BackTracker: Machine learning to identify kinematic phenotypes for personalised exercise management in non-specific...

Area of Science:

• Biomedical Engineering
• Rehabilitation Science
• Artificial Intelligence in Healthcare

Background:

• Low back pain (LBP) is a significant global disability, with most cases being non-specific (NSLBP) and lacking clear causes.
• Current clinical guidelines recommend early active management for NSLBP, but individualized exercise prescription is limited by the variability of impairments and the complexity of existing classification systems.
• Timely and personalized care for NSLBP is hindered by lengthy assessment procedures and the need for extensive clinical training in current diagnostic methods.

Purpose of the Study:

• To develop and validate an Artificial Intelligence (AI) model for identifying the two most common motor control impairments (MCIs) in non-specific low back pain (NSLBP): flexion pattern (FP) and extension pattern (EP).
• To utilize spinal silhouettes extracted from movement videos for self-phenotyping and to guide the selection of personalized exercises for NSLBP patients.
• To assess the diagnostic accuracy and robustness of the AI model in classifying NSLBP based on motor control patterns.

Main Methods:

• Ninety NSLBP participants, classified by an expert physiotherapist into FP or EP MCIs, were recruited.
• Participants performed standard forward- and backward-bending tasks, with movements recorded via video in the sagittal plane.
• Pose estimation and instance segmentation techniques were employed to extract motion features and spinal silhouettes, which were then used to train a feedforward neural network.

Main Results:

• The AI model achieved a high diagnostic accuracy of 91.91% for backward-bend movements, surpassing inter-examiner reliability rates for trained physiotherapists.
• The model demonstrated robustness with a mean Area Under the Curve (AUC) of 0.9422.
• Classification accuracy was lower for forward-bend images (86.69%) and combined tasks (86.29%), with patient-reported outcome measures (PROMs) alone yielding only 63.82% accuracy.

Conclusions:

• The AI model reliably differentiates between FP and EP NSLBP subgroups, highlighting its potential to support prompt and personalized rehabilitation strategies.
• Integrating PROMs with motion features did not significantly improve classification accuracy, suggesting limited added value for exercise tailoring when physical impairments are the primary focus.
• AI-driven analysis of movement patterns offers a promising avenue for objective phenotyping and personalized exercise selection in NSLBP management.