Cognitive Workload Classification of Upper-limb Prosthetic Devices

Limb amputation can cause severe functional disability in performing activities of daily living (ADLs). Using prosthetic devices as aids for such activities requires substantial cognitive resources. Machine Learning (ML) algorithms can be used to predict cognitive workload (CW) of prosthetic device prototypes early in the design process and serve as a tool for improving device usability. The objective of this study was to explore subsets of input features that can be easily captured during early stages of the design cycle to classify CW of electromyography (EMG)-based upper-limb prostheses. An experiment was conducted with 30 participants to collect task performance and pupillometry data, and to provide a basis for generating cognitive performance model (CPM) outcomes. Three ML algorithms, including the random forest (RF), support vector machine (SVM), and naive Bayesian (NB) classifier were developed. The most important subset of features was selected based on classification accuracy and computational and experimental cost. Findings revealed that the CPM outcomes and prosthetic device configuration were the most important features for reasonably classifying CW responses under low cost. Also, the SVM classifier can be used for near-real time classification of CW. Future studies should include additional data and improve hyperparameter tuning parameters, as well as advanced CPM techniques to improve the performance of algorithms.