A machine learning-based framework for automated clinical measurements, deformity classification and prediction of post-operative outcomes in early onset scoliosis patients

Early Onset Scoliosis (EOS) is a complex spinal and thoracic deformity affecting children under 10 years of age, accounting for approximately 10% of all pediatric scoliosis cases. The etiology of EOS varies widely and can be classified into congenital, idiopathic, neuromuscular, and syndromic scoliosis. Progressive spine deformity in EOS patients leads to altered thoracic cage development, resulting in reduced thoracic size and shape, which may ultimately impair normal lung development and cause thoracic insufficiency syndrome (TIS). TIS can lead to complications such as pulmonary hyperplasia and premature death. The C-EOS classification system is the only available method for grouping EOS patients based on pre-operative clinical indices. However, the cut-offs for major curve (Cobb) angle and kyphosis in C-EOS are not based on a data-driven approach, highlighting the need for a more robust, data-driven method to group EOS patients using pre-operative clinical indices such as age, Cobb angle, and kyphosis. Currently, there are no machine learning (ML) based methods to predict clinical outcomes after surgical intervention for the EOS population. Management strategies for EOS rely on clinician preferences and opinions, and despite the C-EOS system's ability to organize complex EOS pathologies, there is a lack of consensus among surgeons regarding intervention age, instrumentation type, and location. Therefore, an ML-based approach is needed to predict surgical outcomes based on the type of intervention. Manual radiographic measurement of clinical indices is time-consuming and associated with significant inter-observer errors, particularly in the presence of vertebral, lung, and ribcage abnormalities in EOS patients. Automated identification of anatomical structures, such as the vertebrae and lungs, is necessary to facilitate faster and more accurate clinical measurements. This thesis presents novel automated methods for clustering EOS patients using commonly used pre-operative clinical indices and provides a web-based application for accessibility. Machine learning-based regression is employed to predict Cobb angle at three time points post-intervention: immediately post-op, 2-year post-op, and 5-year post-op. The most predictive features contributing to the prediction of post-operative Cobb angle are also identified. Furthermore, a framework for segmenting vertebrae and lungs in adolescent idiopathic scoliosis (AIS) and EOS patients is developed, enabling the automated measurement of clinical indices such as Cobb angle, kyphosis, and lordosis.