Classification of hemiplegia by machine learning interpretation of EMG signals

Gait pattern recognition has a fundamental role in the process of diagnosis and clinical decision-making. Recent approaches have addressed this issue by applying machine learning techniques to achieve a rapid, reliable and objective interpretation of the data collected during gait analysis session. In the study have been proposed two machine learning approaches for surface electromyographic (sEMG)-based classification of the typical walking pattern of able-bodied children, Winters’ type I and Winters’ type II hemiplegic children. sEMG signals were taken from retrospective studies performed at Laboratory of Gait Analysis, Mocalieri (TO). The database consists of gait data related to 30 hemiplegic children (13 Winters’ type I and 17 Winters’ type II children) and 30 able-bodied children. The first approach proposed showed an average classification accuracy of 98.6% for learned subjects and 79.2% for unlearned ones in the discrimination of the three neuromuscular states considered. The second one showed an average classification accuracy of 93.8% and 88.9% in the discrimination between control and hemiplegic children in unlearned and learned conditions, respectively. In the recognition of the two forms of hemiplegia it reported an average classification accuracy of 94.4% for learned subjects and 75.0% for unlearned ones. Finally, for what concerns the discrimination between healthy, Winters type I and Winters type II subjects, it achieved and an average classification accuracy of 90.5% for learned subjects and 79.2% for unlearned ones. These promising performances highlight the reliability of machine learning predictions in the discrimination of different neuromuscular conditions. Moreover, these findings prove a correlation between the Winters’ classification, based on the observation of kinematic data, and muscle recruitment during walking.