We performed a search using PubMed via National Library of Medicine, Embase via Embase, and Cochrane reviews via Cochrane Library; from January 2010 to April 2025. The search strategy was based on Medical Subject Heading (MeSH) for PubMed and Entree for Embase, terms or equivalent using idiopathic scoliosis AND intelligent artificial OR Machine learning AND Systematic reviews OR Metanalysis.

The inclusion criteria were female and male with AIS diagnosis or suspicion (population); Use of AI tools or Machine Learning models (intervention); current standard of care (comparator or control); and outcomes focused on following benefits of AI for (i) diagnosis, (ii) treatment, (iii) follow-up, (iv) quality of life/psychological effects, and (v) possible harms associated with the use of AI (outcome). Studies were systematic reviews, meta-analysis, narrative reviews, scoping reviews, published in English between 01 January 2010 and 01 April 2025.

The exclusion criteria applied were articles not written in English, studies that did not include exclusively paediatric populations, those addressing other types of scoliosis or spinal deformities, and publications limited to conference abstracts, letters to the editor, or responses.

Data extraction was conducted independently by two reviewers using a tested and standardized form. For quality assurance and search strategy peer review, we validated the primary search strategy by testing if known relevant records were retrieved. We used the AMSTAR 2 (A MeaSurement Tool to Assess systematic Reviews) to assess the quality of systematic reviews, including those based on non-randomized studies of healthcare (4). Conflict between reviewers was either resolved between reviewers or by a third reviewer.

The initial literature search retrieved 128 articles from Embase, Cochrane and PubMed. Following the abstract screening, 20 duplicate articles were identified and removed. Continuing with the predefined search strategy, a total of 29 full-text studies were reviewed (Figure 1). After analyzing these 29 articles, only five studies were independently selected by both reviewers to minimize potential selection bias (6–10). The two reviewers independently agreed that these five articles met the initial inclusion criteria, as they were systematic reviews with meta-analyses focusing on artificial intelligence and idiopathic scoliosis.

Critical appraisal of reviews was independently performed by two reviewers using the 16 items of the AMSTAR 2 tool (Table 1) (4). This tool allows to evaluate reviews that include randomized or non-randomized studies of health care intervention. Conflict between reviewers was either resolved between reviewers or by a third reviewer.

Following the AMSTAR recommendations, an overall rating of confidence in the results of the published reviews are displayed in Table 2 (4). Two reviews were rated as low confidence in their results (6, 8), and two as moderate confidence in their results (7, 9). The metanalysis by Zhu et al. (10) was assessed as providing high-confidence results.

Chen K et al. review (6) was one of the first to describe how Machine Learning (ML), an advanced branch of AI, was driving a revolution in the clinical practice of scoliosis. It can impact every stage, from screening and diagnosis to surgical planning and prognosis prediction. For instance, ML algorithms, including Artificial Neural Networks (ANNs) and Convolutional Neural Networks (CNNs), significantly enhance early detection using images, reducing reliance on radiation and automating Cobb angle measurements. In surgical decision-making and assistance, ML helps classify curve types, identify optimal fusion regions, and supports precise pedicle screw insertion and 3D vertebral reconstruction. Furthermore, ML can predict postoperative complications like proximal junctional failure and surgical site infections and forecasting long-term patient outcomes. Despite its vast potential, the authors discuss several challenges including data limitations, issues with generalization, and the “black box” nature of some algorithms. Crucially, ethical concerns regarding patient data privacy and security must be addressed. Finally, the authors conclude that the future aims for a “human-and-machine” approach with multidisciplinary collaboration to fully integrate ML reliably into scoliosis care.

In 2023, Zhang H et al. (8) reviewed the impact of AI—utilizing ML and deep learning (DL)—on the diagnosis and treatment of scoliosis. AI significantly improves early imaging-based screening by enhancing diagnostic accuracy and reducing radiation exposure, including applications involving Large Language Models (LLMs) for patient education and severity assessment. It enables automatic and precise evaluation of spinal parameters, such as the Cobb angle, thereby reducing manual workload. In therapeutic decision-making, AI facilitates rapid classification, progression prediction, and surgical planning, including outcome forecasting. In surgical assistance, AI supports accurate pedicle screw placement and robot-assisted procedures, improving precision and reducing complications. Additionally, AI algorithms can effectively predict postoperative prognosis and potential complications. Despite these benefits, limitations remain, including data scarcity, limited generalizability, and critical ethical concerns related to patient privacy and algorithmic bias. The authors concluded similarly—that future efforts should focus on improving generalizability and ensuring the robust and ethical integration of AI into clinical practice.

Two additional reviews were published in 2024. One, authored by Goldman SN et al. (7) offers a comprehensive examination of the applications of A in the management of AIS. This review specifically explores the utilization of AI—particularly CNNs—in the automated analysis of radiographic images, classification of curve severity, and prediction of curve progression in AIS. The authors underscore the considerable potential of AI to enhance diagnostic accuracy and inform treatment decisions. However, they also identify several critical challenges, including the absence of standardized clinical implementation guidelines, limited model interpretability, and a lack of robust external validation. The review concludes by emphasizing the necessity of future research aimed at addressing these limitations to foster clinician trust and facilitate the effective integration of AI into routine clinical practice.

In the second one from 2024, Li L et al. (9) present a systematic review that highlights the ways in which various ML models—such as Random Forests (RF), Diffusion-Convolutional Neural Networks (DCNNs), Capsule Neural Networks (CapsNets), and Recurrent Neural Networks (RNNs) with Long Short-Term Memory (LSTM) units—enhance predictive accuracy compared to traditional approaches. These models leverage multidimensional input data, including patient demographics, clinical parameters, and two-dimensional imaging modalities (e.g., radiographs, surface topography scans, and smartphone-acquired photographs). The application of ML facilitates early imaging-based screening, automates measurement processes, and supports therapeutic decision-making by predicting the risk of curve progression, thereby enabling timely interventions or minimizing unnecessary radiation exposure. While these technologies offer promising avenues for personalized management of adolescent idiopathic scoliosis, several limitations persist. These include issues related to model interpretability (the so-called “black box” problem), the need for larger and more diverse multi-center validation studies, and the necessity of addressing key ethical concerns, particularly those involving patient data privacy.

The systematic review recently published by Zhu Y et al. (10) investigate DL algorithms for automated Cobb angle measurement on X-rays. The study aimed to provide an overview of DL algorithms, identify their limitations, and summarize possible solutions for improving their performance in scoliosis assessment. The Cobb angle is the gold standard for scoliosis diagnosis; however, manual measurement is time-consuming and prone to observer errors, highlighting the need for efficient automated tools. The researchers conducted a comprehensive literature search across six databases, ultimately including 50 studies in the systematic review and 17 in the meta-analysis. The meta-analysis, which primarily used the circular mean absolute error (CMAE) as the evaluation metric, reported an overall CMAE estimate of 2.99 (95% CI: 2.61–3.38). A significant finding was that segmentation-based methods demonstrated greater accuracy (CMAE 2.40) compared to landmark-based methods (CMAE 3.31). Despite high heterogeneity among the included studies, DL algorithms showed relatively high accuracy in automated Cobb angle measurement. The study also summarized potential strategies for improving model design in future research, such as using two-stage models, combining segmentation and landmark methods, or incorporating multi-view X-ray information. However, despite their accuracy, automated Cobb angle estimation methods are not yet widely adopted in routine clinical practice. The authors conclude that DL algorithms are promising, with segmentation-based methods potentially outperforming landmark-based approaches but emphasize the need for standardization to enable broader clinical adoption.

Table 3 summarized the years screened, articles included and scope of each of the five reviews.

This study possesses multiple limitations that must be acknowledged when evaluating its results.

This study was performed as a quick review, which necessarily entails methodological simplification relative to a comprehensive systematic review. While the PRISMA-RR standards were adhered to and essential databases were examined, restricting the search to three databases and English-language publications may have led to the exclusion of pertinent research, thereby creating selection bias. This review exclusively examined published reviews, including systematic reviews, meta-analyses, narrative reviews, and scoping reviews, rather than primary studies. Thus, the results rely on the methodological rigor and reporting requirements of the incorporated reviews. The AMSTAR 2 evaluation indicated variety in quality, with numerous reviews lacking established processes, thorough risk-of-bias assessments, or sufficient consideration of heterogeneity (4).

Third, significant variation was noted among the included evaluations regarding artificial intelligence models, data sources, input factors, outcome measures, and performance indicators. This variability hinders direct comparisons among studies and prevents rigorous quantitative synthesis, therefore limiting the generalizability of the results. Moreover, several AI models discussed in the provided reviews lacked external validation with independent datasets. Much research relied on retrospective, single-center data, frequently characterized by small sample numbers, which heightens the danger of overestimating model efficacy and constrains real-world clinical relevance (14).

A significant restriction is the irregular incorporation of essential clinical data, including indications of skeletal maturity and psychological aspects, into predictive models. Moreover, numerous deep learning methodologies operate as “black box” systems with restricted interpretability, thereby undermining clinician trust and obstructing integration into collaborative clinical decision-making.

The lack of standardized clinical implementation guidelines, restricted interoperability with current health information systems, and insufficient evidence concerning clinical impact and cost-effectiveness continue to pose substantial obstacles to the integration of AI-based tools into standard care for adolescent idiopathic scoliosis (15).

Despite the growing body of literature on the application of artificial intelligence (AI) in scoliosis care, several critical limitations persist, currently impeding its routine clinical implementation. One of the foremost concerns is accuracy. While some AI models demonstrate promising performance in tasks such as Cobb angle estimation or vertebral landmark detection, many still lack the level of precision required for safe and effective clinical decision-making (16). This raises the risk of misdiagnoses or diagnostic omissions, particularly in borderline or complex cases.

In our rapid review, we identified several key issues across studies: insufficient external validation, absence of clinical implementation guidelines, limited incorporation of skeletal maturity indicators, and the opaque nature of many AI models.

Looking ahead, several priorities must be addressed to enable the responsible and effective integration of AI into scoliosis management. Chief among these is the need to improve model generalizability. This will require the development of large, diverse, and multi-center datasets that reflect the full clinical spectrum of scoliosis, including varying curve patterns, severities, and patient demographics (16).

Equally important is the rigorous validation of AI models across independent datasets and real-world clinical environments. Future research should emphasize interdisciplinary collaboration among spine specialists, data scientists, software engineers, and ethicists.

Multimodal data integration represents another critical research frontier. Incorporating imaging data alongside electronic health records, free-text clinical notes via natural language processing, and wearable sensor outputs can create more holistic and personalized AI models. Additionally, mechanisms for dynamic model updating will be necessary to ensure continued accuracy as clinical practices evolve.