Classification of Stair Ascent and Descent in Stroke Patients

Abstract

Wearable sensor units are a promising technology to assess ambulatory activities such as level walking, stair ascent and descent in the home environment, shedding light into the recovery process and independence of stroke survivors. However, algorithms for the identification of ambulatory activities were optimized for healthy subjects, and show limitations when considering the reduced walking speed and altered gait patterns found in patients. We present a method to identify ambulatory phases and distinguish stair ascent and descent from level walking in daily activity recordings of stroke survivors. A realistic dataset was captured with inertial and barometric pressure sensors worn at 5 anatomical locations. Statistical and wavelet based acceleration features fed into a Support Vector Machine were used to identify walking phases, while a k-Nearest-Neighbor classifier was used to discriminate between level walking, stair ascent and descent based on barometric pressure and acceleration features. Combining data from multiple sensor modules resulted in walking classification sensitivities and specificities of up to 96%. Looking at sensor modules individually, the module placed at the nonparetic ankle showed the best performance, increasing sensitivity of walking identification by almost 10% compared to the module at the paretic ankle. Level walking was identified with 97% sensitivity and 91% specificity, stair ascent with 94% sensitivity and 99% specificity and stair descent with 87% sensitivity and 99% specificity in the multi-sensor setup. Again, sensor modules placed at the ankles displayed the best performance when looking at modules individually.