Normative Alignment Goals Using Machine Learning Finds the Sweet Spot Between Pseudarthrosis and Proximal Junctional Kyphosis in Adult Spinal Deformity
Clinical Researcher Columbia University Medical Center/NewYork-Presbyterian Och Spine Hospital
Disclosure(s):
Justin Reyes, MS: No financial relationships to disclose
Introduction: Traditional age-adjusted spinopelvic alignment formulas risk under-correction in ASD patients. Leveraging machine learning, this study develops surgical targets by analyzing alignment in asymptomatic volunteers.
Methods: A predictive model was built for PI-LL from 468 asymptomatic adults(80% training, 20% validation) across multiple centers/ethnicities. The eXtreme Gradient Boosting algorithm utilized PI, age, & sex. To validate targets, we analyzed 458 ASD patients with 2Y follow-up. These patients were classified as under-(UC), adequately-(AC), or over-corrected(OC), based on the model’s targets ±5°. Key outcomes were pseudarthrosis/implant breakage & PJK. Outcomes were analyzed using multivariable regression models, adjusted for significant variables identified in univariate analyses. Data shown as [UC vs AC vs OC,P(ANOVA)].
Results: Mean absolute error between observed & predicted PI-LL were 3.04° & 5.02° for training & validation groups. In the surgical ASD cohort, 149(32.5%), 159(32.8%), & 150(34.7%) patients were UC, AC, & OC respectively. Differences were observed in instrumented levels(12.31 vs 12.69 vs 13.8,P=0.0028), baseline PI-LL(30.3° vs 22.1° vs 17.8°,P < 0.0001), & T1PA(30.9° vs 26.0° vs 23.4°,P < 0.0001). Pseudarthrosis rate was 9.82%(45/548), with highest incidence in UC cohort(15.4% vs 8.18% vs 6.0%,P=0.0161). PJK rate was 10.0%(46/458), most prevalent in OC group(19.3% vs 6.04% vs 5.03%,P < 0.0001). In an adjusted multivariable model(P < 0.0001, AUC=0.76) found that AC(aOR: 0.45,P=0.046), & OC(aOR: 0.41,P=0.044) had lower odds of pseudarthrosis compared to UC patients. In an adjusted PJK model(AUC=0.687,P < 0.0001), AC had lower odds of PJK compared to OC (OR: 0.45,P=0.0034. Both models found the current classification supersedes baseline alignment and magnitude correction in association with pseudarthrosis & PJK.
Conclusion : Machine learning-derived PI-LL targets demonstrate a critical balance in deformity correction. Deviation from these tailored benchmarks increases risk of pseudarthrosis when under-corrected and PJK when over-corrected.