Novel video-centric explainable AI models for accurate low back pain classification

Liu, Zebang 2025. Novel video-centric explainable AI models for accurate low back pain classification. PhD Thesis, Cardiff University. Item availability restricted.

Preview	PDF (Thesis) - Accepted Post-Print Version Download (3MB) | Preview
	PDF (Cardiff University Electronic Publication Form) - Supplemental Material Restricted to Repository staff only Download (208kB)

Abstract

Low back pain (LBP) is a leading cause of disability worldwide, yet its diagnosis and treatment remain complex due to its multifactorial nature and the lack of objective biomarkers. This thesis proposes a novel video-centric, explainable artificial intelligence (XAI) framework for the automated classification of non-specific low back pain (NSLBP), with a focus on spinal function phenotyping. The proposed system leverages 2D side-view videos of patients performing standardised trunk flexion (forward bend) and extension (backward bend) tasks to extract motion features using human pose estimation (HPE). These features are then used to classify spinal function into clinically relevant subgroups, such as high vs. low level function, movement impairment (MI) vs. motor control impairment (MCI), and extension pattern (EP)-MCI vs. flexion pattern (FP)-MCI, guided by the clinically accepted multidimensional classification system (MDCS). Three deep learning pipelines were developed: a multimodal architecture integrating patient-reported outcome measures (PROMs) and motion features (for high/low function), an attention-enhanced Dual Attention model combining spatial and temporal cues (for MI/MCI) and a convolutional neural network (CNN)-based model using key posture images (for EP-/FP-MCI). Each model was rigorously evaluated multiple times on a proprietary patient dataset, achieving classification accuracy ranging from 91.91% to 98.75%. Multiple explainability methods, such as gradient-based saliency methods (Grad-CAM), feature attribution techniques, such as Shapley Additive Explanations (SHAP), integrated gradients (IG), and modality-level attention visualisation were used in combination to highlight salient time frames, key anatomical regions, and main motion features (e.g., lumbar flexion angle, velocity). By systematically addressing three key NSLBP classification tasks, this thesis advances the state-of-the-art in automated spinal function assessment. The proposed models demonstrate high accuracy across diverse classification tasks, confirming their generalisability and clinical validity. Unlike previous studies that rely on wearable sensors or static images, this work introduces a video-centric, non-invasive methodology grounded in the MDCS, enabling the nuanced stratification of patient function and movement quality. Moreover, explainable AI techniques are leveraged to provide clinically meaningful insights into model decisions, fostering transparency and aligning predictions with observable kinematic features. These findings lay a foundation for scalable, explainable, and personalised rehabilitation tools that can support clinical reasoning and improve outcomes in NSLBP management. Limitations include a relatively modest sample size and potential sensitivity to variations in video acquisition conditions. Future work should focus on expanding dataset diversity, enhancing robustness to real-world variability, and validating performance in clinical deployment settings.

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Schools > Engineering
Uncontrolled Keywords:	1. Non-specific low back pain (NSLBP) 2. Artificial intelligence 3. Computer vision 4. Explainable AI 5. Deep learning 6. Clinical classification 7. Motion analysis 8. Rehabilitation
Date of First Compliant Deposit:	8 May 2026
Last Modified:	08 May 2026 13:17
URI:	https://orca.cardiff.ac.uk/id/eprint/186818

Actions (repository staff only)

Edit Item