A Two-Stage Tremor Detection Algorithm for Wearable Inertial Sensors During Normal Daily Activities.

Continuous monitoring of tremor with wearable wrist sensors during normal daily activities is more difficult than in a clinical setting when subjects perform prescribed activities because some normal daily activities resemble tremor, many normal movements contain frequency content that overlaps with the tremor frequency, and the tremor amplitude has a large dynamic range during normal daily activities. We describe a novel two-stage algorithm that offers improvement at discriminating tremor from other activities. Some of this improvement is attained by using prior domain knowledge that tremor occurs over a narrow range of frequencies for an individual, but the mean tremor frequency may vary significantly between individuals in a study population. We validated the algorithm in continuous recordings from people with Parkinson's disease and matched control subjects. The algorithm has good face validity, a low rate of false positives on recordings from control subjects (<; 1.1%), and good correspondence with the constancy of rest tremor as measured by this question on the MDS-UPDRS (ρ = 0.54).