Artificial intelligence applications in refractive error management: A systematic review and meta-analysis.
systematic_review · Level I
Where this comes from
- Record sourced from PubMed, PMID 40997051.
- Also identified by DOI 10.1371/journal.pdig.0000904 and PMC identifier 12463214.
- Licence recorded as CC BY.
- The licence permits redistribution, so the abstract is shown in full and the full text is available from the publisher.
Abstract
Artificial intelligence (AI) has transformed healthcare, and is becoming increasingly useful in eye care. We conducted a systematic review and meta-analysis of the use of AI in the diagnosis, detection, prediction, progression, and treatment of refractive errors (REs). The study adhered to the PRISMA checklist to ensure transparent reporting. The following databases were searched from inception to January 2025, with an English language restriction: PubMed, Web of Science, Embase, Scopus, Cochrane Library and Google Scholar. Two independent reviewers performed study screening, data extraction, and quality assessment, with a third author resolving discrepancies. All original studies on the use of AI techniques in RE were identified and the effectiveness of these techniques was compared. A critical appraisal was conducted using the QUADAS-2 risk-of-bias tool. A meta-analysis was performed using R software (version 4.5.0). Of 6,288 records retrieved, 45 met eligibility for systematic review, with 19 included in meta-analysis. Among these 45 studies, 55.5% (25/45) applied deep learning (DL) approaches, while 44.4% (20/45) employed machine learning (ML) techniques. The pooled sensitivity, specificity, diagnostic odds ratio (DOR), and summary of receiver operating characteristic (SROC) for detection and/or diagnosis studies were 0.94 (95%CI, 0.90-0.97), 0.96 (95%CI, 0.92-0.98), 382.56 (95% CI 111.91 -1307.77) and 0.98 (95%CI, 0.91-0.97), respectively. For prediction of REs, the pooled sensitivity, specificity, DOR, and SROC were 0.87 (95%CI, 0.73-0.94), 0.96 (95%CI, 0.90-0.980), 159.94 (95% CI, 40.17-636.85) and 0.96 (95%CI, 0.85-0.95), respectively. Among studies focused on progression, performance metrics ranged from AUC = 0.845-0.99, R² = 0.613-0.964, and MAE = 0.119D-0.49D. In treatment studies, performance varied more widely, with AUC values between 0.60-0.94 and MAE from 0.17D-0.54D. Collectively, AI technologies, particularly DL and ML, achieved high diagnostic and predictive accuracy in RE management. Future research should focus on developing generalizable models trained on diverse datasets to ensure broad clinical relevance.