Small-incision lenticule extraction (SMILE) is a safe and effective surgical procedure for refractive correction. However, the nomogram from the VisuMax femtosecond laser system often overestimates the achieved lenticule thickness (LT), leading to inaccurate estimation of residual central corneal thickness in some patients. In order to improve the accuracy of predicting achieved LT, we used machine learning models to make predictions of LT and analyze the influencing factors of LT estimation in this study. We collected nine variables of 302 eyes and their LT results as input variables. The input variables included age, sex, mean K reading of anterior corneal surface, lenticule diameter, preoperative CCT, axial length, the eccentricity of the anterior corneal surface (E), diopter of spherical, and diopter of the cylinder. Multiple linear regression and several machine learning algorithms were employed in developing the models for predicting LT. According to the evaluation results, the Random Forest (RF) model achieved the highest performance in predicting the LT with an R2 of 0.95 and found the importance of CCT and E in predicting LT. To validate the effectiveness of the RF model, we selected additional 50 eyes for testing. Results showed that the nomogram overestimated LT by 19.59% on average, while the RF model underestimated LT by -0.15%. In conclusion, this study can provide efficient technical support for the accurate estimation of LT in SMILE.
Keywords: Random Forest; SMILE; lenticule thickness; machine learning.