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

Gait analysis commonly uses Inertial measurement unit (IMU) sensors because of their mobility, reliability, and applicability for real-time applications. Current research on gait phase identification mostly concentrates on enhancing classification accuracy through the use of many sensors or intricate algorithms, but often overlooks challenges such as sensor drift and recalibration in wearable medical settings. To perform real-time gait phase classification and identify irregularities linked to knee osteoarthritis (KOA), this work proposes a wearable sensor-based system that uses dual IMU sensors mounted on the femur and tibia. Walking, stair climbing, and sit-to-stand exercises were observed in a total of 15 participants, including both healthy people and KOA patients, resulting in 637 gait cycle samples. MPU9250 sensors were used to gather IMU data at 100 Hz, and min-max normalization and outlier elimination were used during processing. Real-time gait features (orientation, angular velocity, and acceleration) were extracted and used to train eight machine learning classifiers. Out of the above models being analysed, ensemble classifiers exhibited high performance. The classification accuracy of Random Forest was 97 %, that of Gb and Stacking was also 96 %. Statistical analysis using the Friedman test (χ² = 18.9, p < 0.01) and post-hoc Nemenyi comparisons confirmed Random Forest's significant advantage. After training, the models were paired with TinyML-ready hardware to ensure gait phase classification operates reliably and efficiently. Results demonstrate the feasibility of affordable, real-time wearable IMU devices for accurate gait monitoring and early KOA detection. The proposed method has immense potential for applications in online surveillance, clinical rehabilitation, and personalized mobility assessments.