O, Housheng H. Biomedical Signal Processing and Control 2,- 2007: 275-294 [4] [5] [6] [7] [8]