Cobb angle prediction for adolescent idiopathic scoliosis via an explainable machine learning model

Abstract

This study aims to build an accurate and interpretable machine learning model capable of adolescent idiopathic scoliosis prognostication. A tree-based gradient boosting machine is incorporated with a recently proposed Shapley-value-based explanation method-TreeExplainer. Anthropometric training data are collected from a public orthopedics clinic, and each instance is characterized by nine features with a prediction target. We adopt a transfer-learning strategy that takes advantage of the additive property of tree-based gradient boosting, allowing a gradient boosting machine regressor to be trained with limited labeled examples. Cross-validation estimation shows a satisfactory performance for predicting future spine curvature (Cobb angle). The root mean square error (${}^{\circ }$), the mean absolute percentage error (${}^{\circ }$), and the Pearson correlation coefficient are 3.69 ± 1.23, 2.81 ± 1.69, and 0.92 ± 0.01, respectively. Moreover, the overfitting has been largely removed, and the model may be generalized well to new patients. A well-trained model is taken as the input to the TreeExplainer. The output of the TreeExplainer provides us a richer understanding that demonstrates how a feature’s value impacts the model’s prediction for every instance. The patterns identified can substantially improve the human-artificial intelligence collaboration in the clinical management of patients with adolescent idiopathic scoliosis by preventing serious scoliosis progression and reducing healthcare costs.