Machine Learning Prediction Of Surgically Induced Astigmatism For Toric Intraocular Lens Power Calculations

The prediction of surgically-induced astigmatism (SIA) affects the accuracy of toric intraocular lens power calculations and therefore postoperative refractive outcomes. Currently, the use of a centroid SIA value is recommended. However, SIA varies with factors including the magnitude of preoperative corneal astigmatism and location of the incision relative to the steep axis. This study aimed to assess if a machine learning-predicted (MLP) SIA value will (i) more accurately predict the achieved SIA value, compared to a centroid SIA value, and (ii) improve refractive outcomes.

Cathedral Eye Clinic, Belfast, United Kingdom

63 right eyes undergoing cataract surgery with implantation of a non-toric monofocal IOL were included in this study. Patients underwent preoperative biometry using swept-source optical coherence tomography. A 2.5mm superior clear corneal incision was used. Incision axis was verified using intraoperative computer-aided alignment, or postoperative measurement at the slit-lamp. At 1 month postoperatively, keratometry was repeated and a manifest refraction was recorded. The Barrett Toric Calculator v2.0 was used to retrospectively calculate the predicted residual refractive astigmatism using (i) a centroid SIA value or (ii) a MLP SIA value. Refractive astigmatism and SIA prediction errors were calculated using vector analysis.

Mean preoperative corneal astigmatism was 1.02D (range 0.13 – 3.46D). The centroid SIA value for these eyes was 0.54D @ 16^o + 0.68D. MLP SIA values were generated for each eye. SIA prediction error was 0.20D @ 50^o + 0.65D using a centroid value vs 0.37D @ 92^o + 0.56D with a MLP value (p = 0.04). Refractive astigmatism prediction error was 0.15D @ 37^o + 1.04D using a centroid value vs 0.25D @ 0^o + 1.59D using a MLP value (p < 0.01). The percentage of eyes within 0.25, 0.50, 0.75 and 1.00D of target postoperative cylinder were 31.7, 58.7, 81.0 and 92.1% for a centroid value and 34.9, 63.5, 82.5 and 92.1% for a MLP value.

This study demonstrated that the use of machine learning-predicted SIA values in toric IOL power calculations may improve clinical outcomes, with a higher percentage of eyes within 0.25 and 0.50D of predicted postoperative cylinder. It may therefore be a valid alternative to inputting a centroid SIA value for toric IOL calculations. The greater centroid SIA prediction error produced by the machine-learning prediction may be due to errors for outliers where the model is less robust, and the model is likely to be improved when trained on a larger dataset. This will be the aim of future studies.