An automated algorithmic approach for activity recognition and step detection in the presence of functional compromise

Wearable technology is a potential stepping stone towards personalised healthcare. It provides the opportunity to collect objective physical activity data from the users and could enable clinicians to make more informed decisions and hence provide better treatments. Current physical activity monitors generally work well in healthy populations but can be problematic when used in some patient groups with severely abnormal function.
We studied healthy volunteers to assess how different algorithms might perform for those with normal and simulated-pathological conditions. Participants (n=30) were recruited from the University of Leeds to perform nine predifined activities under normal and simulated-pathological conditions using two MOX accelerometers on wrist and ankle (Maastricht Instruments, NL). Condition classification was performed using a Support Vector Machine algorithm. Activity classification was performed with five different Machine Learning algorithms: Support Vector Machine, k-Nearest Neighbour, Random Forest, Multilayer Perceptron, and Naive Bayes. A step count algorithm was developed based on pattern recognition approach, using two main techniques, Dynamic Time Warping and Dynamic Time Warping-Barycentre Averaging. Finally, synthetic acceleration signal was generated that represented walking activities since there was limited access to patient data and to refine synthetic data generation in this field. Three dynamic coupled equations were used to represent the morphology of the desired signal.
Wrist and ankle locations performed similarly and the wrist location was used for further analysis. Both condition and activity classification algorithms achieved good performance metrics i.e. that the volunteer has been correctly classified in the right condition, and the activities performed have been correctly recognised. Additionally, the novel step count algorithm achieved more accurate results for both conditions in comparison to existing algorithms from the literature. Finally, the signal generation approach seems promising since the normal condition synthetic signals matched closely to their associated original signals.
Algorithms developed for a specific group or even person with functional pathology, using techniques such as Dynamic Time Warping-Barycentre Averaging produce better results than traditional algorithms trained on data from a different group.