Optimizing IOL Calculators with Deep Learning Prediction of Total Corneal Astigmatism

Background/Objectives: This study aims to identify the most accurate regression model for predicting total corneal astigmatism (TCA) from anterior corneal astigmatism (ACA) and to fine-tune the best model's architecture to further optimize predictive accuracy. Methods: A retrospective review of 19,468 eyes screened for refractive surgery was conducted using electronic medical records. Corneal topography data were acquired using the Pentacam HR. Various types (7) and subtypes (21) of regression learners were tested, with a deep neural network (DNN) emerging as the most suitable. The DNN was further refined, experimenting with 23 different architectures. Model performance was evaluated using root mean squared error (RMSE), R², average residual error, and circular error. The final model only used age, ACA magnitude, and ACA axis to predict TCA magnitude and axis. Results were compared to predictions from one of the leading TCA prediction formulas. Results: Our model achieved higher performance for TCA magnitude prediction (R² = 0.9740, RMSE = 0.0963 D, and average residual error = 0.0733 D) compared to the leading formula (R² = 0.8590, RMSE = 0.2257 D, and average residual error = 0.1928 D). Axis prediction error also improved by an average of 8.1° (average axis prediction error = 4.74° versus 12.8°). The deep learning approach consistently demonstrated smaller errors and tighter clustering around actual values compared to the traditional formula. Conclusion: Deep learning techniques significantly outperformed traditional methods for TCA prediction accuracy using the Pentacam HR. This approach may lead to more precise TCA calculations and better IOL selection, potentially enhancing surgical outcomes.