Inertial Measurement Unit-Based Frozen Shoulder Identification from Daily Shoulder Tasks Using Machine Learning...

Area of Science:

• Biomedical Engineering
• Rehabilitation Technology
• Machine Learning in Healthcare

Background:

• Frozen shoulder (FS) causes shoulder pain and limited range of motion (ROM), with current assessments being subjective.
• Objective evaluation methods are needed to improve the accuracy of FS diagnosis and monitoring.
• Inertial Measurement Units (IMUs) offer a potential solution for objective motion analysis.

Purpose of the Study:

• To develop and evaluate an IMU-based system for automatically identifying shoulder tasks in individuals with and without FS.
• To compare the effectiveness of different machine learning (ML) and deep learning (DL) models for FS identification.
• To analyze the impact of feature types and sensor placement on identification performance.

Main Methods:

• Utilized two IMUs attached to the arm and wrist of 24 FS patients and 20 healthy subjects performing five daily shoulder tasks.
• Extracted and analyzed two feature groups: time-domain statistical features and kinematic features.
• Applied seven ML techniques and two DL models (Convolutional Neural Network, Multilayer Perceptron) for task identification.

Main Results:

• The DL-based system achieved the highest identification performance, with the CNN reaching 88.26% accuracy and MLP achieving an 89.23% F1 score.
• Wrist-based IMU features yielded higher accuracy than arm-based features.
• Time-domain statistical features demonstrated better discriminability for FS identification compared to kinematic features.

Conclusions:

• An IMU-based system using ML is feasible for objective FS assessment in clinical practice.
• Deep learning models, particularly CNNs, show significant potential for accurate FS task identification.
• Optimizing sensor placement (wrist) and feature selection (time-domain) can enhance system performance.