Introduction: Low back pain is prevalent and severe, yet difficult to treat.  Chronic low back pain (cLBP) – is particularly challenging, in that both the source of the pain, as well as its severity may be multifactorial and patient-specific, depending on various aspects of the patient’s psychological and social status, biological factors, and previous injury history [2].  One promising area of research is the use of changes in patient motion patterns to biomechanically phenotype different sources of low back pain.  To this end, previous studies have attempted to differentiate between subjects with and without cLBP through quantitative analyses of spinal motion with promising results.  The eventual goal of this study is to apply the same process to differentiate more specific motion characteristics, likely at the segmental level, and then to objectively track treatment progress by means of return to healthy biomechanical function.  The purpose of this research is to continue the analysis of healthy and symptomatic spinal motion by including additional features of interest and a broader range of functional movements to determine which features statistically differ between healthy and cLBP subjects and how well those features detect cLBP.

Methods: Data were collected from 25 subjects (10 healthy and 15 with cLBP) as they performed a variety of single planar and functional movements with inertial measurement units (IMUs).  The IMUs were placed on the sacrum, L4 spinous process, and C7 spinous process to record the subjects’ angular velocity, acceleration, and jerk (the rate of change of acceleration) at each of these spinal levels.  The motion features from each level for each movement were grouped by health status (healthy and cLBP) and tested for statistical significance differences between group means using t-tests analyses (a p-value of 0.05 or less was considered statically significant).  Additionally, a classification machine learning model was developed using a k-nearest neighbor algorithm to predict subject health status.

Results: The t-tests revealed statistically significant differences between the healthy and cLBP motion features during several functions, especially for the L4 spinous process velocity and jerk.  Figure 1 shows example data for a single activity (Flexion+Left Axial Rotation). The machine-learning model was able to predict subject health with 76% accuracy.  The model was evaluated with a 10-fold cross validation using an optimized k-value of 5.  Additional subject data will be useful for further model refinement and testing.

Discussion: Our results indicate that spinal kinematic information, specifically velocity and jerk during functional movements, clearly differentiates between healthy and cLBP subjects.  We hypothesize that with a higher sample size and more detailed phenotyping information, kinematic information may also be used to identify specific sources of pain and treatment paradigms for cLBP patients.

Acknowledgements: This research was supported by NIH grant UH3AR076723.