Assessing the correlation between swept-source optical coherence tomography lens density pattern analysis and...

To assess linear correlation between swept-source optical coherence tomography (SS-OCT) lens density variation and patients' best-corrected visual acuity (BCVA).

Linear densitometry was performed on horizontal lens images from 518 eyes, obtained using SS-OCT. All densities from the anterior to the posterior side of the cataract were exported for detailed analysis. The algorithm used a classical random forest regression machine learning approach with fourfold cross-validation, meaning four batches of data from 75% of the eyes with known preoperative best-corrected visual acuity (poBCVA) were used for training a model to predict the data from the remaining 25% of the eyes. The main judgement criterion was the ability of the algorithm to identify linear correlation between measured and predicted BCVA.

A significant linear correlation between poBCVA and the algorithm's prediction was found, with Pearson correlation coefficient (R)=0.558 (95% CI: 0.496 to 0.615, p<0.001). Mean BCVA prediction error was 0.0965±0.059 logarithm of the minimal angle of resolution (logMAR), with 312 eyes (58%) having a BCVA prediction correct to ±0.1 logMAR. The best algorithm performances were achieved for 0.20 logMAR, with 79%±0.1 logMAR correct prediction. Mean, anterior cortex, nucleus and posterior cortex pixel density were all not correlated with patient BCVA.

Pixel density variations based on axial lens images provided by SS-OCT biometer provide reasonably accurate information for machine learning analysis to estimate patient BCVA in all types of cataracts. This study demonstrates significant linear correlation between patients' poBCVA and the algorithmic prediction, with acceptable mean prediction error.