Effective Clinical Gait Analysis using Supervised Learning Techniques

Abstract

Gait analysis has become a popular trend to solve critical problems in different application domains. It has demonstrated crucial significance in the clinical domain also. Quantification of gait pattern through subtle analysis of salient features have surpassed the pitfalls of prevailing qualitative assessment techniques. Computational intelligence techniques, specifically machine learning algorithms, have demonstrated competing performance in modeling non-linear data relationships of gait variables. Another issue that needs to be emphasized is the expensiveness of the prevailing gait assessment systems. High end sensors make the overall system costly, which is not affordable for most clinics, especially in a developing country. High expenditure causes limited expansion of gait laboratories across the world. Compulsorily pathologists follow the error-prone qualitative techniques to assess gait. Hence, a low-cost arrangement to analyze gait is hugely needed. This thesis has proposed four contributions to address some challenging problems of human gait in the clinical domain using some low-cost system setups. Few vital issues,like gait event detection, abnormality detection, feature assessment, etc., were explored and investigated. Basically, this thesis targets to construct a few affordable automatic gait abnormality detection systems after addressing some pre-requisite issues. Validation of a sensor before using it for clinical purposes is an important issue. Studies reported that the skeletal data stream of Kinect is not suitable to estimate joint kinematics. Although the joint angle time series follows the pattern of the corresponding ground truth, it differs substantially in terms of magnitude. Hence, as an alternative, the color image data stream of Kinect was investigated for joint kinematics in the first contribution. The point cloud feature of Kinect was used to extract lower limb joint positions, which were then converted to joint angles using extended Kalman filter and a kinematic model. The process was validated against the gold standard cameras. The obtained joint angles were significant to use for medical purposes. Event annotation is considered as an initial work for constructing a gait abnormality detection system. Most of the clinics follow the manual annotation technique of gait events on a time series data. However, this method is error-prone and laborious. On the contrary, automatic gait event detection systems are gradually becoming popular. This thesis proposes an event detection system using a state-space model. A multi-Kinect architecture for overground walking was established. Data were collected from both pathological and normal populations. A state-space model was constructed where the temporal evolution of gait signal was modeled by quantifying feature uncertainty. The inter-state transition frames were marked as the gait events. In addition, an attempt was made for treadmill gait also. Here, an unsupervised approach was proposed to detect gait events using a multi-Kinect system. Cerebral Palsy is a widespread disease across the world. The activity of daily life of patients suffers from distorted gait. Numerous features have been extracted to characterize the gait pattern of this population. However, there exists a high variability in the recommended features. Prior information on the most important gait feature would help to construct a population-specific gait abnormality detection system. Hence, a well-known statistical approach called meta-analysis, which is generally used in medical science to estimate the effect of an intervention, has been used to select the most important gait features in Cerebral Palsy population. Features were ranked according to their importance level. Automatic abnormality detection systems, specifically for the Cerebral Palsy patients, are expensive. On the other hand, systems based on a low-cost sensor, like Kinect, suffer from several problems. This thesis addresses some of those issues. A clinically relevant walking track was constructed using a multi-Kinect architecture. An algorithm to remove outliers from the multi-Kinect data has been proposed. Features, generally used to detect the Cerebral Palsy gait, are influenced by the gait velocity. A speed-invariant feature might be beneficial for such systems. Hence, this thesis used a handcrafted speed invariant feature and compared its performance against the best feature set. Different supervised models were established to construct the abnormality detection systems.