An Unsupervised Learning Approach for Multimodal Low Back Pain Stratification.

Area of Science:

• Biomedical research
• Radiology
• Data science

Background:

• Current low back pain (LBP) risk stratification uses tools like the Örebro Musculoskeletal Pain Screening Questionnaire (ÖMPSQ) short and STarT Back Tool (SBT).
• These tools capture psychosocial factors but LBP's complexity requires a more comprehensive approach for effective risk stratification.

Purpose of the Study:

• To propose a novel stratification framework for individuals with low back pain (LBP).
• To integrate deep learning (DL)-based imaging biomarkers with non-imaging data for a multimodal approach.
• To validate the framework's utility and demonstrate superior net benefit over a
• treat-all
• strategy.

Main Methods:

• Developed a multimodal unsupervised patient stratification method integrating imaging biomarkers (disc degeneration, facet tropism) from DL-analyzed lumbar spine MRI with non-imaging data (smoking, demographics, ÖMPSQ short, SBT).
• Utilized K-Means clustering for stratification, characterizing clusters by LBP frequency and bothersomeness.
• Validated cluster robustness with multi-class logistic regression and assessed net benefit using decision curve analysis.

Main Results:

• Identified three distinct subgroups characterized by LBP frequency and bothersomeness.
• Subgroups were dominated by psychosocial characteristics, physical degenerative changes (disc degeneration), or a mix of both.
• Predictive models for cluster assignment showed high accuracy (SBT-based: 0.89; ÖMPSQ-short-based: 0.87), with superior net benefit compared to a "treat-all" strategy.

Conclusions:

• A novel framework integrating multimodal data was developed to identify distinct subgroups in a population-based cohort.
• These subgroups are differentiated by physical and psychosocial characteristics.
• The framework shows potential for advancing personalized care in low back pain management.