BACKGROUND

This article describes the development of an add-on module for wheeled walkers dedicated to sensor-based posture and gait pattern recognition with the goal to develop an everyday aid for fall prevention. The core contribution is a clinical study that compared single gait parameter assessments coming from medical staff to those obtained from an automatic classification algorithm, i. e. the Mahalanobis distance over time series of sensor measurements.

METHODS

The walker-module described here extends an off-the-shelf wheeled walker by two depth cameras that observe the torso, pelvic, region and legs of the user. From the stream of depth images, distance measurements to eight relevant feature points on the body surface (shoulders, iliac crests, upper and lower legs) are combined to time series that describe the individual gait cycles. For automatic classification of gait cycle descriptions 14 safety-relevant gait parameters (gait width, height, length, symmetry, variability; flection of torso, knees (l/r), hips (l/r); position, distance to walker; 2‑value, 5‑value gait patterns [While the two-value gait pattern differentiates a gait cycle into physiological and pathological, the five-value gait pattern distinguishes between antalgic, atactic, paretic, protective, and physiological gait]), single classifier algorithms were trained using machine learning techniques based on the mathematical concept of the Mahalanobis distance (distance of individual gait cycles to class averages and corresponding covariance matrices). For this purpose, training and test datasets were gathered in a clinical setting from 29 subjects. Here, the assessment of gait properties given by medical experts served for the labelling of sensorial gait cycle descriptions of the training and test datasets. In order to evaluate the quality of the automated classification in the add-on module a final comparison between human and automatic gait parameter assessment is given.

RESULTS

The gait assessment conducted by trained medical staff served as a comparator for the machine learning gait assessment and showed a relatively uniform class distribution of gait parameters over the group of probands, e. g. 57% showed an increased and 43% a normal distance to the walker. Of the subjects 51% positioned themselves central to the walker, while 41% took a left deviating, and 8% a right deviating position. A further 12 gait parameters were differentiated and evaluated in 2-5 classes. In the following, single gait cycle descriptions of each subject were assessed by trained classification algorithms. The best automatic classification rates over all subjects were given by the distance to walker (99.4%), and the 2-value gait pattern (99.2%). Gait variability (94.6%) and position to walker (94.2%) showed the poorest classification rates. Over all gait parameters and subjects, 96.9% of all gait cycle descriptions were correctly classified.

DISCUSSION/OUTLOOK: With an average classification rate of 96.9%, the described gait classification approach is well suited for a patient-oriented training correction system that informs the user about false posture during every day walker use. A second application scenario is the use in a clinical setting for objectifying the gait assessment of patients. To reach these ambitious goals requires more future research. It includes the replacement of depth cameras by small size distance sensors (1D Lidar), the design and implementation of a suitable walker-user interface, and the evaluation of the proposed classification algorithm by contrasting it to results of modern deep convolutional neural network output.