Advanced machine learning model for Cobb angle progression in adolescent idiopathic scoliosis with surface...

Excessive radiographic exposure in the follow-up of adolescent idiopathic scoliosis (AIS) remains a clinical concern. Surface topography (ST) and angle of trunk rotation (ATR) have shown promise for non-radiographic monitoring, although their ability to detect clinically meaningful Cobb angle changes (> 5°) remains limited. This study aimed to validate a machine learning model for predicting Cobb angle progression using ST parameters and ATR obtained at three and six months.

A prospective observational study was conducted in 43 AIS patients (57 curves) recruited from two centers. Baseline and six-month radiographic Cobb angles were recorded along with ATR and five ST asymmetry parameters (MaxDev, RMS, LatDev, hump volume, asymmetry patch area). A random forest (RF) model was used to predict Cobb angles at 3 and 6 months and then to estimate progression over 6 months (ΔCobb). Outcomes were classified as improvement (ΔCobb < -5°), stabilization (-5° ≤ ΔCobb ≤ + 5°), or progression (ΔCobb > + 5°).

The RF model predicted the six-month radiographic Cobb with MAE of 7.03° (three-month input) and 6.91° (six-month input). The progression model integrating both time points achieved an overall accuracy of 80.7%, with detection accuracies of 100% for stabilization, 53.8% for improvement, and 37.5% for progression. Quantitatively, 84.2% of the curves had a progression prediction error of less than 5°.

The model accurately identified stabilized cases, suggesting that non-radiographic follow-up combining ST and ATR could reliably detect non-progressive AIS within six months. This approach could potentially reduce up to 80% of unnecessary follow-up radiographs in stable patients.