From 3 to 8 clusters: a machine learning-based classification of operated adult spinal deformity patients
Spine Journal, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.spinee.2026.05.003
- Dergi Adı: Spine Journal
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, CINAHL, EMBASE, MEDLINE, Academic Search Ultimate (EBSCO)
- Anahtar Kelimeler: Adult spinal deformity, Clustering, K-means algorithm, Machine learning, Phenotyping, Predictive modeling, Surgical planning
- Acıbadem Mehmet Ali Aydınlar Üniversitesi Adresli: Evet
Özet
BACKGROUND CONTEXT: Adult spinal deformity (ASD) is a heterogeneous condition encompassing diverse etiologies, clinical presentations, and surgical challenges. While previous unsupervised learning models have stratified ASD patients into 3 broad phenotypes, these often lack clinical granularity for surgical decision-making. PURPOSE: To refine the classification of operated ASD patients using unsupervised machine learning and assess whether expanding from 3 to 8 clusters improves clinical discrimination and predictive robustness. STUDY DESIGN: Retrospective multicenter study using machine learning clustering. PATIENT SAMPLE: 471 adult patients who underwent surgery for ASD across 6 specialized spine centers. OUTCOME MEASURES: ODI, SRS-22, spinal alignment metrics, surgical strategy, and postoperative complications. METHODS: A k-means clustering algorithm was applied to a selected set of 12 demographic, radiographic, and functional variables (C12). Clustering solutions with 3 and 8 groups were compared. Each cluster was analyzed for age, etiology, disability scores (ODI, SRS-22), spinal alignment, surgical strategy, and complication rates. Predictive models using LDA and KNN assessed classification accuracy for new patient assignment. RESULTS: The 8-cluster model identified clinically distinct phenotypes, including subgroups of young patients with idiopathic scoliosis, structural hyperkyphosis, and early sagittal decompensation. In elderly patients, clusters differentiated pain profiles, alignment, and frailty. The most severe clusters distinguished coronal-dominant vs sagittal collapse deformities, with differing risks and outcomes. LDA maintained robust accuracy (91.4%) in predicting cluster assignment across 50 testing splits, outperforming KNN. CONCLUSIONS: To our knowledge, this is among the most granular unsupervised ML-based phenotypic classifications of operated ASD patients reported to date. This refined 8-cluster model enhances clinical phenotyping in ASD surgery, offering more precise subgroup stratification than traditional 3-cluster models. It supports the integration of unsupervised learning into personalized surgical planning, risk stratification, and multicenter outcome trials.