Using Partial Least Squares Regression And Random Forest Estimation For Intraocular Lens Power Calculation

Published 2024 - 42nd Congress of the ESCRS

Reference: FP06.03 | Type: Free paper | DOI: 10.82333/xagk-7v35

Authors: Vinzenz Rudnay* ¹ , Lorenz Vock ¹ , Armin Ettl ¹ , Lisa Tasch ² , Leon Pomberger ² , Matthias Bolz ² , Nino Hirnschall ²

¹Ophthalmology and orbital surgery ,University clinic St.Poelten ,St.Poelten,Austria, ²Ophthalmology,Kepler university clinic,Linz,Austria

Purpose

To use a combination of partial least squares regression and random forest estimation for intraocular lens position prediction.

Setting

Prospective study.

Methods

This prospective study observed patients undergoing cataract surgery within the last two years prior to recruitment. Axial length (AL), corneal radii (K1 and K2), anterior chamber depth (ACD), lens thickness (LT), aqueous depth (AQD), central corneal thickness (CCT), white-to-white (WTW) distance and Chang-Waring chord (CW) were measured pre- and postoperatively using swept-source optical coherence tomography-based biometry. Additionally, postoperative refraction was measured objectively using an autorefractor-keratometer as well as subjectively using the Jackson cross-cylinder method. A combination of partial least squares regression and random forest estimation was used the predict the post-operative IOL position.

Results

A total of 73 patients were observed pre- and post-operatively. Their AL ranged from 21.3 mm to 27.6 mm. The post-operative mean spherical equivalent was -0.36 ± 0.52 D with a maximal refractive error of 1.75 D. The best corrected distance visual acuity (BCDVA) ranged from 0.63 to 1.25 with an average of 1.0. The mean absolute error of the calculated postoperative refraction with different IOL-power formulae were 0.614 ± 0.628 D (Haigis), 0.537 ± 0.549 D (HofferQ) and 0.495 ± 0.522 D (SRK/T). The combination of PLSR and AI had a mean absolute error of 0.018 ± 0.41 D.

Conclusions

A combination of partial least squares regression and random forest estimation was found to be useful for IOL position prediction. However, even an improved IOL position prediction cannot completely eliminate the risk of refractive surprises.