Abstract
PURPOSE: To develop and validate a machine learning system for the preoperative prediction of small effective optical zone (EOZ; diameter < 5.5 mm) after small incision lenticule extraction (SMILE). METHODS: In this multicenter cohort study, 1030 multimodal combinations of preoperative parameters (PP), anterior corneal curvature maps (AACM), surgery video frames, and three-month postoperative EOZ diameter from 1030 eyes (634 patients) undergoing SMILE were divided: 677 for training, 85 for primary validation, 85 for internal test, and 183 for external test. The AACM-PP-Model integrating AACM and PP was developed and compared against parameter-only or image-only models, with primary performance evaluated by the area under the receiver operating characteristic curve (AUROC) and macro F1 score. RESULTS: In the primary validation set, the AACM-PP-Model achieved an AUROC of 0.897 (95% confidence interval [CI], 0.819-0.975) and macro F1 of 0.823. It outperformed the best parameter-only model (PP-CatBoost, AUROC of 0.885 [95% CI, 0.813-0.957], macro F1 of 0.821) and best image-only model (model based on suction-initiated, centering-checked frames and anterior axial curvature map [SIF-CCF-AACM-Model], AUROC of 0.837 [95% CI, 0.745-0.928], macro F1 of 0.612). Superiority of AACM-PP-Model over PP-CatBoost and SIF-CCF-AACM-Model was maintained in internal (AUROC 0.957 vs. 0.915 vs. 0.773) and external (0.863 vs. 0.814 vs. 0.660) test sets. Intraoperative frame models performed poorly (AUROC < 0.7 on all test sets). CONCLUSIONS: Multimodal machine learning using standard preoperative data can accurately predict small post-SMILE EOZ, demonstrating superior generalization to support preoperative decision-making. TRANSLATIONAL RELEVANCE: Objective preoperative risk stratification could aid in personalized surgical planning and help manage patient expectations regarding postoperative visual quality.