4. Systematic clustering analysis using multimodal data in a chronic low back pain cohort: a preliminary baseline analysis in the ComeBACK Study

Chronic low back pain (cLBP) is a major global cause of disability, with health care costs exceeding $100 billion annually in the US. Its treatment poses a complex challenge due to the interplay of biological, psychological, and social factors. Understanding the diversity of cLBP is crucial for effective treatment, especially precision medicine for nonspecific cLBP. Unsupervised clustering methods have been increasingly used for patient phenotyping, yet the outcomes and interpretations of these methods heavily depend on the implementation and decisions made during analysis. Utilizing multimodal data from a cLBP prospective cohort, our aim was to systematically evaluate different clustering algorithms for unsupervised patient phenotyping. We critically assessed the effectiveness and reliability of these algorithms in our analysis. N/A N/A N/A We analyzed multimodal data (demographics, clinical characteristics, imaging, biomechanics, and psychosocial factors) from a preliminary 244 participants in the UCSF ComeBACK cohort study, collected between 2019 to 2023. After consulting domain experts, 679 variables were selected, and missing value imputation was performed. The data underwent dimensional reduction using Multiple Factor Analysis (MFA) or Isometric Mapping (ISOMAP). Systematic clustering algorithms applied included K-means (KM), Fuzzy K-means (Fuzzy), Agglomerative Clustering (AGC), Gaussian Mixture Model (GMM), Bayesian Gaussian Mixture Model (BGMM), and Spectral Clustering (SPC). After grid search for number of clusters (2-10), we plotted gap statistics (GAP) and silhouette scores (SS), two distinct metrics of clustering performance, exploring optimal cluster numbers and evaluating findings. Visually, common clustering tendencies across different algorithms were observed. Specifically, a commonality around 6 clusters was noted, with shared clusters identifiable in 6 regions (top, right, center, top-left, bottom-left, and bottom) of the plot, except for AGC. Subsequently, the effectiveness of systematic cluster metrics was calculated. Both GAP and SS, metrics considering the density within clusters and separation between clusters, were used to measure clustering quality. The inflection point where the increase in GAP values slows down indicates the appropriate number of clusters, while higher SS signify better cluster separation and cohesion. A common trend was observed across both MFA and ISOMAP processed data for each metric. Particularly in ISOMAP,