Akademik Veri Yönetim Sistemi
European Spine Journal, 2025 (SCI-Expanded, Scopus)
Background: Mechanical complications requiring revision represent one of the main issues after surgical management for adult spine deformity (ASD). The heterogeneity of the population and the multifactorial nature of mechanical complications have not yet allowed to identify patient-specific risk factors for this complication. The aim of this study was to utilize machine learning (ML) to create homogeneous patient clusters, which could then be analyzed with classical statistical methods to evaluate with granularity the type and rate of mechanical complications and the relative risk factors. Methods: All patients who underwent surgical management for ASD and had a minimal follow-up of two years were clustered into three homogeneous groups. For each, the type and rate of mechanical complication requiring revision were analyzed and logistic regression and marginal effect study were used to evaluate possible risk factors. Results: For Cluster 1 (N = 192, younger subjects with coronal deformity), the revision rate for mechanical complication was 6% (33% pseudarthrosis, 13% screw loosening); the main risk factors were implant density, coronal balance and pelvic incidence at 6 weeks and presence of pelvic fixation. In Cluster 2 (N = 455, older patients with mild coronal and sagittal imbalance and mild disability) the revision rate was 19% (41% pseudarthrosis, 24% proximal junctional kyphosis – PJK); the main risk factors were thoracic kyphosis (TK) at 6 weeks, n. of instrumented levels, ODI, preoperative sacral obliquity, TK and pelvic tilt, and presence of pelvic fixation. In Cluster 3 (N = 172, older patients with severe imbalance and mild disability), the revision rate was 30% (58% pseudarthrosis, 14% screw loosening); the main risk factors were n. of instrumented levels and global tilt at 6 weeks. Conclusion: ML allows to cluster patients into homogeneous groups, which showed differences for mechanical complication rates requiring revision, with specific risk factors in each group. This ML approach might improve preoperative risk assessment.