A machine learning based approach for quantifying muscle spasticity level in neurological disorder patients

Abstract

Muscle spasticity is a condition that occurs in patients with neurological disorders when their muscles are stiff, tight, and resistant to stretching. The current assessment method, relying on the subjective judgment of therapists using the Modified Ashworth Scale (MAS), introduces variability that may affect the rehabilitation process. In addition, many existing computational models are not aligned with clinical standards such as MAS, limiting their practical adoption in clinical settings. They also often overlook direction-specific movement phases and fail to distinguish muscle responses across different axes and muscle groups. To address the limitations, this research aimed to develop a quantitative assessment method by muscle spasticity characteristics based on mechanomyography (MMG) signals and MAS levels, utilising machine learning techniques. A Quantitative Spasticity Assessment Technology (QSAT) platform has been developed which consists of two sensors that were tri-axial accelerometer mechanomyography (ACC-MMG) functioning in measure acceleration of biceps and triceps muscle contraction and potentiometer to measure the angular position of forearm during flexion and extension movement. A comprehensive investigation was conducted to assess muscle spasticity level by recording ACC-MMG signals from patients' forearm musculature during flexion and extension movements using QSAT platform. A total of 30 patients with neurological disorders were classified into five MAS levels (0, 1, 1+, 2, and 3), along with 10 healthy subjects serving as a baseline group. The pre-processed data comprised 48 extracted features from ACC-MMG signals along the x, y, and z axes for both flexion and extension movements of the biceps and triceps. These features corresponded to the longitudinal, lateral, and transverse muscle orientations. For both flexion and extension movements, machine learning models were trained using the selected subset of 25 significant features and the full set of 48 features respectively, with performance comparisons made to identify the most effective approach. Various machine learning models algorithms, including Linear Discriminant Analysis (LDA), Decision Tree (DT), Support Vector Machine (SVM), and K-Nearest Neighbour (KNN), were tested. The KNN-based classifier demonstrated the highest performance using a 90% training and 10% testing data split surpassing the performance of other classifiers. Specifically, at k = 15 using Euclidean distance, the KNN achieved an accuracy of 91.29% for flexion using the significant features, with corresponding precision, recall, and F1-score of 91.64%, 91.25%, and 91.47%, respectively. For extension, the same configuration resulted with 96.30% for extension using the full feature set, with precision, recall, and F1-score of 96.53%, 96.30%, and 96.33%, respectively. These results indicate high classification performance, with minimal false positives or false negatives, particularly in distinguishing between different MAS levels. This research suggests that the muscle characteristic model embedded in the QSAT can serve as a standardised and objective assessment tool for measuring the spasticity level of the affected limb using computational method, leading to support clinical evaluations and enabling more effective rehabilitation strategies.