Classification of Transhumeral Movements Using Deformable CNN With Magnitude-Based STFT From Tripartite EMG Sensor Data

Transhumeral movement classification via Electromyographic (EMG) signals is a critical step forward in the quest to develop advanced prosthetic limbs that closely mimic natural arm movements, thereby significantly enhancing the quality of life for amputees. This paper introduces a novel framework that utilizes Deformable Convolutional Neural Networks (DCNN) in conjunction with a magnitude-based Short-Time Fourier Transform (STFT) approach. This method uses three EMG sensors to precisely examine muscle movements in the medial deltoid, biceps, and triceps. It is a notable improvement compared to conventional methods. DCNN offer an unusual ability to handle intricate data patterns, such as shifts, scales, and distortions. This sets them apart from other Convolutional Neural Networks (CNN), particularly in tasks that demand precise feature extraction. In order to enhance the functionality and user experience of prosthetics, we conduct an experimental study involving 20 healthy participants. The study focuses on six fundamental arm movements: flexion, extension, abduction, adduction, pronation, and supination. The outcomes are encouraging, as the DCNN-based classifier had a remarkable accuracy of
$82.03\pm 0.44$
% and F1-score
$81.82\pm 0.46$
%, overcoming the standard CNN benchmark of
$79.47\pm 1.49$
% and F1-score
$79.27\pm 1.49$
%. The results emphasize the considerable potential of DCNN in advancing prosthetic limb technology, leading to the development of more intuitive and efficient prosthetic solutions.