Decoding Attempted Hand Movements in Stroke Patients Using Surface Electromyography

Abstract

Brain- and muscle-triggered exoskeletons have been proposed as a means for motor training after a stroke. With the possibility of performing different movement types with an exoskeleton, it is possible to introduce task variability in training. It is difficult to decode different movement types simultaneously from brain activity, but it may be possible from residual muscle activity that many patients have or quickly regain. This study investigates whether nine different motion classes of the hand and forearm could be decoded from forearm EMG in 15 stroke patients. This study also evaluates the test-retest reliability of a classical, but simple, classifier (linear discriminant analysis) and advanced, but more computationally intensive, classifiers (autoencoders and convolutional neural networks). Moreover, the association between the level of motor impairment and classification accuracy was tested. Three channels of surface EMG were recorded during the following motion classes: Hand Close, Hand Open, Wrist Extension, Wrist Flexion, Supination, Pronation, Lateral Grasp, Pinch Grasp, and Rest. Six repetitions of each motion class were performed on two different days. Hudgins time-domain features were extracted and classified using linear discriminant analysis and autoencoders, and raw EMG was classified with convolutional neural networks. On average, 79 ± 12% and 80 ± 12% (autoencoders) of the movements were correctly classified for days 1 and 2, respectively, with an intraclass correlation coefficient of 0.88. No association was found between the level of motor impairment and classification accuracy (Spearman correlation: 0.24). It was shown that nine motion classes could be decoded from residual EMG, with autoencoders being the best classification approach, and that the results were reliable across days; this may have implications for the development of EMG-controlled exoskeletons for training in the patient's home.

Keywords: EMG; brain-computer interface; myoelectric control; pattern recognition; stroke.

Figure 1

Rectified (only for visualization) and bandpass filtered surface EMG for the nine different motion classes for a single repetition and a single participant. Hand Close (HC), Hand Open (HO), Wrist Flexion (WF), Wrist Extension (WE), Supination (Sup), Pronation (Pro), Lateral Grasp (Lat), and Pin (Pinch Grasp). Flexor (Fl.), Extensor (Ex.). Clear EMG activity can be seen for most motion classes except the Lateral Grasp.

Figure 2

Overall classification accuracy for all motion types. The results are presented as mean ± standard deviation across participants. “Day12” indicates training on data from day 1 and testing on data from day 2. “Day21” indicates training on data from day 2 and testing on data from day 1. LDA (linear discriminant analysis), AE (autoencoders), and CNN (convolutional neural network).

Figure 3

Rectified (only for visualization) and bandpass filtered surface EMG for the Hand Open motion class for the subject with the highest (subject 4) and lowest (subject 7) classification accuracy. The highest and lowest overall classification accuracies were 91% and 54% (classified with linear discriminant analysis), respectively. The amplitude of the EMG for the motions performed by the best subject is higher compared to the worst subject. Moreover, there is a smaller EMG amplitude for the resting state between the movements for the best subject.