Multimodal Predictive Modeling for Visual Quality Recovery After Keratorefractive Lenticule Extraction

To investigate the factors influencing visual quality recovery after keratorefractive lenticule extraction (KLEx) and to establish a predictive model for postoperative visual quality using multimodal data integration.

A prospective, non-interventional single-center study was conducted on 210 eyes from 105 myopic patients undergoing KLEx surgery. Preoperative and postoperative visual quality was assessed using the Objective Scatter Index (OSI). High-precision segmentation of posterior lenticule scanning images was performed using a U-net-based deep learning model to extract radiomics features. These features, along with clinical data, surgical parameters, corneal topography, and biomechanics, were integrated into machine learning models for predicting OSI at 1 day and 1 week postoperatively.

The study found that 30.95% (65/210) of eyes experienced poor early visual recovery. Significant differences in spherical equivalent, corneal curvature, and cutting depth were observed between good and poor visual quality groups. The ExtraTrees model demonstrated high efficacy in predicting OSI on the first postoperative day (area under the curve [AUC] = 0.96), whereas the AdaBoost model excelled at 1 week (AUC = 0.99). Regression models RidgeCV and LassoLarsIC performed best for predicting OSI values at 1 day and 1 week, respectively. Radiomics features such as original topology Gaussian Curvature Min and original gldm Dependence Variance were significantly correlated with the OSI.

The study establishes a predictive framework that integrates intraoperative scanning image features, corneal biomechanics, and densitometry parameters, offering a quantitative basis for personalized surgical planning. The synergistic use of radiomics and clinical parameters can pre-identify high-risk patients, optimize surgical settings, and improve postoperative visual quality.