The potential role of mobility assistive devices (e.g., canes, walkers) to promote functional independence in the home and community is significant in view of the rapid increase of the number of elderly individuals with diminished capacities and multiple chronic disabilities. Among all mobility assistive devices, canes are the most used by individuals at risk for falling. In addition to improving balance, canes can prevent injuries related to falls and negative behaviors associated with fear of falling (e.g., decreased mobility). Once taken home, these mobility assistive devices are commonly misused or not used at all. A means for assessing the use of a cane in the home setting is needed to aid clinicians in the prescription of such devices.
In this study, we studied the use of wearable sensors to monitor the quality of use of a cane. Wearable sensors were used to identify the motor task (e.g., level walking, walking on an incline, stair climbing) that a subject was performing, whereas sensors on the cane were used to evaluate the use of the cane in the context of the identified task. Fifteen subjects that used a mobility assistive device for arthritis affecting the knee or hip were recruited in the study. A triaxial accelerometer and a single axis gyroscope were integrated into a small circuit board that was attached bilaterally to the subjects’ ankles and wrists. Analog signals (n = 16) from the sensors were digitized and sampled on a PDA at a rate of 100 Hz. A single-point cane was equipped with a load cell and two accelerometers parallel to the ground. Analog signals from these sensors were sampled and wirelessly transmitted using a dedicated RF transceiver. Data from the PDA were transmitted to a PC through an 802.11b TCP/IP link and synchronized with the signals from the cane.
Each subject performed the following set of motor tasks: level walking, walking carrying an object, walking on an uneven surface, walking up a ramp, walking down a ramp, walking up a flight of stairs, walking down a flight of stairs, walking over an object, pivoting, and opening a door. The raw data from the wearable sensors were processed and features were derived. Such features were fed into a neural network that was trained to identify the motor tasks of interest. The neural network was trained with an output neuron for each task, with only the output neuron of the correct task being activated. A threshold was used to determine whether the output pattern matched closely enough to a target result to be counted.
Results demonstrated that the motor tasks of interest could be reliably identified. For average sensitivity equal to 95%, specificity was greater than 95%. The distribution of load values and features derived from the sensors on the cane indicated that the proposed technique was highly sensitive to differences in quality of use across individuals and differences in the dynamics of loading the cane across motor tasks.