Impact of Segmentation Parameters on Pattern Recognition in Amputated Subjects for Prosthetic Hands

The myoelectric prosthetic hand functions based on electromyography (EMG) signals generated by the skeletal muscles, reflecting the user's intentions and their control via pattern recognition strategy increases their dexterity. The pattern recognition process involves three main steps: segmentation, extraction of features, and classification. There are two types of segmentation: overlapping segmentation & disjoint segmentation. Disjoint segmentation splits signal data in distinct segments of a predetermined span. Overlapping segmentation, on the other hand, involves new segments that include a part of the previous segment. Achieving both high classification accuracy and less time to respond is crucial for live continuous classification and the advantage of using overlapping segmentation is that it increases the decision-making process. However, effectiveness of overlapping segmentation is determined by two factors: the length of the segments known as window length (WL) and the increment used known as window increment (WI). These factors would impact the characteristics of the extracted feature set and consequently, the classification accuracy. Surface EMG (sEMG) signals of five trans-radial amputees from Ninapro Database and six commonly used gestures were used in this study. Window lengths ranging between 100 ms~300 ms and window increments of 25,50 and 75% were examined in this study. Four time-domain features were extracted: Zero Crossing (ZC), Waveform Length (WL), Slope Sign Change (SSC), and Mean Absolute Value (MAV). K-Nearest Neighbor and Random Forest Classifier were used for classification. The investigation revealed that classification accuracy increases with an increase in window length, and while maintaining a constant window length, classification accuracy decreases with an increase in window increment. As a result of these findings, it is recommended to maintain the length and increment of the window at 200 ms and 25%, respectively, to optimize classification accuracy.