Ensemble transformer-based framework for predicting finger angles

Abstract More than 50% of individuals with upper-limb amputation reject the use of prosthetic devices, primarily due to limitations in functionality, comfort, and intuitive control. Accurate estimation of finger joint angles is a key requirement for proportional prosthetic control, as it enables the generation of continuous and natural hand movements that more closely mimic biological hand function. However, real-time prediction of finger joint angles remains challenging because of the nonlinear relationship between electromyographic signals and joint motion, as well as the computational constraints associated with embedded prosthetic systems. In addition, electromyographic signal quality is affected by biological factors such as skin properties, bone structure, tendons, and muscle variability. In this study, multimodal electromyographic and inertial measurement unit signals collected from able-body subject were used to predict metacarpophalangeal joint angles. Features including variance, mean, median, and muscle synergies were extracted, and a sliding-window technique was applied to improve training stability and output smoothness. Dimensionality reduction was performed using the minimum redundancy maximum relevance method, and a hybrid convolutional neural network –Transformer architecture was employed. To capture similarities and differences across finger motion ranges, the hand was divided into three functional regions using an ensemble learning strategy. The proposed model achieved a coefficient of determination of 0.90 and a root mean square error of 7.92° during testing using a window size of 150 samples, an overlap of 25 samples, and 18 selected features. The proposed framework has the potential for use in applications related to proportional prosthetic control.