A Predictive Model of Progression for Adolescent Idiopathic Scoliosis Based on 3D Spine Parameters at First Visit

Marie-Lyne Nault, MD, PhD; Marie Beauséjour, PhD; Marjolaine Roy-Beaudry, MSc; Jean-Marc Mac-Thiong, MD, PhD; Jacques de Guise, PhD; Hubert Labelle, MD; Stefan Parent, MD, PhD


Spine. 2020;45(9):605-611. 

In This Article

Materials and Methods

Study Design and Participants

Our model was derived from a prospective cohort recruited at Sainte Justine University Hospital Center in Montreal, Canada, from January 2006 to May 2010, a tertiary care pediatric teaching hospital and a referral center for scoliosis. The study was approved by the institutional review board, and all participants provided signed informed consent. Inclusion criteria were the following: first visit with an orthopedic surgeon, diagnosis of AIS with Cobb angle between 11° and 40°, age 10 and over, and a Risser sign of 0 or 1. We excluded patients with: congenital, neuromuscular, or syndromic scoliosis, Risser sign of 2 or greater (because they were not likely to progress[8]), curves greater than 40° (because some surgeons would consider surgical instrumentation and fusion).

The final Cobb angle was measured at the endpoint for the study, which corresponds to being fully skeletally mature (Risser 4 or higher) or having a fusion surgery, whichever came first. Brace treatment was allowed according to the treating physician, but brace had to be removed the night before each visit. Brace treatment indications followed the SRS guidelines.[9]


At the first and all subsequent visits, patients had posteroanterior (PA) and lateral spine X-rays. They were followed by one of four spine surgeons, with the follow-up interval between 4 and 6 months chosen by the treating surgeon. There was no modification to the standard clinical protocol for patients included in this study, thus the amount of irradiation was not for research purpose but to ensure an appropriate follow-up.

At first visit, the curve type was defined as: single right thoracic, double with main right thoracic and left lumbar, single left thoracolumbar, or other. Because of the small magnitude of the curves, Lenke and King classifications could not be used. The Risser sign and triradiate cartilage status (open or closed) were evaluated. To limit radiation exposure, skeletal maturation was evaluated from the initial spinal X-rays.

Skeletal maturation was set as either stage 0 (Risser 0 with open triradiate cartilage) or stage 1 (Risser 0 with closed triradiate cartilage or Risser 1).[8] Radiological assessments were made by the same experienced evaluator for all participants. An inter- intra-observer reliability study was performed on 35 cases and published in 2010.[8]

A 3D spinal reconstruction was done for each patient, using the PA and lateral standing radiographs. A single experienced research assistant blinded to the goals of the study did all the reconstructions. Two different software programs were used, each matching the specifications of the two radiographic imaging systems used in the study: Spine 3D (LIS3D, Montreal, Canada) for the Fuji system (Tokyo, Japan) and IdefX (LIO, Montreal, Canada) for the EOS system (Paris, France). The two programs generate 3D reconstructions of comparable accuracy. Pomero et al[10] showed no clinically significant difference in mean error between 3D vertebral models and CT-scan reconstructions. The precision of the reconstructions has been demonstrated as satisfactory, with mean point-to-surface error less than 1.5 mm and mean angular error less than 2° as compared with conventional CT-scan reconstructions.[11–13]

All measurements stemming from the 3D reconstructions were calculated using the same custom software IdefX (LIO, Montreal, Canada), without further human intervention, thus avoiding measurement bias.


Parameters included global (T1-S1 spine), regional (scoliotic segment), or local (vertebra) descriptors. Six types of descriptors were examined (Cobb angles, plane of maximal curvature, 3D vertebral body (Figure 1A) and disk wedging (Figure 1D), axial vertebral rotation (Figure 1C), torsion (Figure 1E), and slenderness (Figure 1B)). Measurements are further detailed in a previous paper by the same authors.[6]

Figure 1.

Three-dimensional spine parameters. (A) Maximal 3D wedging of vertebral body is measured where wedging is greatest around the vertical axis. (B) Slenderness measurement of a single vertebral body, showing height/width (h/w) ratio. (C) Intervertebral rotation measurement. The rotation of the superior vertebra with respect to the inferior vertebra was calculated by projecting its local x-axis onto the x-y plane of the inferior vertebra. (D) Mean measure of two apical 3D disk wedgings. (E) Torsion measurement: χ, mean; Σ, sum; θ, angle.

Statistical Analysis

Sample Size. The study period is sufficient for inclusion of over 100 patients, which would be a minimal requirement for linear modeling. It was chosen to ensure a sample size of at least five times the number of tested variables in the model.

Prediction Model. The outcome for the model was the Cobb angle measured on a PA standing X-ray at the end-point.

Due to the large number of descriptors, we performed univariate analyses to select the most relevant candidate predictors to include in multivariate analyses. We performed Pearson's and Spearman's correlations between the final Cobb angle at skeletal maturity and descriptors of the spine to identify those associated with a P value of 0.1 or less.

The next step consisted of creating the predictive model based on a general linear model (GLM) using a backward selection approach. The initial model includes all candidate predictors selected from the correlation study. Multiplicative interactions for clinically relevant interactions were also tested. The algorithm removes the variable with the smallest contribution to the model, and this procedure is repeated until all remaining included variables contribute significantly to a stronger coefficient of determination R2. The model estimated the beta (β) coefficients with 95% confidence interval (CI).

All statistical analyses described above were done with IBM SPSS Statistics (IBM Corp. Released 2011. IBM SPSS Statistics for Windows, Version 20.0. Armonk, NY: IBM Corp). Patients with missing values (such as calibration errors) or who were lost to follow-up (for comparisons with the final outcome) were removed from the analyses. Model performance was assessed by goodness-of-fit evaluated by R2; calibration with the Bland–Altman method comparing predicted and observed values of Cobb at maturity[14,15] and computation of the mean standard error of prediction; classification capacity and clinical usefulness by the positive and negative predictive values.