Artificial intelligence in early onset scoliosis: a scoping review

TL;DR

This paper reviews how artificial intelligence is being used to diagnose and manage early onset scoliosis in children under 10, highlighting promising results and areas needing improvement.

Contribution

The study provides a comprehensive scoping review of AI applications in early onset scoliosis, identifying current methods and limitations.

Findings

Most studies used convolutional neural networks for image analysis in early onset scoliosis.

AI models achieved mean accuracy of 91.2% in predicting outcomes and analyzing spinal deformities.

Common limitations included small sample sizes and lack of external validation.

Abstract

Early onset scoliosis comprises spinal deformities in children younger than 10, creating challenges in diagnosis, risk assessment, and management. Timely intervention is vital, because untreated deformity can lead to cardiopulmonary compromise. Artificial intelligence and machine learning are reshaping orthopaedic care by improving detection, forecasting progression, and guiding treatment. This scoping review maps current use in this patient population.

Following PRISMA ScR standards, we systematically searched PubMed, Embase, Web of Science, Cochrane, and Scopus for studies that developed, applied, or validated AI models to diagnose, manage, or predict outcomes in EOS.

After removing duplicates, 352 records were screened, 22 full texts were reviewed, and 11 studies met inclusion criteria. Most investigations (63.6%) employed convolutional neural networks (CNNs) such as Mask R CNN, EfficientNet, and U Net. Ensemble learning with gradient boosting, random forest, and logistic regression (9.1%), Gaussian Naïve Bayes (9.1%), sparse additive machines (9.1%), and unsupervised clustering (9.1%) were also used. Image analysis dominated (72.7%), automating radiographic measurements (Cobb angle, skeletal maturity) and monitoring growing-rod distraction. Predictive models (27.3%) estimated prolonged hospital stay, unplanned reoperation, or postoperative complications. Mean accuracy was 91.2% (range 86.1% to 94.0%). Common limitations were small sample sizes, single-centre data, and limited external validation.

AI shows promise for EOS imaging and risk prediction, yet translation is hindered by methodological heterogeneity and scarce external validation. Future work should adopt standardised reporting, aggregate multicentre datasets, and test models prospectively in large cohorts.

Figures