Machine Learning–based Identification of Eyes With a History of Previous Myopic Laser Vision Correction

Purpose: To develop a machine learning model for identifying eyes with previous myopic laser vision correction (LVC). Methods: A total of 41,508 (35,269 eyes in 20,399 patients) and 5,517 (5,517 eyes in 3,210 patients) examinations performed at Seoul National University Bundang Hospital (SNUBH) and Seoul National University Hospital (SNUH), respectively, were included in this study. Using the SNUBH dataset, a machine learning model for identifying eyes with previous myopic LVC (ML-LVC) was developed. The SNUBH and SNUH datasets, respectively, were used as internal and external validation sets to investigate the performance of the ML-LVC model and to compare it with that of the Cooke–Riaz–Wendelstein (CRW1) index. Results: The areas under the receiver operating curve of the ML-LVC model were 0.9970 (95% confidence interval [CI]: 0.9957 to 0.9981) and 0.9960 (95% CI: 0.9929 to 0.9982) in the internal and external validation sets, respectively. The accuracies of the ML-LVC model were 0.9908 (95% CI: 0.9886 to 0.9929) and 0.9927 (95% CI: 0.9906 to 0.9949) in the internal and external validation sets, respectively. The sensitivities of the ML-LVC model were 0.9528 (95% CI: 0.9342 to 0.9691) and 0.9222 (95% CI: 0.8802 to 0.9600) in the internal and external validation sets, respectively. The ML-LVC model exhibited significantly better prediction results, compared with the CRW1 index, in both the internal and external validation sets (McNemar's test, P = .004, P = .005, respectively). Conclusions: The ML-LVC model showed better performance in identifying eyes with myopic LVC, compared with the CRW1 index. This model will benefit many ophthalmologists and is currently available on a web application.