Principles Underlying Surgical Decision-Making in Lowest-Instrumented Vertebra Level Selection of Adolescent Idiopathic Scoliosis: From Traditional Landmarks to Emerging Predictive Models—A Narrative Review

Background: Adolescent idiopathic scoliosis (AIS) affects approximately 0.5–5.2% of adolescents. Only a minority of curves progress, and surgery is generally reserved for curves exceeding 45–50°. Selection of the lowest instrumented vertebra (LIV) remains contested and directly influences postoperative coronal balance, distal adding-on (DA) and revision risk. Aim: This narrative review synthesizes the biomechanical principles governing LIV selection, compares traditional anatomical landmarks with contemporary prediction formulas and machine learning (ML) tools, and defines what these tools can and cannot currently deliver at the bedside. Methods: A structured but deliberately non-systematic search of PubMed/MEDLINE, EMBASE and the Cochrane Library (January 2000–December 2024) was supplemented by purposive inclusion of seminal pre-2000 work. No PRISMA-compliant screening, risk-of-bias appraisal or quantitative synthesis was undertaken, and the article is therefore reported as a narrative review. Results: Traditional landmarks—end, neutral, stable, last touched (LTV) and last substantially touched vertebra (LSTV)—are compared; fusion at or distal to the LTV/LSTV lowers DA incidence, yet 8–28% of patients still develop DA. Prediction formulas estimate the residual lumbar curve to within approximately 6°, but each was derived within a single surgical team and none nominates a vertebral level. A composite index combining (LIV− Neutral vertebra) and (LIV− Stable vertebra) reported sensitivity 100% and specificity 92–94% in its derivation cohort, without external validation. Hypercorrection of the main thoracic curve (>53%) combined with postoperative LIV tilt < 10° is associated with DA. Machine learning models predict three-dimensional alignment within 5–7°, yet every published model predicts an outcome; none has shown that acting on a model-derived recommendation prevents DA. Conclusions: LIV selection is best framed as a multidimensional judgment integrating coronal balance, vertebral body rotation, sagittal alignment, skeletal maturity, lumbopelvic morphology and non-radiographic factors such as paraspinal muscle quality, bracing and rehabilitation. Prediction models and artificial intelligence remain hypothesis-generating: external validation and prospective outcome trials are prerequisites before they can be considered practice-changing.