Machine Learning Predicts the 3D Outcomes of Adolescent Idiopathic Scoliosis Surgery Using Patient-Surgeon Specific Parameters.

STUDY DESIGN

Retrospective descriptive, multicenter study.

OBJECTIVE

The aim of this study was to predict the three-dimensional (3D) radiographic outcomes of the spinal surgery in a cohort of adolescent idiopathic scoliosis (AIS) as a function preoperative spinal parameters and surgeon modifiable factors.

SUMMARY OF BACKGROUND DATA

Current guidelines for posterior spinal fusion surgery (PSF) in AIS patients are based on two-dimensional classification of the spinal curves. Despite the high success rate, the prediction of the 3D spinal alignment at the follow-ups remains inconclusive. A data-driven surgical decision-making method that determines the combination of the surgical procedures and preoperative patient specific parameters that leads to a specific 3D global spinal alignment outcomes at the follow-ups can lessen the burden of surgical planning and improve patient satisfaction by setting expectations prior to surgery.

METHODS

A dataset of 371 AIS patients who underwent a PSF with two-year follow-up were included. Demographics, 2D radiographic spinal and pelvic measurements, clinical measurements of the trunk shape, and the surgical procedures were collected prospectively. A previously developed classification of the preoperative global 3D spinal alignment was used as an additional predictor. The 3D spinal alignment (vertebral positions and rotations) at two-year follow-up was used as the predicted outcome. An ensemble learner was used to predict the 3D spinal alignment at two-year follow-up as a function of the preoperative parameters with and without considering the surgeon modifiable factors.

RESULTS

The preoperative and surgical factors predicted three clusters of 3D surgical outcomes with an accuracy of 75%. The prediction accuracy decreased to 64% when only preoperative factors, without the surgical factors, were used in the model. Predictor importance analysis determined that preoperative distal junctional kyphosis, pelvic sagittal parameters, end-instrumented vertebra (EIV) angulation and translation, and the preoperative 3D clusters are the most important patient-specific predictors of the outcomes. Three surgical factors, upper and lower instrumented vertebrae, and the operating surgeon, were important surgical predictors. The role of surgeon in achieving a certain outcome clusters for specific ranges of preoperative T10-L2 kyphosis, EIV angulation and translation, thoracic and lumbar flexibilities, and patient's height was significant.

CONCLUSION

Both preoperative patient-specific and surgeon modifiable parameters predicted the 3D global spinal alignment at two-year post PSF. Surgeon was determined as a predictor of the outcomes despite including 20 factors in the analysis that described the surgical moves. Methods to quantify the differences between the implemented surgeon modifiable factors are essential to improve outcome prediction in AIS spinal surgery.Level of Evidence: 3.