Abstract
A major clinical challenge in adolescent idiopathic scoliosis (AIS) is the difficulty of predicting curve progression at initial presentation. The early detection of progressive curves can offer the opportunity to better target effective non-operative treatments, reducing the need for surgery and the risks of related complications. Predictive models for the detection of scoliosis progression in subjects before growth spurt have been developed. These models accounted for geometrical parameters of the global spine and local descriptors of the scoliotic curve, but neglected contributions from biomechanical measurements such as trunk muscle activation and intervertebral loading, which could provide advantageous information. The present study exploits a musculoskeletal model of the thoracolumbar spine, developed in AnyBody software and adapted and validated for the subject-specific characterization of mild scoliosis. A dataset of 100 AIS subjects with mild scoliosis and in pre-pubertal age at first examination, and recognized as stable (60) or progressive (40) after at least 6-months follow-up period was exploited. Anthropometrical data and geometrical parameters of the spine at first examination, as well as biomechanical parameters from musculoskeletal simulation replicating relaxed upright posture were accounted for as predictors of the scoliosis progression. Predicted height and weight were used for model scaling because not available in the original dataset. Robust procedure for obtaining such parameters from radiographic images was developed by exploiting a comparable dataset with real values. Six predictive modelling approaches based on different algorithms for the binary classification of stable and progressive cases were compared. The best fitting approaches were exploited to evaluate the effect of accounting for the biomechanical parameters on the prediction of scoliosis progression. The performance of two sets of predictors was compared: accounting for anthropometrical and geometrical parameters only; considering in addition the biomechanical ones. Median accuracy of the best fitting algorithms ranged from 0.76 to 0.78. No differences were found in the classification performance by including or neglecting the biomechanical parameters. Median sensitivity was 0.75, and that of specificity ranged from 0.75 to 0.83. In conclusion, accounting for biomechanical measures did not enhance the prediction of curve progression, thus not supporting a potential clinical application at this stage.