Machine Learning-based Cluster Analysis Identifies Three Unique Phenotypes of Patients With Adult Spinal Deformity With Distinct Clinical Profiles and Long-term Recovery Trajectory: A Development Study

Study design: A retrospective review of a prospective adult spinal deformity data.

Objective: To identify distinct patient clinical profiles and recovery trajectories in patients with adult spinal deformity (ASD).

Summary of background data: Patients with ASD exhibit a diverse array of symptoms and significant heterogeneity in clinical presentations, posing challenges to precise clinical decision-making. Accurate patient selection may provide further insight to personalized management strategies.

Methods: Latent profile analysis (LPA) was performed to determine possible patient phenotype. Goodness-of-fit indices were used to determine the optimal cluster profiles. Outcome differences were evaluated using Analysis of Variance (ANOVA) and subsequent post hoc Tukey's test, while significant predictors of group membership were identified through multinomial logistic regression.

Results: A total of 204 ASD patients (mean age of 60.3 ± 11.8 years, comprising 62.3% females) with complete 1-year and 2-year follow-up outcome were included. LPA identified three phenotypes: 51 patients in phenotype 1, 73 patients in phenotype 2 and 80 patients in phenotype 3, respectively. Each phenotype exhibited a unique symptom profile and distinct functional recovery trajectories. Patients in phenotype 3, although demonstrated the worst Scoliosis Research Society-22 questionnaire (SRS-22r) domains at baseline, patients in this cluster exhibited the most substantial Δchange in SRS-22r domains except for self-image at both 1-year and 2-year follow-up. Remarkably, a relative large proportion of patients (58.8%) who were dissatisfied at 1-year follow-up transited to satisfied at 2-year follow-up. Advanced age, longer symptom duration, severe preoperative pelvic incidence-lumbar lordosis (PI-LL) mismatch, higher preoperative sagittal vertical axis (SVA), fusion extending to sacrum/pelvis and grade ≥ 3 osteotomy predicted membership in the phenotype 3.

Conclusions: LPA enabled the delineation of three distinct phenotypes among ASD patients, each characterized by unique clinical profiles and distinct long-term recovery trajectories. By pinpointing the crucial variables that uniquely distinguish and predict membership in different phenotypes, the study provides valuable guidance for patient stratification.