Enhancing rehabilitation in stroke survivors: a deep learning approach to access upper extremity movement using accelerometry data

TL;DR

This study uses deep learning to analyze wrist sensor data and classify upper limb movements in stroke survivors, offering a more accurate way to assess rehabilitation progress in real-life settings.

Contribution

A novel deep learning approach using CNNs and Dense layers with raw accelerometry data to classify functional movements in stroke survivors.

Findings

The intrasubject model achieved an average accuracy of 0.90 for classifying paretic upper extremity movements.

Incorporating non-paretic arm data improved the intersubject model's accuracy to 0.88.

The method outperformed previous approaches by eliminating manual feature extraction and using raw data.

Abstract

Upper Extremity (UE) rehabilitation is crucial for stroke survivors, aiming to improve the use of the paretic UE in everyday activities. However, assessing the effectiveness of these treatments is challenging due to a lack of objective measurement tools. Traditional methods, such as clinician-rated motor ability or patient self-reports, often fail to measure UE performance in real-life settings accurately. Evidence suggests that currently used clinical assessments do not reliably capture actual UE use at home or in the community. This study investigates the application of Convolutional Neural Networks (CNNs) combined with Dense layers using accelerometry data from wrist-worn sensors to classify functional and non-functional UE movements of stroke survivors. Two types of models were developed: one trained on data from individual subjects (intrasubject model) and another trained on data across all subjects (intersubject model). The intrasubject model for the paretic UE achieved an average accuracy of 0.90 ± 0.05, while the intersubject model reached an accuracy of 0.79 ± 0.06. When incorporating signals from the non-paretic arm, the intersubject model’s accuracy improves to 0.88 ± 0.10. Notably, this method utilized raw accelerometry data, eliminating the need for manual feature extraction, which is commonly required in traditional machine learning, and yielded higher accuracy than previously reported methods. This proposed deep learning approach incorporates CNNs with Dense layers, offering a cost-effective and adaptable method for monitoring UE functionality in real-world settings. The results from this study have the potential to inform the development of personalized rehabilitation strategies for stroke survivors, offering valuable insights for clinical practice.

Figures