The paper presents a nonlinear regression model built on the coronal thoracic curvature, the lumbar lordosis and the slope of the first lumbar vertebra in order to estimate the thoracic kyphosis measure between T2 and T12. To train the proposed model, a large database containing scoliotic spines demonstrating several types of scoliotic deformities was used to train the proposed system by a cross-validation method. Validation was performed on patients exhibiting three different types of sagittal thoracic profiles: normal, hypo-kyphotic, and hyper-kyphotic. Results show that a multivariate regression model based on dependent variables is able to predict with a reasonable accuracy the sagittal thoracic kyphosis for the automatic assessment and classification of the spinal curve.