<![CDATA[Introduction: Keratoconus is a degenerative corneal disorder leading to vision impairment. It is important to detect it early to prevent its progression by corneal cross-linking (CXL). Keratoconus is diagnosed using videokeratography and Scheimpflug tomography, which provide valuable data on the corneal surface. However, distinguishing keratoconus from normal variations remains challenging. Recent advances in artificial intelligence (AI) offer promising improvements in detecting subtle corneal changes, enhancing keratoconus detection and diagnosis. Purpose: To analyze AI as a diagnostic modality for keratoconus by calculating the pooled sensitivity and specificity to evaluate its accuracy. Methods: Databases involved PubMed, Scopus, Google Scholar, Embase, and Science Direct, from 2018 to March 2024. Also, to include unpublished works, the grey literature was searched, using the OpenGrey repository. Studies were included when they met the inclusion criteria. Results: We involved a total of 19 studies in this meta-analysis. The pooled sensitivity for detecting keratoconus was 95% confidence interval (CI) (91% to 98%), with a pooled specificity of 98% CI (96% to 99%). Additionally, the random forest model had a pooled sensitivity of 98.11% (CI, 96.77% to 99.44%), with a pooled specificity of 99% (CI, 98.24% to 99.76%). On the other hand, the convolutional neural network (CNN) model had a pooled sensitivity of 89.73% CI (79.77% to 99.69%), with a pooled specificity of 95.27% CI (91.88% to 98.66%). Conclusion: The results confirmed the reliability of different AI models in diagnosing keratoconus, especially the random forest model. This is important, as the early and accurate detection of keratoconus provides opportunities to reduce risk factors and offer treatments, including CXL, which can potentially slow its progression and improve the patient’s quality of life.]]>
Copyrights © 2025
