Machine Learning for Automating Subjective Assessment of Arm Movement Abnormality After Acquired Brain Injury

Neurology Information Science Artificial Intelligence Medicine Orthopedics Bioengineering Medical Imaging Life Sciences
machine learning acquired brain injury gait assessment movement analysis neurorehabilitation

Automated Arm Movement Anomaly Assessment Based on Machine Learning

Automated Clinical Assessment of Abnormal Walking Movements in ABI Patients Through Image Extraction and Classification Systems

Academic Background

Walking disability is a common physical impairment following Acquired Brain Injury (ABI). ABI typically includes stroke and traumatic brain injury, with a global incidence of approximately 1.5 million cases. Walking disabilities in ABI patients affect not only the lower limbs but also the trunk and upper limbs, limiting participation in daily activities and significantly reducing quality of life. In addition to functional impairments, these apparent movement abnormalities may also cause aesthetic issues, negatively impacting patients’ body image, self-esteem, mental health, and social integration.

Research Motivation

Traditional assessment of movement abnormalities in ABI patients usually relies on subjective evaluation through visual observation by experienced physiotherapists. However, the International Classification of Functioning, Disability and Health (ICF) movement abnormality assessment scale shows strong consistency within the same assessor but only moderate reliability between different assessors, limiting its application in clinical practice. To address this issue, researchers attempted to introduce a two-tier machine learning model to automatically assess upper limb movement abnormalities during walking in ABI patients.

Article Source

This research was jointly authored by Ashleigh Mobbs, Michelle Kahn, Gavin Williams, Benjamin F. Mentiplay, Yong-Hao Pua, and Ross A. Clark. They are from the School of Health and Sport Sciences at the University of the Sunshine Coast, Australia; Department of Physiotherapy at Epworth Healthcare; Department of Physiotherapy at the University of Melbourne; School of Allied Health, Human Services and Sport at La Trobe University; and Department of Physiotherapy at Singapore General Hospital. The paper was published in the Journal of NeuroEngineering and Rehabilitation in 2024.

Research Process

Experimental Design and Methods

Participants: The study included 42 ABI patients and 34 healthy controls (HCs). ABI patients were recruited from brain injury rehabilitation centers and private physiotherapy clinics in Melbourne, with an average age of 48 years and an average time since injury of 6.2 years. HCs were from the research team’s staff, family, and friend networks, with an average age of 37 years.

Video Collection: Participants were asked to walk barefoot on a 10-meter walkway, with dynamic video recording of the frontal plane as they approached the camera. Videos were captured using a Microsoft Kinect v2 camera, with RGB image resolution of 1920×1080 and video length covering a 10.5-meter walking segment.

First-tier Machine Learning: Anatomical Landmark Network Development (DeepLabCut™) The open-source software DeepLabCut™ was used for frame selection and labeling, with the network trained for 500,000 iterations. The network was used to calculate two-dimensional spatial motion angles of bilateral shoulder, elbow, and wrist joints during participants’ walking. Nearly 50%, 75%, 90%, and 100% of ABI participant videos were used for training.

Second-tier Machine Learning: Predictive Algorithm Development (Random Forest Network) Three neurological rehabilitation specialist physiotherapists with >15 years of experience scored participant videos using the ICF scale. These scores, along with joint angle data from the first tier, were used to train a random forest network and perform nested cross-validation. Machine learning model predictions for all videos were compared with clinical assessor scores using quadratic weighted kappa coefficients and single-sample t-tests.

Research Results

The results showed that machine learning predictions performed similarly in terms of consistency with experienced human assessor scoring, with no significant differences. There were no significant differences among the four different networks, although there was slight under-prediction in some scores. This indicates that machine learning models can provide reliable assessment results even with limited sample sizes.

One-way ANOVA: A one-way analysis of variance applied to the four networks (50%, 75%, 90%, 100%) showed no statistically significant differences in predictions between networks (f=0.119, p=0.949).

Research Conclusions

The study demonstrates that machine learning can perform comparably to experienced clinicians in subjectively assessing upper limb movement abnormalities during walking in ABI patients. Although the small sample size may lead to under-prediction of certain scores, the effect size is small, and there are no significant performance differences. Future large-scale studies may help further validate the effectiveness of this method, especially in local and remote rehabilitation assessments, using smartphone and edge computing technologies to reduce measurement errors and inequalities in healthcare access.

Significance of the Research

The main contribution of this research is to demonstrate how machine learning technology can be used to address problems in traditional clinical assessments, such as assessment consistency and technical burden. Machine learning can achieve more efficient and reliable assessment processes, which is significant for practical clinical applications and telemedicine. Additionally, this study points the way for future research, suggesting that increasing sample size and assessing complex dynamic tasks could further improve model prediction accuracy and clinical application value.

Research Highlights

  • Methodological Innovation: The study uses a two-tier machine learning model, including DeepLabCut™ and random forest networks, to automatically assess upper limb movement abnormalities in ABI patients.
  • Practical Application: The model demonstrates high efficiency and accuracy on a relatively small dataset and is not affected by changes in assessors.
  • Future Potential: This research provides a theoretical and practical basis for promoting the use of machine learning technology in actual clinical and remote assessments, with broad application prospects.

Other Important Information

  • Data Openness and Privacy Protection: While raw data cannot be made public to protect participant privacy, anonymized data can be provided if needed.
  • Ethical Approval: The study received ethical approval from Epworth Healthcare and the University of the Sunshine Coast, and complies with the Declaration of Helsinki.

This study demonstrates the potential of machine learning in clinical assessment through innovative methods and rigorous validation, providing a solid foundation for future applications in broader medical practices.