Computational methods to detect step events for normal and pathological gait evaluation using accelerometer H.-K. Lee, J. You, S.-P. Cho, S.-J. Hwang, D.-R. Lee, Y.-H. Kim and K.-J. Lee The presented study highlights the feasibility and accuracy of novel computational methods based on a morphological filter and a least square acceleration filter to detect step events for evaluating normal and pathological gait parameters using a single accelerometer. This is the first evidence that demonstrates the feasibility and accuracy of the novel accelerometer-based system and methods in both normal and pathological populations. Introduction: The development of a method for gait monitoring systems provides accurate and reliable motion analysis in ergonomics, sports and rehabilitation applications. Contemporary equipments using either optoelectronics or ultrasound have been widely used to assess bio- mechanical characteristics of human locomotion. However, cost, lack of portability, as well as real-time monitoring and accuracy issues for such systems have made them impractical for many clinicians and researchers under ergonomic workplaces or clinics, field or sporting events and rehabilitation settings [1]. Recently, miniature sensors, such as inertial, magnetic or gyroscopic sensors, combined with accel- erometers along with optimal methods have been successfully used to characterise the biomechanical patterns of locomotor behaviours and postural control in the normal population [2]. However, there is a dearth of clinical evidence that has successfully elucidated important characteristics of the pathological locomotor behaviours in stroke and cerebral palsy using a single accelerometer. This may be owing to the issues associated with method optimisation featuring different or aber- rant locomotor patterns and movement artifacts, as well as magnetic interference. Hence, we have developed two computational methods to measure step events during locomotion in both normal and pathological populations. The methods are based on 1. the morphological filter (MF) and 2. the least square acceleration filter and morphological filter (LSAF-MF). Participants: A convenient sample of 18 participants (four healthy adults, six adults with chronic hemiparetic stroke and eight children with cerebral palsy) were recruited from a community hospital. The demographic and clinical characteristics of the participants are presented in Table 1. Informed consent forms were obtained from all participants prior to participation in this study. Table 1: Results of comparative errors of each participant group Subject Sex Age Height (cm) Weight (kg) Clinical information Comparative results (steps/minute) Markers Accel. (MF) Error Accel. (LSAF-MF) Error NO1 M 26 163.0 55.8 – 36.1 +1.0 35.6 +0.7 0.5 +0.3 36.0 +0.5 0.1 +0.5 NO2 M 20 170.2 67.4 – 39.5 +0.3 39.8 +0.7 0.3 +0.4 39.8 +0.7 0.3 +0.4 NO3 M 23 185.5 76.3 – 41.1 +2.0 41.1 +1.9 0.0 +0.1 41.0 +1.9 0.0 +0.1 NO4 M 22 178.0 78.9 – 39.9 +0.6 40.0 +0.5 0.1 +0.1 39.9 +0.6 0.0 +0.0 (Average) 22.8 +2.5 174.2 +9.7 69.6 +10.4 – – – 0.23 +0.22 – 0.10 +0.14 ST1 F 54 156.6 57.1 L.H.S. 47.1 +1.5 47.2 +1.7 0.1 +0.2 47.0 +2.3 0.1 +0.8 ST2 M 67 162.0 61.6 L.H.S. 37.7 +0.9 37.5 +0.9 0.2 +0.0 37.5 +0.9 0.2 +0.0 ST3 F 54 166.9 80.1 R.H.S. 47.5 +1.3 47.6 +1.5 0.1 +0.2 47.5 +1.6 0.0 +0.0 ST4 M 48 171.2 70.0 R.H.S. 42.2 +1.2 41.3 +1.2 0.9 +0.0 41.6 +1.3 0.6 +0.1 ST5 M 71 178.2 62.8 R.H.S. 52.0 +3.5 52.1 +3.5 0.1 +0.0 52.1 +3.5 0.1 +0.0 ST6 F 46 157.5 56.6 L.H.S. 58.9 +5.5 59.0 +5.4 0.1 +0.1 59.0 +5.4 0.1 +0.1 (Average) 56.7 +10.1 165.4 +8.4 64.7 +8.7 – – – 0.25 +0.32 – 0.18 +0.21 CP1 M 9 126.1 21.8 S.D.C.P 27.2 +3.4 27.1 +3.4 0.1 +0.0 27.1 +3.3 0.1 +0.1 CP2 F 6 121.0 24.9 R.S.H.C.P. 29.5 +4.0 29.7 +4.0 0.2 +0.0 29.7 +4.1 0.2 +0.1 CP3 F 6 111.5 25.0 S.D.C.P 28.1 +3.3 28.7 +4.1 0.6 +0.8 28.8 +4.4 0.7 +1.1 CP4 F 5 114.0 22.8 L.S.H.C.P. 28.9 +1.1 28.9 +1.1 0.2 +0.0 28.9 +0.9 0.2 +0.2 CP5 M 7 121.0 31.3 S.D.C.P 26.9 +1.5 27.3 +2.1 0.4 +0.6 26.8 +1.4 0.1 +0.1 CP6 F 7 122.2 27.0 R.S.H.C.P. 34.8 +1.9 34.2 +1.8 0.6 +0.1 34.2 +1.7 0.6 +0.2 CP7 F 5 85.5 8.5 R.S.H.C.P. 26.1 +4.3 26.0 +4.4 0.1 +0.1 26.0 +4.4 0.1 +0.1 CP8 F 6 111.0 18.6 R.S.H.C.P. 28.1 +1.9 28.3 +1.9 0.2 +0.0 28.3 +1.9 0.2 +0.0 (Average) 6.3 +1.3 114.0 +12.8 22.5 +6.8 – – – 0.30 +0.21 – 0.28 +0.24 Total – – – – – – 0.27 +0.24 – 0.21 +0.21 L.H.S.: left hemiplegia stroke; R.H.S.: right hemiplegia stroke; S.D.C.P.: spastic diplegic cerebral palsy; R.S.H.C.P.: right spastic hemiplegic cerebral palsy; L.S.H.C.P.: left spastic hemiplegic cerebral palsy Instruments and data acquisition: The accuracies of the methods using a single three-axis accelerometer (CXL02LF3, Crossbow, USA) were evaluated using a six optical camera motion capture system (Vicon, Oxford Metrics Ltd, UK) as the reference. The accelerometer was placed on the second sacrum and one reflective marker was placed on each heel. The accelerometer and reflective marker data were synchro- nously captured at a sample rate of 120 Hz while the participants walked on an eight-metre walkway at a self-selected speed. Data for four successful walking trials were obtained from each participant. Computational methods: 1. Method based on MF: The MF is a non- linear signal transformation method that modifies the shape information of signals [3], which is based on set operations. The method based on MF is applied to detect the maximal peak acceleration point value obtained from the accelerometric measurement at the initial contact of each gait cycle. The MF consists of four basic operations [3]: Erosion: (f ⊖ k )(m)= min n=0,...,M−1 f (m + n)− k (n), for m = 0, ... , M − N (1) Dilation: (f ⊕ k )(m)= max n=m−M+1,...,m f (n)− k (m − n), for m = M − 1, M , ... , N − 1 (2) Opening: ( f W k )(m) = [(f ⊖ k )⊕ k ](m) (3) Closing: ( f †k )(m) = [(f ⊕ k )⊖ k ](m) (4) where f (m) is the measured data and k(m) is a structuring element (SE). The block diagram of the method based on MF and the results of each method of operation are presented in Fig. 1. First, a vector magnitude of the accelerometer data ( f1) was computed. Secondly, a closing operation (C1) was computed using f1 and SE with a sample length of 24 and a constant value of 1. Thirdly, an opening operation (O1) was computed using the C1 and SE with a sample length of 12 and a constant value of 1. Finally, we obtained data (R1) to detect the gait steps via subtrac- tion operation, which was operationally expressed as the resultant R1. To compute the peak point featuring the acceleration envelope for each gait step, we applied an experimental threshold (0.005) to the vector magni- tude-based data using R1. The gait steps were determined by calculating the peak points over the threshold point. i = x,y,z f1 = Σ ACC 2 i C1 = f1 k1 O1 = C1 k1 R1 = C1– O1 f 1 C1 O1 350 400 450 500 550 600 time (samples) R1 Fig. 1 Block diagram and data morphology of method based on MF 2. Method based on LSAF-MF: The LSAF is a simple mathematical method which requires differentiating the discrete-time signals to extract the sharpest peak of the accelerometer data [4]. The MF was then applied as described above. The block diagram using the method based on LSAF-MF is presented in Fig. 2. First, a vector magnitude of the accelerometer data ( f2) was computed. Secondly, the fourth- order LSAF was applied to amplify the most morphologically distinct data points (LSA). Thirdly, a closing operation (C2) was calculated using f and the SE with a sample length of 24 and a constant value of 1. Fourthly, an opening operation (O2) was computed using the C2 and the SE with a sample length of 12 and a constant value of 1. Finally, we obtained the data (R2) required to detect the gait steps via a subtraction operation, which was operationally expressed as the resultant R2. To compute the peak point featuring the acceleration envel- ope for each gait step, we applied an experimental threshold (0.005) to the vector magnitude-based data using R2. The gait steps were then determined by computing the peak points over the threshold point. ELECTRONICS LETTERS 19th August 2010 Vol. 46 No. 17