Unsupervised Identification of Population Clusters With Distinct Low Back Pain Profiles.

BACKGROUND CONTEXT: Low back pain (LBP) is a leading cause of disability worldwide, yet population-level stratification of LBP risk remains limited. Unsupervised machine learning offers a data-driven approach to identify latent subgroups with distinct biopsychosocial profiles.

PURPOSE: To identify population groups within a nationally representative survey using machine learning-based clustering algorithms and describe their LBP prevalence and associated factors.

STUDY DESIGN/

SETTING

Cross-sectional secondary analysis of the Chilean National Health Survey 2016-2017 (ENS 2016-2017).

PATIENT SAMPLE: A total of 6,233 individuals aged 15 years and older completed the main forms, and 5,520 of these individuals also completed laboratory measurements.

OUTCOME MEASURES: Weighted means and prevalence of LBP, sociodemographic characteristics, psychosocial factors (depression suspicion), and clinical biomarkers (metabolic, cardiovascular, and musculoskeletal profiles). Feature importance was assessed to determine the primary drivers of cluster assignment and LBP risk.

METHODS

A machine learning pipeline was applied, using the K-prototypes clustering algorithm for mixed-type variables. The optimal number of clusters was determined using the elbow method and silhouette score. Cluster interpretability and feature importance were assessed using SHAP (SHapley Additive explanations) values analysis. Weighted prevalence estimates and descriptive comparisons were calculated across clusters using survey expansion factors. One-way ANOVA and Chi-square tests were conducted to identify differences among clusters (P = .05), and Odds Ratios (ORs) were calculated and adjusted for age and sex.

RESULTS

Three clusters emerged: (A) healthy young adults, (B) middle-aged adults with higher depressive symptom burden, and (C) socioeconomically disadvantaged older adults with multimorbidity. Cluster B exhibited the highest weighted prevalence of LBP (26.86%; 95% confidence interval [CI], 24.77 to 29.06), followed by Cluster C (24.38%; 95% CI, 22.32 to 26.56) and Cluster A (10.78%; 95% CI, 9.72 to 11.93) (P<.001). Pain intensity (mean [SD], 7.1 (1.97); 95% CI, 6.97 to 7.23) and depressive symptoms (37.48%, 95% CI, 35.03 to 40.00) were key differentiators for Cluster B, while Cluster C showed increased cardiometabolic risk (High CV risk, 60.53%; 95% CI, 57.41 to 63.57), the longest pain exposure (103.74 months [191.89]; 95% CI, 89.64 to 117.84) and highest multisite pain burden (2.47 sites [1.61]; 95% CI 2.34 to 2.60).

CONCLUSIONS

Unsupervised learning identified distinct subgroups within the population with heterogeneous LBP patterns. These findings contribute to a stratified understanding of population-level patterns of low back pain, highlighting the relevance of psychosocial and metabolic factors.