Exploring unsupervised feature extraction of IMU-based gait data in stroke rehabilitation using a variational...

Area of Science:

• Biomedical Engineering
• Rehabilitation Science
• Machine Learning in Healthcare

Background:

• Inertial Measurement Units (IMUs) offer rich gait data but require dimensionality reduction.
• Variational Autoencoders (VAEs) can potentially extract salient information from complex sensor data.
• Assessing the psychometric properties of VAE-derived gait features is crucial for clinical translation.

Purpose of the Study:

• To investigate the efficacy of VAEs in reducing IMU-based gait data from post-stroke individuals into a low-dimensional latent space.
• To evaluate the reliability, group differentiation, and responsiveness of these latent features compared to traditional gait speed measures.

Main Methods:

• Collected 2-minute walk test data using IMUs from post-stroke patients and healthy controls.
• Applied VAEs to segment IMU data into 12 latent features.
• Assessed test-retest reliability (ICC), inter-group differences (t-test), and rehabilitation responsiveness.

Main Results:

• VAE successfully reconstructed gait data with minimal error.
• Five latent features demonstrated good-to-excellent test-retest reliability.
• Seven latent features significantly differentiated between stroke survivors and controls, though responsiveness analysis showed ambiguity.

Conclusions:

• VAEs provide an effective method for reducing IMU gait data into a concise set of reliable latent features.
• Certain latent features show promise for distinguishing between healthy individuals and stroke survivors.
• Further research is necessary to validate the clinical utility and interpretability of these VAE-derived gait biomarkers.