Unsupervised machine learning for clinical risk subtyping and re-fracture risk analysis in osteoporotic vertebral compression fractures

To investigate the application of unsupervised machine learning analyzing re-fracture risks in osteoporotic vertebral compression fractures.

A cohort of 867 patients (658 females and 209 males) with 1,255 fracture cases was analyzed. Fractures predominantly occurred at vertebrae T12, L1, and L2. Patients were categorized into once-fracture (n = 730) and multi-fracture (n = 137) groups. Bone mineral density (BMD) was measured, and t-SNE was used for dimensionality reduction. K-means clustering was applied for clustering fracture cases into four risk clusters based on patient-level clinical characteristics and regional fracture distribution (Clusters A: male-predominant low-risk, Cluster B: female low-risk, Cluster C: female high-risk, thoracolumbar, and Cluster D: female high-risk, mid-thoracic).

Multi-fracture patients had a median re-fracture interval of 10 months and significantly lower BMD compared to once-fracture patients. Cluster A fractures were exclusive to males with a re-fracture probability of 19.31%. Female fractures were categorized into Clusters B, C, and D, with re-fracture probabilities of 19.00%, 36.71%, and 32.77%, respectively. Cluster B fractures in females showed a significantly lower re-fracture risk compared to Clusters C and D. Significant differences in age, height, weight, and fracture location were observed among subtypes.

The study highlights the potential of unsupervised machine learning for identifying exploratory, potentially clinically relevant risk phenotypes in OVCF patients, which may inform future risk-stratified follow-up strategies.

The findings underscore the potential of unsupervised machine learning in identifying exploratory data-driven subtypes of spinal fractures and may inform risk stratification, thereby warranting further investigation.