The use of accelerometry to detect heel contact events for use as a sensor in FES assisted walking

Transcription

1 Medical Engineering & Physics 25 (2003) Technical note The use of accelerometry to detect heel contact events for use as a sensor in FES assisted walking Avril Mansfield, Gerard M. Lyons Biomedical Electronics Laboratory, Department of Electronic and Computer Engineering, University of Limerick, Limerick, Ireland Received 1 November 2002; received in revised form 31 March 2003; accepted 20 June 2003 Abstract Current sensors for the control of functional electrical stimulation (FES) assisted walking in hemiplegic individuals are not wholly satisfactory, as they are either not implantable or ineffectual in the detection of heel contact events. This study describes the use of an accelerometer placed on the trunk to detect heel contact events of both legs based on the examination of the anterior posterior horizontal acceleration signal. Four subjects wore an accelerometer over their lumbar spine. Footswitches placed on the sole of one foot recorded the heel contact and heel off times for that foot. The acceleration signal was reduced to a series of pulses by studying the negative positive changes in acceleration. It was found that there was approximately a 150 ms delay between heel contact and the negative positive change in acceleration. This delay was consistent across different walking speeds, but was different between subjects and when hemiplegic gait was simulated. Therefore, accelerometers placed on the trunk are valid sensors for the detection of heel contact events during FES assisted walking IPEM. Published by Elsevier Ltd. All rights reserved. Keywords: FES; Accelerometry; Hemiplegia; Gait events; Sensors 1. Introduction Since the 1960s, functional electrical stimulation (FES) has been used to correct foot drop in stroke patients [1]. It works by delivering an electrical stimulus to the tibialis anterior during the swing phase of gait to allow proper foot clearance and prevent the need for hip circumduction or other unstable gait compensations. While FES systems have proved successful in improving walking ability in stroke patients by improving walking speed, functional mobility, step length on the affected side, ankle flexion, muscle strength, motor coordination and balance [2 8], there are some fundamental problems associated with their use. Baker [9] has identified two such problems; the complexity of the gait cycle and the need for implantation of the system, and both problems put demands on the sensor and control system used. For any FES system to work in an effective manner, Corresponding author. Tel.: ; fax: address: gerard.lyons@ul.ie (G.M. Lyons). it requires a sensor to detect when the stimulus should be delivered. At first, the system can be initiated manually by an external switch. However, during the gait cycle itself, stimulation should be automatic. Therefore, a sensor is required to detect the phases of the gait cycle, so that the stimulation can be delivered during the swing phase of gait on the affected side. In order to do this effectively, the sensor must be able to detect heel off of the affected limb to trigger stimulation just prior to the swing phase and either heel contact or heel off of the contra-lateral leg to stop stimulation. Originally, footswitches, usually based on force sensitive resistors, were placed on the sole of the foot and used as sensors for FES [2]. This seems like the natural choice of sensor; heel loading and unloading can be detected with a footswitch. However, the footswitch as a sensor for FES assisted walking correction has many problems associated with it, most notably, reliability and the lack of implantation potential. The reliability of footswitches is often diminished due to their tendency to detect heel contact events during the swing phase as small forces are exerted on the heel during swing. Similarly, the reliability of the footswitch is reduced in /$ IPEM. Published by Elsevier Ltd. All rights reserved. doi: /s (03)

2 880 A. Mansfield, G.M. Lyons / Medical Engineering & Physics 25 (2003) patients who shuffle their feet when they walk [10,11]. The implantability of the stimulator is important as patients often reject systems that require a lot of donning time or are not aesthetic. This problem may be addressed through the use of implanted cuff electrodes placed on the sural nerve to use the patient s own sensory nerves to detect heel contact events [12]. However, such sensors are still ineffective in patients who shuffle their feet when they walk [10,11]. Hand switches have also been used to trigger stimulation, which may be preferable in the event of encountering an obstacle or uneven terrain [13]. However, hand switches increase the attentional demands of walking, which may reduce the patient s capacity to respond to perturbations to their balance [14], leading to an increased risk of a fall event. Accelerometers have the potential to be used as sensors for FES control systems. Accelerometers have been used previously to detect phases of the gait cycle [15], although these authors have placed the accelerometer on the leg, which requires more involved data processing. Implantation of the accelerometer could be carried out in the abdominal cavity of the patient, which would allow for implantation of both stimulator and sensor at the same site. Placement of the accelerometer on the trunk facilitates detection of phases of the gait cycle using changes in measured acceleration patterns and a trunk-mounted accelerometer has also been shown to measure instantaneous walking speed [16], which could provide another control input for stimulus intensity. From the placement of the accelerometer on the trunk, it may be possible to detect phases of the gait cycle by changes in acceleration patterns. It was the purpose of this study to determine the potential of accelerometry to detect gait events for their prospective use in a FES stimulator for the correction of drop foot in hemiplegic patients. 2. Theoretical background Changes in the kinematic parameters of gait are cyclical. That is, the shape of the acceleration curve recorded for one step is repeated for every step during continuous walking. Therefore, by studying the acceleration signal throughout the gait cycle, it is possible to determine gait events. This occurrence has already been exploited by Currie et al. [17] to detect gait cycle events by examining the acceleration time curve during walking. However, these authors used manual inspection to determine heel contact events rather than an automated computer algorithm. Fig. 1A shows a sample of an anterior posterior (AP) approximated centre of mass (CoM) acceleration signal to illustrate that this signal is cyclical during continuous walking. In order to mark the beginning and end of a step on this figure, a simultaneous footswitch output is shown. The footswitch signal measured heel contact events from one foot alone. According to this output, the heel is in contact with the ground, when it reads 1 and unloaded when it reads 0. It is evident from this graph that positive peaks in AP acceleration during walking occur just prior to unloading of the heel, whereas negative peaks in AP acceleration occur immediately following heel contact. By observing the positive/negative changes in AP acceleration throughout the gait cycle, one can predict heel contact and heel unloading times. The acceleration signal as presented in Fig. 1A is difficult to interpret without using complicated rules. Therefore, in order to simplify the algorithm, the acceleration signal can be reduced to a pulse-like signal similar to the footswitch signal. This is done by assigning a value 1 if the acceleration signal is increasing (a negative positive change) and a value 0 if the acceleration signal is decreasing (representing a positive-negative change). This results in the acceleration pulse presented in Fig. 1B. The data presented in this figure is the same as that presented in Fig. 1A, with the acceleration signal reduced to pulses. There are a few important points evident upon inspection of Fig. 1B. The first is that there are two acceleration pulses for every footswitch pulse. Therefore, every alternate acceleration pulse recorded coincides with the stance phase of alternate feet. Furthermore, it seems that analysis of the negative positive change in the acceleration signal is most relevant to the prediction of heel contact events, with each negative positive change occurring after heel contact of the alternating feet. We can therefore assume that each negative positive change occurs at a constant time after the heel contact as measured by the footswitch. Moreover, if this constant can be determined experimentally, it is possible to predict heel contact events from both feet from the accelerometer output. 3. Methods This study aimed to assess the potential use of accelerometry as a sensor for the control of FES assisted walking for the correction of drop foot in hemiplegic patients. Therefore, the ability of the accelerometer to detect heel contact events was compared to that of a footswitch. In order to test the ability of the accelerometer to detect heel contact events, experimental trials were performed on four subjects. Ethical approval for this study was first obtained from the University of Limerick Research and Ethics Committee. All subjects gave their informed consent to participation. The subjects were aged between 22 and 24 years and had no musculoskeletal impairment or injury that affected their balance or gait. Table 1 details the subject information.

3 A. Mansfield, G.M. Lyons / Medical Engineering & Physics 25 (2003) Fig. 1. (A) (top) shows changes in AP horizontal acceleration during walking. The broken line shows the output from the footswitch (with 1 indicating on and 0 indicating off ) marking the beginning and end of each stride. The lower figure 9 (B) shows the reduction of the acceleration curve presented in (A) into a pulse-like signal. A value of 1 indicates a negative positive change (or increase) in the acceleration signal, while a value of 0 indicates a positive-negative change (or decrease) in the acceleration signal. The footswitch output is also shown for comparative purposes. Table 1 Information on subjects tested Subject Gender Age (years) Height (m) A Female B Female C Male D Female One accelerometer (ADXL202, Analog Devices Inc.) was attached to the lumbar region of the subjects. This accelerometer is a dual-axis accelerometer, but acceleration from only one axis was considered, that representing anterior posterior (AP) acceleration of the subject during walking. A 100 Hz anti-aliasing filter was applied to the accelerometer signal prior to sampling. The accelerometer was placed in a plastic box for stability, which was mounted on a plastic board and attached to some Velcro straps, allowing secure fastening to the subject with minimal movement. The accelerometer was placed in alignment with the lumbar spine, with the active axis perpendicular to the spine. The beginning of each stride was marked with a footswitch. This consisted of a force sensing resistor, configured in a voltage divider arrangement, placed on the heel of one foot; the affected side in simulated hemiplegic walking. The footswitch signal was processed using a threshold value set at 97.5%, so that all values, 97.5% or greater of the maximum value represented on, and all values less than 97.5% of the maximum represented off. It was found that this threshold level resulted in no false heel off detections (i.e. heel off being detected during stance) and fewer false heel contact detections (i.e. heel on being detected during swing). The heel contact time for each stride was then found by recording the time at which each on condition occurred. The accelerometer and footswitch were connected to a portable data logger (BM42, 32MB version, Biomedical Monitoring Ltd.), which allowed for the collection of data from walking trials. The three channels were sampled at 500 Hz for subject C, and 200 Hz for all other subjects. The data was then transferred via an external disk drive (Securmate SDDR-33, Sandisk) onto a CPU for post processing using MATLAB (version 4.2b, The Mathworks, Inc.). Gait trials were performed along a 40 m length of corridor. The surface of the corridor was covered in shallow carpeting. The subject was asked to walk in a straight line during the trials. Three different groups of trials were performed which, for the purposes of this analysis, were grouped as either steady, changing or hemiplegic walking. Steady trials involved the subject walking either as normal, or at a constant speed determined by a metronome. Cadences from 60 to 140 steps

4 882 A. Mansfield, G.M. Lyons / Medical Engineering & Physics 25 (2003) per minute were tested using the metronome. During changing trials, subjects were asked to alternate their walking speed between normal, faster than normal or slower than normal about half way down the corridor. Hemiplegic walking was simulated by halving the step length on the unaffected side. Subjects were also asked to walk slower than normal during these trials and as if they had a limp. Forty trials in all were performed with each subject; 24 steady trials, 8 changing trials and 8 hemiplegic trials. Post processing using MATLAB low-pass filtered the acceleration data at 2 Hz and compensated for the component of gravity measured during the trials by measuring the average angle of tilt of the accelerometer during quiet stance at the start of the trial. The timing of negative positive changes in acceleration values were calculated by reducing the acceleration signal to a series of pulses as described in the previous section. The times of heel contact as determined by the footswitch were also calculated. However, there were two negative positive acceleration changes for every stride, as each negative positive change occurs synchronously with foot contact of each foot. Therefore, the MATLAB routine deleted every second value from the negative positive timing vector produced. It did this by detecting the first acceleration pulse immediately following heel contact as detected by the footswitch. This was the first event of note from the accelerometer, and all acceleration pulses recorded previous to this were assumed to be due to postural sway or noise. The acceleration pulse immediately following this first event was deleted, and every second subsequent acceleration pulse was deleted to give unilateral information. Initially, the gait timing data was not synchronised exactly. In some instances, either the footswitch or the accelerometer falsely detected a heel contact event (i.e. detected more than one event per gait cycle). These values were removed manually using Miscrosoft Excel (Version 9.0, Microsoft Corporation). The percentage of incorrect foot event detections for both the footswitch and the accelerometer was recorded. Following this, the delay between foot contact timing as calculated by the footswitch and the negative positive acceleration changes was found for each step. A oneway ANOVA on SPSS (release , SPSS Inc.) was used to determine if these differences were the same for all subjects and for the three different types of trial (i.e. steady, changing and hemiplegic). Post-hoc analysis using Tukey s HSD determined which, if any, of the three groups of trials were different to each other. 4. Results A total of four subjects completed 40 gait trials each. Six trials were removed from analysis due to a fault with the recording device, leaving 154 trials for analysis. This resulted in a total of 2626 strides. Both the accelerometer and the footswitch were not wholly accurate in the prediction of heel contact events. In some instances, they falsely indicated a heel contact event. The error in detecting heel contact events was determined manually on Microsoft Excel, and the results of this determination are found in Table 2. From this table, it is evident that the accelerometer was between 98.2% and 99.8% reliable in the detection of heel contact events, whereas the footswitch is between 92.4% and 98.7% reliable when all the trials are considered together. Analysis of steady trials alone for all subjects pooled, revealed that there was on average a 147 ms (±91 ms standard deviation) delay in heel contact as measured from the footswitch and the negative positive change in AP acceleration. Therefore, there was a consistent and repeatable relationship between the output from the footswitches and the output from the accelerometer in terms of predicting heel contact events. Analysis of the mean difference in timing of rise in footswitch voltage and negative positive change in the accelerometer output between subjects revealed that there is large variation in this mean difference between subjects (p 0.001). This analysis included all trials. Table 3 shows the average difference for each subject. While the variation in the average delay for all subjects pooled was quite large (standard deviation of 91 ms), this variation greatly reduces when the results for each different subject are presented individually. It is possible that a large proportion of this variation is due to differing placement of the footswitch during the trials. While every effort was made to ensure that the footswitch was placed in approximately the same place for every subject, differences in foot and shoe size and the type of shoe worn between subjects may have meant that there were slight delays in the loading of the footswitch. However, it is also likely that the mean difference is due to differences in walking strategy between subjects. The three different trial conditions, steady, changing and hemiplegic, were then compared within subjects in terms of the delay in heel contact identification and negative positive acceleration change. Analysis revealed that these delays were statistically different across trial conditions (p 0.001). However, post-hoc using Tukey s HSD analysis revealed that the steady and changing conditions were similar to each other (p = 0.991). Fig. 2 illustrates this for all trials for all subjects. 5. Discussion Results from this study indicate that there was a negative positive change in acceleration approximately 150 ms after heel contact as identified by the foot switch. However, there was large variation observed in this figure, and while this variation was reduced when individ-

5 A. Mansfield, G.M. Lyons / Medical Engineering & Physics 25 (2003) Table 2 Error in heel contact event prediction from the footswitch and the accelerometer for all trials (steady, changing and hemiplegic), normal walking alone and simulated hemiplegic walking. The values displayed reflect the percentage of times either the footswitch or the accelerometer correctly detected a heel contact event Subject Correct footswitch detection (%) Correct accelerometer detection (%) All Normal Hemiplegic All Normal Hemiplegic A B C D Table 3 Average delay in heel contact as determined by the footswitch and the negative positive change in AP horizontal acceleration Subject Delay (ms) Standard deviation (ms) A B C D Fig. 2. The average delay in heel contact and negative positive acceleration change for all four subjects for each type of trial. Significant differences in these delays were found between subjects (p 0.001). Significant differences were also found between the simulated hemiplegic trials and the steady/changing trials (p 0.001). ual subjects data was considered, the standard deviations observed were still quite high (see Table 3). This variation may be largely due to the limitations of the use of thresholding for determination of heel contact events from the footswitch. That is, the use of an absolute voltage, above which is considered on and below which is considered off. The value used in this study (97.5% of maximum voltage) yielded fewer false gait event detections, although there were still more detected than with the accelerometer (from Table 2). More sophisticated algorithms may be used for processing the footswitch data and subsequently determining heel contact events, although the methods used here reflect those typically used in FES applications [5]. Furthermore, different timing in negative positive changes in acceleration was found when hemiplegic walking was simulated. These different timing patterns observed indicate that when gait is altered significantly (e.g. made asymmetrical), the timing of trunk acceleration/deceleration in relation to heel contact changes. This same difference is not observed with simply a change in walking speed. This occurrence has a significant effect on the clinical use of the accelerometer in an FES system. That is, in order to determine the temporal relationship between trunk acceleration changes and FES stimulation, pathological subjects will have to be used. However, further research is required to determine the exact nature of this relationship. The concept of the precise detection of heel contact events is somewhat misleading. The transitions between the phases of the gait cycle are gradual and very often two different and equally experienced investigators may disagree on the exact moment of, for example, heel contact based on video recordings. This proves problematic in the automatic detection of gait events using artificial sensors. While footswitches can provide a simple on/off output for that part of the foot they occupy, the relatively slow loading and unloading of that part of the foot can distort the results obtained. Therefore, no sensor can be deemed wholly accurate in the precise detection of gait events. The inadequacy of footswitches in detecting foot contact events is evident in this paper. Footswitches, on average, detected heel contact falsely % of the time (see Table 2). It is thought that pressure from the sole of the shoe on the footswitch caused some on footswitch signals being output during the swing phase. Furthermore, relatively slow unloading of the heel may be responsible for false heel contact detection immediately following heel off. Indeed, these false heel contact detections are evident in Figs. 1 and 2 immediately following the first and second stride. Shuffling of the feet, leading to an unclear heel contact, can also cause these false heel contact detections. By contrast, the accelerometer was

6 884 A. Mansfield, G.M. Lyons / Medical Engineering & Physics 25 (2003) less likely to output a negative positive acceleration change at an inappropriate time in the gait cycle. Noise in the acceleration signal can obviously cause such contamination of the signal, but filtering at 2 Hz removed most of this noise and resulted in % false negative positive changes detected for all trials. Furthermore, positive/negative changes recorded during quiet stance due to postural sway could result in the false detection of foot contact events. Therefore, some rule should be implemented in the sensor, stating that walking speed should be above a predetermined amount before initiation of the stimulus. Despite the shortcomings in footswitches for the detection of heel contact events, it is generally accepted in the research literature that footswitches are the gold standard in heel contact timing detection, particularly for use as FES sensors [2]. However, their use is limited in stroke patients and other such individuals whose gait is such that they shuffle their feet when they walk [10,11]. Indeed, this study supports this assertion, with improved gait event detection of the accelerometer in two subjects when compared with the footswitch during simulated hemilpegic walking (subjects C and D, Table 2). The improved performance was more prevalent in subject D, who from observation during the trials, seemed to have a less defined heel contact than the other subjects. Therefore, the accelerometer seems a better choice of sensor for individuals with a poorly defined heel contact, such as those with drop foot. Even during the normal walking in our non-pathological subjects, the accelerometer did not detect any false gait events, but the footswitch did, again possibly due to undefined heel contact or insole pressure during swing. Decreased performance of the accelerometer compared with the footswitch was observed for subjects A and B during simulated hemiplegic walking. While every effort was made to ensure that the subjects were familiar with the simulated hemiplegic walking patterns, this could be a result of unnecessary accelerations and decelerations of the trunk caused by the atypical walking patterns required. This is not the first study to use accelerometry to detect heel contact events during walking. As previously mentioned, Currie et al. [17] used manual inspection of acceleration patterns observed during walking to record the timing of heel contact events. Other studies [15,18,19] have used accelerometers placed on the leg to detect heel contact events. Indeed, Williamson and Andrews [15] have stated that a sensor which is capable of detecting heel contact events and measuring joint angle information may be necessary and the accelerometer is capable of meeting both of these criteria. However, this is the first study to report the automated determination of heel contact timing events based on the analysis of acceleration curves of an approximation of the centre of mass. By placing the accelerometer on the trunk, the detection algorithm is simplified; the acceleration signals presented in this study, in Figs. 1 and 2, are much easier to interpret than those reported by Williamson and Andrews [18]. Additionally, the accuracy of heel contact detection found in this study is comparable to accuracies obtained previously, when more involved rules for heel contact detection were used [15]. Furthermore, the accelerometer may be implanted in the abdomen to facilitate a fully implantable FES system. One of the most important issues when dealing with accelerometers is the need to correct for the gravitational component of measured acceleration. While this was corrected for in this study, compensation for the tilt of the accelerometer may not be necessary when the accelerometer is used solely for the detection of heel contact events as it is the shape of the acceleration curve, rather than the values obtained that are studied. 6. Conclusions The results of this study indicate the possibility of a lightweight, implantable sensor for the detection of gait events. An accelerometer may be placed on the trunk of an individual and a simple yet effective algorithm can be used to automatically detect heel contact of both feet. Acknowledgements The authors wish to acknowledge the assistance of Ms Karen Culhane and Ms Elizabeth Egan in the completion of this project and the preparation of this manuscript. References [1] Liberson WT, Holmquest HJ, Scott D, Dow M. Functional electrotherapy: stimulation of the peroneal nerve synchronized with the swing phase of the gait of hemiplegic patients. Arch Phys Med Rehabil 1961;42: [2] Brandell BR. Development of a universal control unit for functional electrical stimulation (FES). Am J Phys Med 1982;61(6): [3] Cozean CD, Pease WS, Hubbell SL. Biofeedback and functional electrical stimulation in stroke rehabilitation. Arch Phys Med Rehabil 1988;69: [4] Burridge J, Taylor P, Hagan S, Swain I. Experience of clinical use of the Odstock dropped foot stimulator. Artif Organs 1997;21(3): [5] Burridge JH, Taylor PN, Hagan SA, Wood DE, Swain ID. The effects of common peroneal stimulation on the effort and speed of walking. A randomised controlled trial with chronic hemiplegic patients. Clin Rehabil 1997;11(3): [6] Stefanovska A, Vodovnik L, Gros N, Rebersek S, Acimovic- Janezic R. FES and spasticity. IEEE Trans Biomed Eng 1989;36(7): [7] Glanz M, Klawansky S, Stason W, Berkey C, Chalmers TC. Functional electrostimulation in poststroke rehabilitation: a metaanalysis of the randomised controlled trials. Arch Phys Med Rehabil 1996;77(6):

7 A. Mansfield, G.M. Lyons / Medical Engineering & Physics 25 (2003) [8] Van Griethuysen CM, Paul JP, Andrews BJ, Nicol AC. Biomechanics of functional electrical stimulation. Prosthet Orthotics Int 1982;6(3): [9] Baker LL. Electrical stimulation to increase functional activity. In: Nelson RM, Hayes KW, Currier DP, editors. Clinical electrotherapy. 3rd ed. Stamford: Appleton and Lange; p [10] Aminian K, Rezakhanlou K, De Andres E, Fritsch C, Leyvraz PF, Robert P. Temporal feature estimation during walking using miniature accelerometers: an analysis of gait improvement after hip arthroplasty. Med Biol Eng Comput 1999;37(6): [11] Von Schroeder HP, Coutts RD, Lyden PD, Billings E, Nickel VL. Gait parameters following stroke: a practical assessment. J Rehabil Res Dev 1995;32(1): [12] Haugland MK, Sinkjaer T. Cutaneous whole nerve recording used for correction of footdrop in hemiplegic man IEEE. Trans Rehabil Eng 1995;3(4): [13] Bajd T, Kralj A, Turk R, Benko H, Šega J. The use of a fourchannel electrical stimulator as an ambulatory aid for paraplegic patients. Phys Ther 1983;63(7): [14] Lundin-Olsson L, Nyberg L, Gustafson Y. Stops walking when talking as a predictor of falls in elderly people. Lancet 1997;349:617. [15] Williamson R, Andrews BJ. Sensor systems for lower limb functional electrical stimulation (FES) control. Med Eng Phys 2000;22: [16] Lyons GM, Wilcox DJ, Grace PA. Real-time measurement of walking speed using an integrated accelerometer. In: Proceedings of the 6th Annual Conference of the International Functional Electrical Stimulation Society, Cleveland, USA p [17] Currie G, Rafferty D, Duncan G, Bell F, Evans AL. Measurement of gait by accelerometer and walkway: a comparison study. Med Biol Eng Comp 1992;30(6): [18] Williamson R, Andrews BJ. Gait event detection for FES using accelerometers and supervised machine learning. IEEE Trans Rehabil Eng 2000;8(3): [19] Willemsen AThM, Bloemhof F, Boom HBK. Automatic stance swing phase detection from accelerometer data for peroneal nerve stimulation. IEEE Trans Biomed Eng 1990;37(12):