Evaluating generalization of arm movement identification using machine learning: From structured to semi-structured...

Area of Science:

• Biomedical Engineering
• Rehabilitation Technology
• Machine Learning in Healthcare

Background:

• Home-based rehabilitation is crucial for stroke survivors' motor recovery and daily activities.
• Wearable technology and machine learning offer potential for advanced home-based arm rehabilitation.
• Current machine learning models often lack testing across diverse environments, limiting real-world application.

Purpose of the Study:

• To evaluate the generalization ability of machine learning models for arm movement identification.
• To compare model performance across structured (lab) and semi-structured (kitchen) environments.
• To assess the impact of sensor configuration (multi-IMU vs. single wrist IMU) on generalization.

Main Methods:

• Investigated two machine learning models: Random Forest and a hybrid deep learning model.
• Trained models in a structured lab environment and tested in a semi-structured kitchen environment.
• Compared performance using four arm-mounted IMUs versus a single wrist-mounted IMU.

Main Results:

• Both models demonstrated good generalization from lab to kitchen environments.
• The four-IMU configuration yielded higher accuracy than the single wrist IMU.
• The Random Forest model showed a smaller accuracy decrease with the single wrist IMU compared to the hybrid model.

Conclusions:

• Arm movement identification algorithms generalize well across different environments, even with minimal sensors.
• These findings support the potential of wearable technology for practical home-based stroke rehabilitation.
• Further development can leverage these algorithms for more accessible and effective remote rehabilitation solutions.