A Hierarchical Three-Class Knee Extension Classifier Using EMG+ IMU Signals

Abstract

In this project, a machine learning classifier was built to evaluate knee
extension exercises using both EMG (electromyography) and IMU (inertial
measurement unit) signals. The goal was to identify a reliable classification
algorithm that can ultimately be implemented on wearable devices, allowing patients to perform physical therapy exercises anywhere while the system automatically checks whether each knee movement is correct. Each trial was classified into one of three categories: (1) Correct Extension, (2) Wrong Extension – Low Lift (LL), and (3) Wrong Extension – Not Fully Extended (NF).

A hierarchical two-stage model using LightGBM with probability calibration
and threshold gating was designed, and the data were split by subject before preprocessing to ensure fair evaluation. The clinical significance of these classifications is noteworthy; a Correct Extension indicates proper technique, while LL may suggest limitations in muscle strength or coordination, potentially leading to delayed recovery. The NF classification indicates inadequate range of motion, which could also extend recovery times. By combining muscle activation (EMG) and movement dynamics (IMU), the model achieved strong performance and demonstrates the potential for robust, on-device movement classification to support safe, unsupervised rehabilitation outside the PT clinic.