Feature Extraction Using Autoencoders For Upper Limb Use Detection

Traditionally, data collected from wearable sensors are preprocessed and undergo feature extraction before being fed into a machine learning model. However, manual feature extraction requires expertise and may not work on all data types. Additionally, movement patterns may vary among individuals and activities, which could also affect the performance of models. Using data from 10 healthy individuals from a previous pilot study, we investigated how automatic feature extraction using autoencoders affects upper limb use detection performance. It is observed that as we used multiple autoencoders, the Youden index value of the model has improved compared to previous studies using traditional feature extraction. We believe that using such advanced machine learning models can lead to efficient quantification of upper limb use and can have significant impacts on rehabilitation outcomes using wearable technologies.