Disability assessment using visual gait analysis Sherif El-Sayed Hussein Computer and Systems Department, Mansoura University, Mansoura, Egypt Tel: +2 0122205369 Fax: +2 0502244690 E-mail: sherif_hussein@mans.edu.eg Abstract The accurate assessment of pathological gait for individual subjects is a major problem in rehabilitation centers. Automated or semi-automated gait analysis systems are important in assisting physicians in the diagnosis of various diseases. However, these systems are not only highly sophisticated but also require superior quality cameras and complex software. Moreover, modern 3D motion capture systems allow large amounts of data from patients with walking difficulties to be recorded, and this data often proves difficult to interpret for clinical staff trying to gain insight into the patient s condition. Automation and simplification of the analysis of gait data is therefore necessary if it is to be used more productively. This research proposes a simple and cost effective approach that utilizes artificial intelligence techniques to automate the analysis and diagnosis processes. It also offers a means to compare different treatment methods and their effectiveness during the course of treatment. Visualization software has also been developed to help increase diagnostic reliability. Keywords: Rehabilitation, gait analysis, and artificial intelligence 1. Introduction Walking is the body s natural means of moving from one location to another. Functional versatility allows the lower limbs to readily accommodate stairs, doorways and obstacles in the path of progression. Efficiency in these endeavors depends on free joint mobility and muscle activity that is selective in timing and intensity. As various types of pathology alter mobility and muscular effectiveness, the patients substitute wherever possible, yield when they must, and accept compensatory reactions from adjacent segments as they occur. The resulting walking pattern is a mixture of normal and abnormal motions that differ in significance. Energy costs are increased and functional versatility is compromised [1]. Gait analysis allows the evaluation of abnormalities of multiple joints and the resulting multilevel concurrent treatment. It provides a way to differentiate primary conditions from secondary compensations and a means for following quantitative treatment results [2]. The physician him/herself can interpret the gait data obtained depending on his experience with a non-automated diagnosis. Since a non-automated diagnosis requires a high level of expertise, only specifically trained orthopedists or physiatrist can use gait data. The automated system is expected to decrease this requirement which may help to increase the number of physicians and patients making use of the laboratory. In addition, automated systems save experts time and decrease the possibility of human error [3]. Qualitative methods for identification and recording of gait deviation have played a role in patient care for decades. In 1937, Boorstein identified 14 disease processes that could be diagnosed with the help of gait assessment. He described seven major gait deficit groups. In the late 1950s, Blair Hangar, the founder of Northwestern University's School of Prosthetics and Orthotics, and Hildegard Myers, a physical therapist at Rehabilitation Institute of Chicago, collaborated to develop the first comprehensive system of clinical gait analysis for persons with trans-femoral amputation. They identified 16 gait deviations and suggested numerous patient and prosthetic causes for each [4]. Gait-based methods, such as the Rancho observational gait assessment, seek to identify and differentiate pathologic versus 3-103
compensatory mechanisms and, as a result, guide specific surgical, therapeutic, orthotic, or prosthetic interventions for a particular patient [5]. For the design of automated diagnoses systems various pattern recognition algorithms were used. These are neural networks (NNs) [6], support vector machines (SVMs) [7], and radial basis functions (RBFs) [8]. There are also studies in which NNs are trained by force platform data to distinguish healthy from pathological gaits [9]. In addition to these, there are studies to recognize walking people among a few subjects (less than 10) by using joint angles as features [10]. These studies produce reasonable results for NNs used in gait classification. However, as the dimension of the features and the size of the data increase, accuracies may not be guaranteed. In similar pattern recognition studies there have been attempts to solve this problem by combining classifiers to increase performance. The objective of this study is to design a software system to supply physicians with accurate and practical ways to diagnose and further classify a musculoskeletal disease and to provide qualitative parameters to help in assessing treatment effectiveness, using a simple system. 2. Gait cycle and phases of gait Normal locomotion is a highly complex activity requiring coordinated controlled movement of all the joints of the lower limbs. Several methods may be used to describe and analyze this activity. The simplest technique is based on an analysis of the pattern of foot contact with the ground. A single step with one leg may be considered as compromising a stance phase, when the foot is in contact with the ground, followed by a swing phase, when the leg is swinging forward above the ground in preparation for the next step. Bipedal locomotion is achieved by the alternate stepping action of the two limbs. When walking, both feet are in contact with the ground for a short period which is referred to as the double support phase. During running, there is a period when both feet are simultaneously off the ground. In normal locomotion, the timing, duration and symmetry of these phases is remarkably consistent; any change is an indication of pathology [11]. While using normal events such as the critical actions separating the phases has proved appropriate for some patients, it has often failed to accommodate the gait deviations of patients impaired by paralysis or arthritis. For example, the onset of stance has customarily been called heel strike; yet the heel of a paralytic patient may never contact the ground or may do so much later in the gait cycle. Analysis of a person s walking pattern by phases identifies the functional significance of the different motions occurring at the individual joints more directly. The phases of gait also provide a means for correlating the simultaneous actions of individual joints into patterns of total limb function. This is a particularly important approach for interpreting the functional effects of disability. The relative significance of one joint s motion compared to another varies among gait phases. Also, a posture that is appropriate in one gait phase would signify dysfunction at another point in the stride, because the functional need has changed. As a result, both timing and joint angle are very significant. This latter fact adds to the complexities of gait analysis [5]. The sequential combination of phases enables the limb to accomplish three basic tasks. These are weight acceptance, single limb support and limb advancement. Weight acceptance begins during the stance period and uses the first two gait phases (initial contact and loading response). Single limb support continues the stance during the next two phases of gait (mid stance and terminal stance). Limb advancement begins in the final phase of the stance (preswing) and then continues through the three phases of swing (initial swing, mid-swing and terminal swing) as shown in Fig. 1. 3-104
Fig. 1. A complete gait cycle can be viewed in terms of the three functional tasks of weight acceptance, single limb support, and limb advancement. 3. Materials and methods The methodology starts by recording the lower limb movement at 25 frames per second during the gait cycle and saving it to an AVI file for further processing using a LabVIEW program. The program is interfaced with a camera using an ActiveX that can acquire, save and display the recorded video. The program converts the AVI file to a sequence of BMP images and can forward or reverse the film frame by frame. This function allows the physician to mark the beginning of each phase on the relevant frame for further processing using MatLab. 3.1. Experimental protocol Three investigators A, B, and C were involved in all the experiments. The procedure for performing the experiments was divided into the following steps: 1. The subject arrived. The experimental procedure was explained to the subject and any questions were answered. 2. A pair of socks in two colours (black for left leg and white for right leg) was fitted to the subject s feet, to be identified in the recorded video. 3. Investigator B observed the gait cycle during the experiments and placed a one meter red bar in the same direction of the motion for calibration purposes. 4. The motion was recorded by investigator C in the sagital plan, using one camera connected to the computer and derived by the LabVIEW program using an ActiveX Control. The recording speed was 25 frames/sec. 5. Investigator C played back the recording and, if necessary, asked the subject to repeat the test. 3.2. Features selection Although there are 8 phases during the gait cycle for each leg, only 5 main event points were selected during the gait cycle; initial contact, loading response, pre-swing, mid swing and terminal swing; each leg, in turn, will have 4 time durations and 4 displacements. This adds up to 12 inputs, representing time ratios for each selected phase calculated between the left and the right leg and the ratios of the distances of each selected phase, relative to the whole cycle. 3.3. The Neuro-Fuzzy classifier Classifiers for both the degree and level of disability were designed using an adaptive neuro-fuzzy inference system (ANFIS), which can be found in the MATLAB Fuzzy Logic Toolbox. The data utilized was collected from normal subjects as well as subjects with pathological gait who performed free walking and was recorded and analyzed using the LabVIEW program. 3-105
ANFIS is functionally based on the Surgeno-type fuzzy rule base and at the same time has an architecture equivalent under some constraints [12] to a radial basis function neural network, allowing the system to learn from the training data. The design started by subtractive clustering to determine the number of rules and the input membership functions. The membership function of choice was the generalized bell function: ( x) 1 x c 1 a Subtractive clustering is an unsupervised algorithm based on a measure of the density of the normalized data points in the feature space. The point with the highest number of neighbors is selected as the center for a cluster. The data points within a pre-specified fuzzy radius are then removed and the algorithm looks for a new point with the highest number of neighbors until all the data points are checked. The following two rules were a part of the first-order Surgeno-type rule base: If u 1 is A 1 and u 2 is B 1 then y 1 =c 11 u 1 +c 12 u 2 +c 10 (2) If u 1 is A 2 and u 2 is B 2 then y 2 =c 21 u 1 +c 22 u 2 +c 20 (3) The fuzzy classifier can interpolate between the two linear rules depending on their state. So, if the firing strengths of the rules are α1and α2 for two inputs u1 and u2, respectively, then the output based on weighted average is: y=( 1 y 1 + 2 y 2 )/( 1 + 2 ) = 1 y 1+ 2 y 2 = 1 (c 11 u 1 +c 12 u 2 +c 10 )+ 2 (c 21 u 1 +c 22 u 2 +c 20 ) = 1 c 11 u 1 + 1 c 12 u 2 + 1 c 10 + 2 c 21 u 1 + 2 c 22 u 2 + 2 c 20 (4) Using the least-squares method, cij, (i =1,2 and j = 0,1,2) could be adjusted in the forward pass while the membership function s parameters ai, bi, and ci could be adjusted by gradient descent using the error signals that propagate in the backward pass [13]. The neuro-fuzzy classifiers were implemented using MATLAB and then integrated into the LabVIEW program. 4. Results and discussion Table 1. The gait classification results from 21 tests. 2b Pathological Normal Number of subjects 14 7 Predicted average degree of disability 0.41 0.05 True average degree of disability 0.47 0 Predicted average level of disability 0.63 0.02 True average level of disability 0.68 0 In this study, 42 normal subjects and 61 subjects with differing degrees and levels of disability representing the output of the classifier were selected, according to the advice of a medical expert. The ratios between the chosen phase durations of the right leg to the durations of the left leg, and the ratios of the phase distances of each leg calculated relative to the whole cycle distance were selected as features of the subject s aforementioned gait, as inputs to the classifier. Two neuro-fuzzy classifiers were then designed using these sets of features to detect the subject s degree and level of disability. The output of the first classifier used to detect the subject s degree of disability ranges from 0 for a normal subject to 1 for a subject with a high degree of disability. The output of the second classifier used to detect the subject s level of disability ranges from 0 for a normal subject to 1 for a subject with a high level of disability. Five bell-shaped membership functions for each of the twelve inputs for each of the classifiers were chosen. A rule base of 72 rules were found with an average learning error of 0.103 for (1) 3-106
the first designed classifier while a rule base of 96 rules were found with an average learning error of 0.112 for the second classifier using ANFIS. These classifiers were integrated with a LabVIEW program and tested using 14 subjects with pathological gait and 7 normal subjects, which resulted in the data summarized in Table 1. The digital image analysis from the LabVIEW program allowed the physician to compute the stride length and stepping speed of each leg during the gait cycle. That extra information provides additional assessment to compare the gait with normal and pathological gait values as shown in Fig. 2. Fig. 2. The LabVIEW program testing a normal subject and showing both the stride data and the degree and level of disability. 5. Conclusion The technique designed in this research has a minimal impact on the natural motion of the subject and allows for the capture of data without the need to link the subject to the data acquisition hardware. As the automated system can give immediate results, subjects will not need to perform unnecessary tests, since once the test is successful the patient is free to go. The automation of the system along with the intelligent abilities programmed and integrated into it not only make the system simple but also make it easy to utilize with a minimum of experience. The system can be also used in places that lack both financial resources and highly experienced personnel, to serve as many patients as possible. The system also provides the physician with stride length and speed of walking data which can be invaluable in supporting the diagnosis of the gait pathology of a patient. In addition, the degree and level of disability can provide qualitative information that can help the physician to assess the progress of a given patient during the course of treatment. References 1. C. Kirtley. Clinical Gait Analysis: Theory and Practice. Churchill Livingstone. 2006. 2. A. Mostayed, M. Mynuddin, G. Mazumder, S. Kim, and S.J. Park. International Conference on Multimedia and Ubiquitous Engineering. 2008, pp. 36-40. 3. R. Noble and R. White. Proc 9-th Int. Conf. on Inf. Visualisation. 2005, pp. 247-252. 4. M. Lusardi, C. Nielsen. Orthotics & Prosthetics in Rehab. Butterworth-Heinemann. 2000. 5. J. Perry. Gait Analysis: Normal & Pathological Function. Slack Incorporated, 1992. 6. N. Koktas and N. Yalabik, Proc. of Int. Symposium on Health Informatics and Bioinformatics. 2005, pp. 174-179. 7. R. Begg, J. Kamruzzaman. Proc. of IEEE TENCON Conference. 2003, pp. 354-358. 8. M. Kohle and D. Merkl. Proc. of the ACM Symp. on Applied Computing. 2000, pp. 41-45. 9. W. Wu, F. Su, Y. Cheng, Y. Chou. Annals of Biomedical Eng. 2001, vol. 29, pp. 83-91. 10. R. Tanawongsuwan, A. Bobick. CVPR. 2001. pp. 726-735. 11. M. Whittle. Gait Analysis: An Introduction. Butterworth-Heinemann; 4 th edition. 2007. 12. R. Nowicki. IEEE Trans. on Knowledge and Data Eng. Apr. 2008, pp. 1-14. 13. Z. Chang, L. Liu, Z. Li. Int. Conf. on Int. Pervasive Computing. Oct. 2007, pp. 437-440. 3-107