Monitoring functional motor activities in patients with stroke.
This manuscript describes the development and results of a series of studies using an automated procedure to identify functional motor activities (FMAs) in a target population of patients with hemiparesis from stroke. The procedure uses artificial intelligence algorithms to analyze surface electromyographic (EMG) and/or accelerometric (ACC) signals recorded during a set of activities used to evaluate functional independence. Eleven and ten patients with hemiparesis from stroke participated in the two phases of studies, respectively. EMG and ACC data (8 channels each) were recorded while the subjects carried out a set of 11 "identification" FMAS derived from the functional independence measure (FIM) scale and 10 "non- identification" FMAs to evaluate misclassification errors. A multilayer feedforward neural network and an adaptive neuro-fuzzy inference system were used to automatically identify the identification FMAs while minimizing misclassification errors of the non-identification FMAs. The neural network and adaptive neuro-fuzzy inference system resulted in a fair classification of the FMAs with a high sensitivity (80%) and specificity (98%) but with a relative high (40%) misclassification rate. The system achieved a successful classification of the FMAS with a high sensitivity (91.18%) and specificity (98.77%), for misclassification error rates arbitrarily set at a 10% while using all 16 channels of information. A 4-hybrid (4 ACC and 4 EMG) sensor configuration yielded a sensitivity of 94.73% and a specificity of 99.32% for the same misclassification threshold. The results demonstrate the feasibility of using a combined EMG and ACC sensor system to automatically monitor functional motor activities in patients with hemiparesis from stroke.