Machine Learning-Guided LIV Selection for Adolescent Idiopathic Scoliosis

Overview

This study will evaluate whether a machine learning-based decision support model, called the Drum Tower Rule, can help surgeons select the lowest instrumented vertebra during corrective surgery for adolescent idiopathic scoliosis. Patients with Lenke type 1 or Lenke type 5 adolescent idiopathic scoliosis who are scheduled for posterior spinal fusion will be randomly assigned to one of two groups. In the model-guided group, surgeons will receive the model-predicted risk of postoperative distal adding-on and a recommendation for lowest instrumented vertebra selection. In the conventional-experience group, surgeons will select the lowest instrumented vertebra according to routine clinical experience and existing surgical principles, without access to the model output. All patients will receive standard posterior spinal fusion. The main outcome is the incidence of distal adding-on at 24 months after surgery, assessed by blinded radiographic reviewers.

Adolescent idiopathic scoliosis is a common spinal deformity in children and adolescents. For patients requiring corrective surgery, selection of the lowest instrumented vertebra is a key surgical decision. An inappropriate distal fusion level may increase the risk of postoperative distal adding-on, coronal imbalance, unnecessary loss of spinal mobility, or revision surgery. The Drum Tower Rule is a machine learning-based decision support model developed to estimate the risk of postoperative distal adding-on and assist with lowest instrumented vertebra selection in patients with Lenke type 1 and Lenke type 5 adolescent idiopathic scoliosis. Before the start of this trial, the model, input variables, risk threshold, and software version will be locked and will not be modified during the study. This is a single-center, prospective, randomized, open-label, parallel-group controlled trial with blinded outcome assessment. Eligible participants will be randomly assigned in a 1:1 ratio to either the model-guided group or the conventional-experience group. Randomization will be stratified by Lenke classification. In the model-guided group, preoperative clinical and radiographic variables will be entered into the locked machine learning model. The model will generate a predicted risk of distal adding-on and a recommendation for lowest instrumented vertebra selection. For Lenke type 1 patients, the decision will focus on selection between one level proximal to the last substantially touching vertebra and the last substantially touching vertebra. For Lenke type 5 patients, the decision will focus on selection between L3 and L4. The surgeon will make the final decision after considering the model output and clinical judgment. In the conventional-experience group, the surgeon will select the lowest instrumented vertebra based on routine clinical experience and existing surgical principles. The model output will not be provided to the surgeon for participants in this group. Both groups will undergo standard posterior spinal fusion with an all-pedicle screw instrumentation system. Postoperative follow-up will be performed at 1 week and at 3, 6, 12, and 24 months after surgery. The primary outcome is the incidence of distal adding-on at 24 months after surgery. Secondary outcomes include number of fused segments, Cobb angle correction rate, coronal balance, sagittal radiographic parameters, Scoliosis Research Society-22 score, visual analog scale score for low back pain, complications, instrumentation failure, revision surgery, and adoption of model recommendations. The primary outcome will be assessed by independent radiographic reviewers who are blinded to treatment allocation.