Abstract
PURPOSE: This study aims to predict postoperative toric implantable collamer lens (TICL) rotation and its effect on refraction and vision using artificial intelligence (AI) models. METHODS: Data from 642 eyes from 371 patients undergoing TICL surgery were included. Regression models predicted rotation degrees, and classification models predicted rotation-related residual astigmatism, visual acuity loss, or the need for realignment surgery. Regression tasks were evaluated using root mean square error (RMSE) and coefficient of determination (R²), while classification tasks were assessed by accuracy and mean area under the curve (AUC). Subgroup analyses were conducted for low, medium, and high astigmatism. The cutoff value for rotation affecting astigmatism or visual acuity was determined using receiver operating characteristic curves. RESULTS: Tabular prior-data fitted network (TabPFN) was the most accurate regression model for predicting postoperative rotation, achieving RMSE (10.672 ± 5.880) and mean absolute error (5.643 ± 2.328) compared to traditional models. For predicting rotation-related complications, TabPFN consistently achieved the highest accuracy across all secondary outcomes (0.906-0.981), particularly excelling in realignment surgery prediction with near-perfect precision (0.990 ± 0.006) and the highest AUC (0.900 ± 0.078). Cutoff values were 7.50° (AUC = 0.65), 4.50° (AUC = 0.68), and 2.50° (AUC = 0.73) for residual astigmatism and 9.50° (AUC = 0.72), 7.50° (AUC = 0.80), and 4.50° (AUC = 0.93) for visual acuity loss in low, medium, and high astigmatism groups, respectively. CONCLUSIONS: AI models effectively predict postoperative rotation stability, providing valuable references for ophthalmologists. TRANSLATIONAL RELEVANCE: This study investigated the quantitative relationship between rotation, astigmatism, and vision, bridging the gap between artificial intelligence and optical theory.