Time Series Classification of Active Range of Motion in Robotics-Assisted Stroke Rehabilitation
| Advisor | Dr. Russell Jeter (Georgia State University) |
Abstract
Background: Stroke therapy is essential to reduce impairments and improve motor movements. This project builds upon previous work that classified stroke residual severity using session-level summary statistics from robotics-assisted rehabilitation. Here, we advance this approach by leveraging high-resolution, raw time-series kinematics data collected during in-home therapy sessions to classify specific movement patterns.
Objective: Our main objective is to use raw angle and pressure data from Motus Nova robotics rehabilitation technology to develop supervised machine learning models. We aim to autonomously classify the active range of motion (ROM) of users undergoing stroke rehabilitation.
Methods: The dataset consists of over 28 million overlapping 30-second samples (at 1 Hz) from 431 unique users, derived from Angle and Pressure sensor channels. The target variable is active ROM, discretized into 10 classes. We evaluated several machine learning architectures, including classical approaches (XGBoost), deep temporal models (LSTM, 1D-CNN, Transformer, InceptionTime, Temporal Convolutional Networks), and automated feature engineering using TSFresh (TSFresh-RF, TSFresh-XGB, TSFresh-DFFNN).
Results: Across eleven architectures scored on an identical 100,000-window evaluation subsample, the Temporal Convolutional Network attained the highest overall accuracy (43.2%), while the recurrent LSTM retained the strongest class-wise discrimination (balanced accuracy 0.201, AUROC 0.670) despite significant class imbalance. TSFresh embeddings built from the minimal feature set produced the lowest Mean Absolute Error (1.58 bins) but did not reproduce the top-line accuracy advantage we previously observed with the comprehensive feature set.
Conclusions: We demonstrated that objectively measured, high-resolution time series data combined with advanced machine learning methods can be used to classify active range of motion. While models relying on TSFresh features showed the highest accuracy, addressing the underlying class imbalance remains a critical area for future study to enhance individualized stroke rehabilitation.