Clinical Interchangeability of Total Corneal Astigmatism Prediction Methods: An Agreement Analysis

Purpose: To compare agreement between total corneal astigmatism (TCA) predictions from a deep learning (DL) model and those from the Barrett Toric Calculator (Barrett) and Abulafia–Koch (AK) formulas. Methods: In 100 eyes with corneal astigmatism, Pentacam HR imaging (Oculus Optikgeräte GmbH) was used to generate TCA predictions with the AK formula, Barrett formula, and a previously validated DL model. Agreement was assessed using Pearson and Spearman correlation, mean absolute error (MAE), root mean square error (RMSE), and Bland-Altman analysis. Clinical interchangeability was predefined as 95% limits of agreement within ±0.25 diopters (D) for magnitude and ±5 degrees for axis. Predictions were also decomposed into Cartesian X and Y vector components. Results: The AK and Barrett formulas demonstrated near-perfect agreement for both magnitude ( r = 0.99; MAE: 0.070 D, RMSE: 0.11 D) and axis ( r = 1.00; MAE: 1.11°, RMSE: 1.68°), with 95% limits of agreement both within the interchangeability thresholds. Compared with both conventional methods, the DL model showed larger systematic differences. Compared with the Barrett formula, magnitude agreement was r = 0.92 (MAE: 0.24 D; RMSE: 0.27 D), and axis agreement was r = 1.00 (MAE: 4.71 degrees). Similar findings were observed versus the AK formula (magnitude r = 0.94, MAE: 0.20 D; axis MAE: 4.92°). The DL model consistently underestimated X-components relative to both traditional methods (RMSE: 0.28 to 0.31 D), whereas Y-component agreement was excellent across all pairs (RMSE ⩽ 0.05 D). DL axis differences followed a sinusoidal pattern with astigmatism orientation ( R 2 = 0.84 to 0.87), peaking at oblique meridians. Conclusions: The AK and Barrett formulas are clinically interchangeable for TCA prediction, with limits of agreement falling entirely within established toric intraocular lens planning thresholds. The DL model exhibits systematic, yet consistent deviations from both traditional methods that demonstrate lack of interchangeability. Prospective studies comparing postoperative refractive outcomes for each method are required to determine whether the DL model's deviations represent superior predictive performance.