Machine Learning-Assisted Algorithm for Parameter Setting in SMILE Surgery

Abstract Objective: In recent years, SMall Incision Lenticule Extraction (SMILE) has demonstrated good safety and effectiveness, among other features, in the treatment of myopia and myopic astigmatism. However, the setting of its surgical parameters is still highly dependent on the surgeon's experience and suffers from insufficient stability and interpretability. Methods: To address this challenge, this study proposes a machine learning-based method for assisted parameter setting in SMILE surgery. This study firstly identifies the patient features that are closely related to the decision-making of surgical parameters based on the ophthalmologists' clinical experience, and subsequently performs two-layer feature selection by Maximum Mutual Information Coefficient (MIC) and Genetic Algorithm, and meanwhile, adopts the SMOTE technique with Cost-Sensitive Learning (CSL) to deal with the problem of data imbalance; selects the five machine learning algorithms, LightGBM, XGBoost, RF, AdaBoost, and BP neural network, were trained and tested under the Optuna hyperparametric framework, and their performances were compared. Results: The experimental results show that LightGBM performs the best with the best accuracy and macro F1 scores of 0.8917 and 0.9018, respectively. In addition, the coefficient of variation (CV) of the prediction probability is used in this study to assess and visualize the uncertainty of the model prediction quantitatively. The importance of the features was also analyzed by SHAP values, and the results of the analysis coincided with the ophthalmologists' clinical experience. Conclusions: The method proposed in this study is expected to improve the stability and interpretability of ophthalmologists' decision-making in the process of setting parameters for myopic refractive surgery.