Boy Sensor Networks to Evaluate Staning Balance: Interpreting Muscular Activities Base on Inertial Sensors Rohith Ramachanran Richarson, TX 75083 rxr057100@utallas.eu Gaurav Prahan Dept. of Computer Science Richarson, TX 75083 gaurav@utallas.eu Lakshmish Ramanna Richarson, TX 75083 lxr062000@utallas.eu Roozbeh Jafari Richarson, TX 75083 rjafari@utallas.eu Hassan Ghasemzaeh Richarson, TX 75083 hxg063000@utallas.eu Balakrishnan Prabhakaran Dept. of Computer Science Richarson, TX 75083 praba@utallas.eu ABSTRACT In this paper, we present a system that integrates inertial sensors an electromyogram (EMG) signals, which measures the muscular activities while performing motions. The objective of our stuy is to investigate the behaviour of the EMG signals to interpret the activity of staning balance. Quantitative parameters for balance are obtaine from an inertial sensor through a boysensor network. These parameters are further use to fin the prominent features in the EMG signal. The inertial sensor use in this system is an accelerometer. The implementation etails an effectiveness of using EMG signals are also provie. Categories an Subject Descriptors J.3 [LIFE AND MEDICAL SCIENCES]: Health General Terms Algorithms, Experimentation, Human Factors Keywors Balance, Inertial Sensors, EMG 1. INTRODUCTION Balance evaluation fins applications in rehabilitation, sports meicine, gait analysis an fall etection. Inertial sensor base systems have been in use for such applications. As in other physiological activities, staning involves the prominent use of muscles. EMG signals have been the most effective source for measuring muscle activity. In this paper, we investigate how parameters for balance obtaine from inertial sensors correlate with that of measurements from EMG signals. We obtain the balance parameters mentione in [1] from experiments conucte on ifferent normal human subjects an classify each of the parameters as low, meium an high. We fin out if features measure from EMG signals can also be classifie base on their correlation with the balance parameters. Linear Discriminant Analysis (LDA) is use for this purpose. Section 2 of this paper escribes how the balance parameters are obtaine from the accelerometer values. Section 3 provies the architecture of the system. The signal processing involve in the system to obtain this correlation is escribe in Section 4. Section 5 escribes how the experiments were conucte. Analysis an interpretation of the experimental results incluing LDA are provie in Section 6. 2. EVALUATION MODEL We use the balance evaluation moel escribe in [1] to erive performance metrics for staning balance. The system uses a single accelerometer place at the approximate height of the centre of mass on the subject s back. All three acceleration components are combine to buil a vector an the path trace by this vector is recore. The calculation of the coorinates of the path trace, as epicte in Figure 1, is as follows: if,, are accelerations in each irection, an is acceleration of gravity, the combine accelerations, is given by A + 2 2 2 = ax + ay a (1) z The irectional angles between A an X, Y, Z are represente by, a 1 a 1 y = cos, β = cos, ε = cos A A az A x 1 α (2) Permission to make igital or har copies of all or part of this work for personal or classroom use is grante without fee provie that copies are not mae or istribute for profit or commercial avantage an that copies bear this notice an the full citation on the first page. To copy otherwise, or republish, to post on servers or to reistribute to lists, requires prior specific permission an/or a fee. HealthNet'08, June 17, 2008, Breckenrige, CO Copyright 2008 ACM ISBN 978-1-60558-199-6/08/06...$5.00 From Figure 1, ε = z D cos (3)
where D is the combine coorinates in the three irections, x, y an z an represents the z coorinate of the en of A (istance to the groun from the sensor), which is assume to be a constant. Hence the coorinates of A at floor level (, ) can be expresse as: D cosα, D cosβ = (4) x y = then Mean Spee can be expresse as, Dt sm = (6) t Mean Raius is given by, m 1 enpo int 2 = = + n startpoint x n N 2 yn r (7) where N is the number of points in the trace path. Mean Frequency can be expresse as, f m Dt 2πr t = (8) m A/P an M/L isplacements are respectively given by, = max( ) min( ) (9) a / p xn x n Figure 1. Obtaining the projecte path. Figure aopte from [1]. = max( ) min( ). (10) m / l yn y n From this trace path, we can obtain the parameters use for evaluation. They are provie in Table 1. No. Table 1. Five Quantitative Features Quantitative Feature 1 Mean Spee 2 Mean Raius 3 Mean Frequency 4 Anterior/Posterior Displacement (A/P) 5 Meial/Lateral Displacement (M/L) Mean Spee, Mean Raius, Mean Frequency, A/P isplacement an M/L isplacement can be calculate. These parameters in combination give the measure of balance. The calculation of these parameters epicte in Figure 2 is as follows: If the Total Distance covere in time t, is given by, D t = enpo int 2 + 2 = ( + ) ( int + ) n startpo y y 1 x n n n xn 1 (5) Figure 2. Extraction of features from the projecte path Figure aopte from [1]. 3. SYSTEM ARCHITECTURE The system consists of two subsystems operating in parallel - the inertial sensor subsystem an EMG sensor subsystem. The inertial sensor subsystem is a boy-sensor network of two noes. One noe is on the boy of the subject an the other is connecte to a esktop PC. Accelerometer values are transmitte to the noe connecte to the PC by the noe on the boy. 3.1 Inertial Sensor Subsystem Our inertial sensor subsystem is a boy sensor network consisting two sensor noes (Moteiv Tmote Sky). The noe place on the
boy has custom-esigne sensor boar with a tri-axial 2g accelerometer as shown in Figure 3. It samples the sensor at 40Hz an sens ata over a wireless channel to a base station. The base station is another mote that relays the information to a PC via USB port. The sensor reaings are collecte an processe in MATLAB. stan on the platform. Figure 5 showss the platform along with an experimental subject wearing motion an EMG sensors. We integrate a HUSKY Digital Level to control the experiment an for coaching purposes (i.e. the subject must tilt the ball 20 egree in an anterior irection). The igital level inicates the amount of inclination when swaying on the platform. Figure 3. Mote with inertial sensors 3.2 EMG Sensor Subsystem We use several EMG sensors to measure the electric activity generate uring muscle contractions that occur while performing the motions. In the EMG suit we use (Delsys Myomonitor III), shown in Figure 4, the EMG signals are acquire using surface electroes attache at the skin surfaces. Each electroe measures the electric flow in the associate muscles. Figure 4. EMG system suit The electroes sample muscle signals in 1000Hz. The signals are amplifie an ban-pass filtere (20-450Hz) by the EMG suit. The ata are transferre to a PC for offline processing. 3.3 Balance Platform We use a balance ball as the platform for assessing the quality of staning balance. The platform is a Both Sies Up (BOSU) Balance Trainer which provies an unstable balance surface. This evice has two functional surfaces integrating ynamic balance with functional or sports specific training. It can be use platform sie up for push-up an seate exercises. We use this configuration which provies an unstable surface when subjects Figure 5. Balance platform an experimental subject wearing motion an EMG sensors 4. SIGNAL PROCESSING FOR FEATURE ANALYSIS Signal Processing involves extracting parameters from the accelerometer an EMG signals, classifying the accelerometer parameters an analysis using LDA. These operations are ivie into five stages as explaine below. Data Collection: Accelerometer values an EMG signals are continuously recore for every trial for uration of 4 secons. The sampling rates of the accelerometer an EMG signals are ifferent. Data from accelerometer is sample at 40Hz an that from the EMG sensors at 1000Hz. Parameter Extraction: Five quantitative features are measure using accelerometer ata as escribe in Table 1 of Section 2. Quantization: For each quantitative feature obtaine from the accelerometer values, the ata obtaine is ivie into three classes- low, meium an high.. The set of values of a particular feature that is greater than the sum of the mean of the feature an stanar eviation is categorize as high. The set of values less than the ifference between the mean an stanar eviation is categorize as low an the rest of the values as meium. Feature Extraction on EMG: To interpret the behavior of the EMG signals epening on the classes efine from accelerometer values, we nee to have exhaustive set of EMG features. An exhaustive set of statistical features are extracte from each EMG signal such as Signal Energy, Maximum Peak, Number of Peaks, Average Peak Value, an Average Peak Rate.
Feature Analysis: Significant features for EMG signals are extracte using Linear Discriminant Analysis (LDA). Given the quantitative metrics measure from the accelerometer, the purpose of feature analysis is to fin out if the EMG signals are representative of the quantitative features for balance evaluation. Figure 6. Signal processing flow EMG electroes on lower leg muscles. The EMG sensors were place on Right-Front leg (Tibialis Anterior muscle), Right-Back leg (Gastrocnemius muscle), Left-Front leg (Tibialis Anterior muscle), an Left-Back leg (Gastrocnemius muscle). The Delsys Trigger Moule enable the EMG subsystem to work synchronously with accelerometer. MATLAB behave as a main controller that sens a trigger to EMG an accelerometer to start acquisitions through the trigger moule (for EMG) an USB (for accelerometer). The process of ata collection was controlle an manage using our MATLAB tool. The EMG signals are obtaine synchronously with the accelerometer signals. The ata, however, are separately processe for the EMG an accelerometer. The accelerometer an EMG ata was recore for 4 secons for nine test conitions per subject. The test conitions are given in Table 2. Two trials for each conitionn were conucte for every subject. The angle of the tilt was measure from the level mounte on the balance platform. 5. EXPERIMENTAL PROCEDURE Experiments were conucte on five male subjects age between 25 an 32 an height between 1.65m an 1.8mm with no previous history of isorers. Subjects with correcte vision wore their glasses. Normal footwear was use for all subjects. Table 2. Test conitions Label Description Quiet staning with tilt limite to lesss than 10 egree S00 on either sie Tilt the ball to the left. Limit the angle of tilt between L10 10 egree an 20 egree Tilt the ball to the left. Limit the angle of tilt to at L20 least 20 egree Tilt the ball to the right. Limit the angle of tilt R10 between 10 egree an 20 egree Tilt the ball to the right. Limit the angle of tilt to at R20 least 20 egree Tilt the ball forwar. Limit the angle of tilt between F10 10 egree an 20 egree Tilt the ball forwar. Limit the angle of tilt to at least F20 20 egree Tilt the ball backwars. Limit the angle of tilt B10 between 10 egree an 20 egree Tilt the ball backwars. Limit the angle of tilt to at B20 least 20 egree A sensor noe with a tri-axial accelerometer was attache to a belt which was worn aroun the waist of the subject. The belt was worn such that the sensor noe was positione on the lower back of the subject. This noe communicate with another noe connecte to the USB port of a esktop computer. A MATLAB tool was evelope to rea the ata from mote connecte to the USB an process it. Although a number of muscles can be potentially active uring an action, currently, we constrain our system in using only four Figure 7. M/L isplacement measure from acceleration ata an quantize into three classes For every trial, the projection of the centre of mass (COM) on the groun was obtaine using the expressions we outline earlier. From the projections, five quantitative features were extracte. These features, liste in Table 1, are the features escribe in [2]. The calculation of these features from the projecte COM is shown in Figure 2. For each feature, the ata obtaine is ivie into three regions as we escribe in Section 4. EMG ata was recore for each trial. For each trial four channels of EMG ata are obtaine with each channel corresponing to a particular muscle. The ata obtaine from each channel was passe through a low-pass filter with a cut-off of 35Hz. From this filtere ata, a set of statistical features were extracte as was escriber earlier in Section 4. 6. EXPERIMENTAL RESULTS In this section, we present our resultss on evaluation of staning balance using the performance metrics. The accelerometer ata was obtaine for ninety trials across five subjects as escribe previously. The three imensional acceleration ata was use to fin projection of the center of mass on the plane. The five acceleration performance parameters were calculate base on the methos state earlier. For each parameter, the ninety trials were mappe into three classes representing quality of observe action in terms of that given parameter. We subjectively quantize every trial into quality level low, meium an high. For example,
with respect to the value of A/P isplacement, measure for each trial, we assign a class label base on its magnitue. This process is one for every accelerometer parameter obtaine in each trial. Each EMG feature set is given the same quality label as its corresponing accelerometer signal. The label assignment is accomplishe because the accelerometer an EMG signals are available for the same uration of the trial. Figure 7 shows a sample istribution of performance parameters across ifferent trials. The values were obtaine for M/L isplacement an were sorte for quantization. The statistical approach explaine before was use to fin threshols on each metric. Table 2. Significant EMG features escribing ifferent performance metrics Performance Metric Low Spee Meium Spee High Spee Low Raius Meium Raius High Raius Low Frequency Meium Frequency High Frequency Low A/P Meium A/P High A/P Low M/L Meium M/L High M/L Significant Features Maximum Amplitue (EMG2) Number of Peaks (EMG3) Maximum Amplitue (EMG2) Number of Peaks (EMG4) Maximum Amplitue (EMG2) Number of Peaks (EMG4) Energy (EMG2) Maximum Amplitue (EMG1) Maximum Amplitue (EMG2) Average Peak Rate (EMG4) Maximum Amplitue (EMG2) The next step in our system is to make EMG signals representative of performance parameters for balance evaluation. To achieve this, we etermine those features from EMG signals that are prominent for each class. We use 50% of the input trials (training set) to fin significant features for EMG an remaining trials (test set) for evaluation of the system. Each EMG trial consists of four signals corresponing to the four muscles. We extracte five features (Signal Energy, Maximum Peak, Number of Peaks, Average Peak Value, an Average Peak Rate) for each EMG signal. These features form a 20 imensional space which represent some properties of muscle activities uring the performe action. The obtaine features are fe to our feature analysis box (shown in Figure 6) where only the most prominent feature is selecte. The feature analysis was performe for each performance parameter. LDA is then use to select the most prominent feature from the subset. The list of prominent features is liste in Table 3. To get insight into the effectiveness of the acquire EMG features, we use k-nn (k-nearest Neighbor) classifier ue to its simplicity an scalability. For each accelerometer parameter class, the corresponing significant feature (extracte from Table 3) was extracte. This feature was use in a binary classifier to ifferentiate between a certain quality levels from the rest. For example, to evaluate how accurate EMG sensors represent performance metric Low A/P, corresponing prominent feature (Number of Peaks by EMG2) was extracte an fe to the k-nn classifier to istinguish between Low A/P an other two levels of A/P isplacement (Meium A/P an High A/P). The outcome of the classification for three values of k is illustrate in Figure 8 for a ranom set of classes. 100.0% 90.0% 80.0% 70.0% 60.0% 50.0% 40.0% 30.0% 20.0% 10.0% 0.0% Low Spee High Spee Meium Raius Low High Frequency Frequency Meium A/P k=1 k=3 k=5 Low M/L High M/L Figure 8. Classification accuracy for staning balance with respect to only EMG signals 7. RELATED WORK Human performance in terms of quality of balance control system has been stuie from ifferent views each taking into account a certain moel an its own evaluation metrics. Cybulski et al. [2] in their stuy of staning performance of paraplegia affecte subjects, euce an use statistical parameters from a center-offorce monitoring platform. Few authors have use accelerometer to measure the parameters use in [2] an stuy balance an control [3], [4] an [5]. Kamen et al. [3] use two uni-axial accelerometers, one each on the forehea an back. Mayagoitia et al. [1] use a single tri-axial accelerometer place on the back at approximate height of the center of mass to evaluate staning balance. Maithe et al. [6] use this evaluation moel to classify basic aily movements. Other authors ([7] an [8]) have concentrate on proviing auio an/or visual feeback on the moel parameters presente in [1] to improve balance. Winter et al. [9] present a kinematic moel of upper boy balance where EMG sensors were obtaine to reinforce the conclusions from the moment of force analyses. A stuy on comparison of EMG an kinetic parameters uring balance responses in chilren was presente by Sunermier et al. [10]. Accoring to their results, the corresponence of muscle activity with measurements of centerof-pressure confirms that muscle activities contribute to the balance. In this paper, however, we investigate methos of learning from inertial sensors to interpret EMG signals for staning balance. To the best of our knowlege, this has not been stuie before. 8. CONCLUSION AND FUTURE WORK We introuce a physiological monitoring system that collects acceleration an muscle activity signals an performs analysis on k=1 k=5
those signals uring staning balance action. The system quantifies performance in terms of five metrics which can be irectly measure from accelerometer ata. For the EMG signals, however, the quality of performe action is represente using a set of prominent features obtaine after processing the EMG signals in conjunction with the accelerometer parameters. To provie a complete evaluation of the system, we plan to investigate methos of integrating a gol stanar balance system with our experiments. We are also working on the eployment of our ata processing techniques in real-time. 9. REFERENCES 1. Mayagoitia, R.E., et al., Staning balance evaluation using a triaxial accelerometer. Gait an Posture, 2002. 16: p. 55-59. 2. Cybulski, G. an R. Jaeger, Staning performance of persons with paraplegia. Arch Phys Me Rehabil, 1986 2(67): p. 103-8. 3. Kamen, G., et al., An Accelerometry-Base System for the Assessment of Balance an Postural Sway. Gerontology, 1998. 44(1): p. 40-45. 4. Moe-Nilssen, R., A new metho for evaluating motor control in gait uner real-life environmental conitions. Part 1: The instrument. Clinical Biomechanics, 1998. 13(4-5): p. 320-327. 5. Rolf, M.-N., Test-retest reliability of trunk accelerometry uring staning an walking. Archives of physical meicine an rehabilitation, 1998. 79(11): p. 1377-1385. 6. Mathie, M., et al., Classification of basic aily movements using a triaxial accelerometer. Meical an Biological Engineering an Computing, 2004. 42(5): p. 679-687. 7. Chiari, L., et al., Auio-biofeeback for balance improvement: an accelerometry-base system. Biomeical Engineering, IEEE Transactions on, 2005. 52(12): p. 2108-2111. 8. Dozza, M., L. Chiari, an F. Horak, Auio-biofeeback improves balance in patients with bilateral vestibular loss. Arch Phys Me Rehabil, 2005. 86(7): p. 1401-3. 9. Winter, D., et al., An integrate EMG/biomechanical moel of upper boy balance an posture uring human gait. Prog Brain Res, 1993. 97: p. 359-67. 10. Sunermier, L., et al., The evelopment of balance control in chilren : comparisons of EMG an kinetic variables an chronological an evelopmental groupings. Experimental Brain Research, 2001. 136(3): p. 340-350.