INTRODUCTION

Adolescent idiopathic scoliosis (AIS) is the most common form of scoliosis, attributing to about 70% of all cases, and is usually diagnosed during puberty [1-3]. One can imagine a curved spine can detrimentally affect the patient’s daily activities, quality of life and even cardiopulmonary function [4]. The progressive nature of AIS warrants an early diagnosis and treatment to prevent progression as soon as possible and to save the patient from more aggressive interventions. Many studies have investigated how structural deformity parameters, and even biomarkers, may contribute to the likelihood of AIS progression. [5-7] But no study has been done so far that utilizes a convolutional neural network (CNN) to combine information from raw two-dimensional (2D) radiographs and one-dimensional (1D) clinical parameters in the research of AIS progression prediction. Thus, by using capsule network as a backbone [8], this study aims to develop a CNN model that is able to predict the probability of AIS progression at the patient’s first visit by integrating 2D radiological images with 1D clinical data.

METHODS

A total of 513 patients have been recruited with the exclusion of 43 patients due to lack of followup. Clinical parameters were recorded at specialist clinics and bi-planar radiographs of the full body were scanned with the EOS system.

Three different crops of the original radiographs are generated -- (1) whole-spine crop, (2) upper end vertebra crop of major curvature, and (3) lower end vertebra crop of major curvature, all of which are used as the 2D inputs of the CNN model.

The clinical parameters used in this study include sex, age, weight, sitting height, standing height, arm span, risser sign, ulna maturity and distal radius maturity. Categorical data are represented as 1D unit vectors, whereas numerical data are normalized to a number between 0 and 1. The processed vector then serves as the 1D input of the CNN model.

The presence of AIS progression is defined by a minimum increase of 5 degrees in the Cobb angle of the major curve, which can happen as soon as 2 months after the initial consultation. This definition is used to verify the predicted outputs from the CNN model.

RESULTS

The CNN model achieved a 92.6% sensitivity and 90.6% negative predictive value (NPV) with an overall accuracy of 72.8%. In addition, it also achieved 56.9% specificity and 63.3% positive predictive value (PPV). The ROC AUC is 0.76.

DISCUSSION

The high sensitivity and NPV make this AI model a great screening tool to exclude those without the risk of AIS progression. With a negative risk predicted by the model, treatment resources can then be spared and reassurance can be given to the patient. With a positive predicted progression risk, the specialist can decide to more closely monitor AIS progression of the patient by increasing the frequency of follow ups to significantly speed up the diagnosis and management of AIS progression and further optimize appropriate allocation of healthcare resources. Limitations of this study include a small data size and occasional missing clinical data.