Prediction models for curve progression in adolescent idiopathic scoliosis: a systematic review with exploratory meta-analysis of discrimination.

BACKGROUND

Predicting curve progression is central to adolescent idiopathic scoliosis (AIS) care, but the robustness and clinical readiness of existing prediction models remain uncertain.

OBJECTIVE

To review AIS progression models across clinically distinct prediction tasks and treatment contexts and, where operational eligibility criteria were met, summarise discrimination using metric-stratified exploratory meta-analyses.

METHODS

PubMed, Web of Science and Embase were searched to 26 April 2026. Full-text studies developing, validating, updating or evaluating prognostic models for AIS progression were eligible. Data extraction was informed by CHARMS, reporting by TRIPOD, and study-level risk of bias was assessed using PROBAST. Binary-outcome AUCs and survival/time-to-event C-statistics were synthesised separately using random-effects models. A combined cross-metric estimate was examined only as a secondary sensitivity analysis.

RESULTS

Forty-eight studies reporting 63 models were included. Models ranged from regression, risk scores and nomograms to machine learning, deep learning, radiomics and multimodal approaches. Thirty-one models reported an AUC or C-statistic, but only 17 provided usable variance information. Six publication-level estimates contributed to the quantitative synthesis. Three binary-outcome AUCs yielded a pooled estimate of 0.794 (95% CI 0.726-0.849; I²=93.9%), whereas three survival/time-to-event C-statistics yielded 0.883 (95% CI 0.850-0.909; I²=0.0%). The two estimates were not directly comparable because the underlying tasks, treatment contexts and outcome structures differed. The secondary combined estimate was 0.839 (95% CI 0.788-0.879; I²=91.8%) and had no single clinical interpretation. Calibration, clinical utility, external validation and implementation reporting were limited. Forty-four of 48 studies (91.7%) were rated at high overall risk of bias.

CONCLUSIONS

AIS prediction evidence is best interpreted through six clinically distinct prediction tasks. Transportability is constrained by treatment context, outcome definition, skeletal maturity and validation design. The pooled estimates were derived from small selected subsets and should be interpreted as exploratory findings secondary to the task-specific narrative synthesis. Future studies should standardise prediction targets, report calibration and clinical utility, and undertake prospective external validation within the intended care pathway.