TinyML-enabled wearable system for early detection of knee osteoarthritis using ensemble gait classification

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Machine Learning in Healthcare

Background:

• Gait analysis commonly employs Inertial Measurement Unit (IMU) sensors for their portability and real-time capabilities.
• Existing research often prioritizes classification accuracy over practical challenges like sensor drift in wearable systems.
• Early detection of knee osteoarthritis (KOA) requires reliable gait monitoring methods.

Purpose of the Study:

• To develop and validate a wearable sensor-based system for real-time gait phase classification.
• To identify gait irregularities associated with knee osteoarthritis (KOA).
• To evaluate the performance of machine learning classifiers for this application.

Main Methods:

• Utilized dual IMU sensors (MPU9250) placed on the femur and tibia, collecting data at 100 Hz.
• Processed IMU data using min-max normalization and outlier elimination for feature extraction (orientation, angular velocity, acceleration).
• Trained and evaluated eight machine learning classifiers, focusing on ensemble methods, and deployed models on TinyML-ready hardware.

Main Results:

• Ensemble classifiers demonstrated high performance, with Random Forest achieving 97% accuracy, and Gradient Boosting (Gb) and Stacking achieving 96%.
• Statistical analysis confirmed Random Forest's significant advantage (Friedman test, p < 0.01).
• The system achieved reliable and efficient real-time gait phase classification on resource-constrained hardware.

Conclusions:

• The proposed dual IMU system offers a feasible, affordable solution for accurate gait monitoring and early KOA detection.
• This technology holds significant potential for online surveillance, clinical rehabilitation, and personalized mobility assessments.
• Real-time gait phase classification using wearable IMUs can be effectively implemented for medical diagnostics.