Prediction of future curve angle using prior radiographs in previously untreated idiopathic scoliosis: natural history from age 6 to after the end of growth (SOSORT 2022 award winner)

Abstract

Purpose Treatment selection for idiopathic scoliosis is informed by the risk of curve progression. Previous models predicting curve progression lacked validation, did not include the full growth/severity spectrum or included treated patients. The objective was to develop and validate models to predict future curve angles using clinical data collected only at, or both at and prior to, an initial specialist consultation in idiopathic scoliosis. Methods This is an analysis of 2317 patients with idiopathic scoliosis between 6 and 25 years old. Patients were previously untreated and provided at least one prior radiograph prospectively collected at first consult. Radiographs were re-measured blinded to the predicted outcome: the maximum Cobb angle on the last radiograph while untreated. Linear mixed-effect models were used to examine the effect of data from the first available visit (age, sex, maximum Cobb angle, Risser, and curve type) and from other visits while untreated (maximum Cobb angle) and time (from the first available radiograph to prediction) on the Cobb angle outcome. Interactions of the first available angle with time, of time with sex, and time with Risser were also tested. Results We included 2317 patients (83% of females) with 3255 prior X-rays where 71% had 1, 21.1% had 2, and 7.5% had 3 or more. Mean age was 13.9 ± 2.2yrs and 81% had AIS. Curve types were: 50% double, 26% lumbar/thoracolumbar-lumbar, 16% thoracic, and 8% other. Cobb angle at the first available X-ray was 20 ± 10° (0–80) vs 29 ± 13° (6–122) at the outcome visit separated by 28 ± 22mths. In the model using data at and prior to the specialist consult, larger values of the following variables predicted larger future curves: first available Cobb angle, Cobb angle on other previous X-ray, and time (with Time2 and Time3) to the target prediction. Larger values on the following variables predicted a smaller future Cobb angle: Risser and age at the first available X-ray, time*Risser and time*female sex interactions. Cross-validation found a median error of 4.5o with 84% predicted within 10°. Similarly, the model using only data from the first specialist consult had a median error of 5.5o with 80% of cases within 10° and included: maximum Cobb angle at first specialist consult, Time, Time2, age, curve type, and both interactions. Conclusions The models can help clinicians predict how much curves would progress without treatment at future timepoints of their choice using simple variables. Predictions can inform treatment prescription or show families why no treatment is recommended. The nonlinear effects of time account for the rapid increase in curve angle at the beginning of growth and the slowed progression after maturity. These validated models predicted future Cobb angle with good accuracy in untreated idiopathic scoliosis over the full growth spectrum.

3D prediction of curve progression in adolescent idiopathic scoliosis based on biplanar radiological reconstruction

“…Machine-learning algorithms are capable of processing complex data and generating more accurate predictions compared to traditional regression models. Serial reconstructions arranged in stepwise layers were found to strongly improve prediction accuracy, 7 , 66 as this allows better extrapolation of growth trajectories. While random forest model and probabilistic classification model were reported as useful prognostication models, 66 , 67 more complex models, such artificial neural network models, 19 , 34 have yet to be explored in 3D analysis.…”

“…It is well established that a larger Cobb angle predisposes to curve progression. 1 , 7 , 18 A systematic review by Wong et al 1 found that initial 2D Cobb angle > 25° and thoracic curves were predictive of curve progression. However, most of our included longitudinal studies involves the first radiograph at early visits, when patients are skeletally immature and have mild curves.…”

“…Nevertheless, as Cobb angle acts as a poor predictor at mild stages, the coronal curvature eventually evolves along with wedging and rotation in other planes, which adds predictive value. To evaluate the predictive power of increasing Cobb angle on growing patients, Parent et al 7 compared the accuracy of predictive models based on 2D Cobb angle assessed at different consultations. Relative skeletal immaturity with larger initial 2D Cobb angle and longer duration of observation were associated with curve progression, though the cohort had a larger range of baseline Cobb angle (20° (SD 10°)).…”

“… 1 - 6 Prediction of curve progression can reduce unnecessary consultations and bracing in non-progressive patients, while allowing earlier intervention and proper prognostication to progressive patients. 7 , 8 …”

“… 29 - 31 While providing extensive quantitative data, commercially available programs for biplanar reconstruction still require considerable manual effort in mapping spinal landmarks prior to the automated measurement sequence. 32 In recent years, we have also seen a rise in transdisciplinary studies using machine learning on clinical data to develop in-house programs for predicting curve progression, 7 , 16 , 33 - 36 which involves specialized terminology that may be challenging to digest.…”