Ambulation Prediction by Fusing Raw IMU Data and Novel sEMG Time-frequency Features

This paper proposed and implemented an experiment to compare the accuracy rate, sensitivity, precision, F1-score, and stability of different feature-level fusion prediction models predicting five lower limb activities of daily living (ADL) in data from 18 healthy human subjects. The results of this work show that a high-level feature fusion CNN model with the time-frequency features of sEMG and raw IMU data has produced significantly stable and more accurately predicted results than SVM models with other time-domain and frequency-domain features, especially the multivariate synchrosqueezing transform (MSST) features, which have maximum accuracy and stability increments of 1.23% and 3.08%, respectively, without more manual feature encoding. It is concluded that the results demonstrate the validity and feasibility of the MSST for extracting muscle activity features and integrating features in a more advanced and novel way to ensure the prediction of daily lower limb ambulation activities.