Optimizing AIS surgery with a digital twin framework integrating AI and personalized biomechanical modeling

Abstract Purpose Posterior spinal fusion (PSF) is the primary treatment for adolescent idiopathic scoliosis (AIS) curves >45°, yet planning remains largely radiographic and empirical, with limited integration of patient-specific biomechanics. AI and biomechanical modeling now offer opportunities for objective, personalized evaluation of instrumentation strategies. This study developed and evaluated a patient-specific model-based optimization framework combining AI-predicted instrumentation with deterministic physics-based biomechanical simulation to explore construct trade-offs within predefined computational constraints. Methods Twenty AIS patients (Lenke 1/2, age 12–19) treated with selective thoracic PSF were analyzed using data from the MIMO clinical trial. Patient-specific digital twin models were generated from validated 3D reconstructions and implemented in a multibody dynamic simulation environment. AI-predicted instrumentation parameters (UIV, LIV, screw number, rod curvature) were varied within clinical ranges, generating 972 construct configurations per patient (19,440 total simulations). A multi-objective optimization algorithm identified AI-derived strategies balancing 3D correction, screw pullout forces, and preservation of unfused vertebrae, which were compared with simulated surgeon-performed constructs. Simulated outcomes included MT Cobb angle, thoracic kyphosis, apical vertebral rotation, screw pullout force, number of screws, and fused levels. Strategies were compared using repeated-measures MANOVA. Results Optimized configurations achieved similar modeled 3D correction to surgeon-performed constructs, with a mean 3° greater coronal correction (95% CI [0.58, 5.81], p = .019) while maintaining comparable axial derotation (AVR) and thoracic kyphosis. Optimized strategies used fewer screws (mean reduction of 4, p < 0.001) and fewer fused levels (median reduction of 2, p < 0.001). All configurations remained within defined pullout force thresholds. Conclusion In this exploratory simulation study, integration of AI-derived instrumentation with a patient-specific digital twin framework combining deterministic physics-based biomechanical modeling and optimization identified alternative constructs using fewer screws and shorter fusion spans while maintaining broadly comparable modeled 3D correction and implant loading. The small differences in modeled correction should not be interpreted as evidence of clinical superiority, and prospective clinical validation is required.