A systematic review of machine learning models for predicting curve progression in teenagers with idiopathic scoliosis

OBJECTIVE: To summarize the development and validation of machine learning (ML) models using data from the first clinical visit to predict ensuing curve progression in individuals with adolescent idiopathic scoliosis (AIS).

DESIGN: Systematic review.

LITERATURE SEACRH: Five databases were searched from inception to 31 March 2023.

STUDY SELECTION CRITERIA: Studies had to include teenagers (10-17 years) with AIS, and internally/externally validated ML models for predicting curve progression.

DATA SYNTHESIS: We used the CHARMS checklist to guide data extraction, the IJMEDI checklist for quality assessments, and the GRADE system for evidence certainty.

RESULTS: Seventeen studies with 37 ML models (3,701 participants) were included. The models predicted risk, threshold progression, Cobb angles, and curve variations. The IJMEDI checklist identified critical shortcomings in data understanding (60%), preparation (82%), validation (95%), and deployment (95%). Thirty-four ML models showed fair to excellent accuracy, with the area under the Receiver Operating Characteristic curve ranging from 0.70-0.93 in most categories.

CONCLUSION: While most models achieved moderate to good accuracy, the certainty of evidence was very low, rendering most models not clinically deployable and lacking external validation. Therefore, these findings should be interpreted with caution. Future research should address these limitations to improve the clinical applicability and external validity.