Ophthopedia (@ophthalmology.bsky.social)

Ophthopedia Update: Prediction of long-term uncorrected distance visual acuity in surgically SMILE corrected myopic eyes using machine learning: Background This study aimed to create machine learning (ML) models to predict the long-term uncorrected distance visual acuity (UDVA) in myopic eyes…

loading . . .

Prediction of long-term uncorrected distance visual acuity in surgically SMILE corrected myopic eyes using machine learning Background This study aimed to create machine learning (ML) models to predict the long-term uncorrected distance visual acuity (UDVA) in myopic eyes corrected by small incision lenticule extraction (SMILE). Methods This was a retrospective cohort study conducted in Zhongshan Ophthalmic Center, Sun Yat-sen University. Participants who underwent SMILE between 2012 and 2016 were invited for the final follow-up examinations in 2019. Medical records and surgical parameter data were collected for analysis. Multicollinearity analysis and feature importance ranking were used to select the most predictive features. The following algorithms were used: least absolute shrinkage and selection operator, random forest, extremely randomised regression trees (extraTrees), gradient boosting machine and extreme gradient boosting. The root mean square error (RMSE) and mean absolute error (MAE) for each ML model were evaluated. Results In total, 873 eyes from 440 patients with complete records were included in this study. The long-term UDVA (logarithm of the minimum angle of resolution) distribution at the final follow-up ranged from –0.1760 to 0.7960. The extraTrees model outperformed the other four models, with RMSE and MAE of 0.1162 and 0.0850, respectively. Additionally, some features, including spherical equivalent, lenticular optical zone, added manifest refraction, preoperative corrected distance visual acuity and cap thickness, had moderate-to-strong effects on the average UDVA prediction using the extraTrees model. Conclusion Long-term UDVA in myopic eyes corrected by SMILE can be effectively predicted using ML technologies, particularly the extraTrees model. However, more features and samples for the prediction model need to be explored to improve accuracy. Otherwise, there was a limitation in this research that sphere and cylinder refraction were treated as independent variables. But, the proportion of astigmatism to spherical refraction is relatively low, less than 1/5. Consequently, this does not lead to the incorrectness of our results, but they are weakened by this. http://dlvr.it/TNJKYv