# Predictive Correction Model for Corneal Back Surface Astigmatism With IOLMaster700 Keratometry Data in a Cataractous Population

## TL;DR

This study develops and compares models to predict corneal astigmatism measurements from pre-surgery data, aiming to improve cataract surgery outcomes.

## Contribution

The paper introduces and evaluates multiple predictive models for corneal back surface astigmatism using IOLMaster 700 data.

## Key findings

- Correction models reduced prediction error by 40-42% compared to uncorrected data.
- Segmented and linear models performed slightly better than global and constant models.
- A simple global constant model is recommended for practical implementation due to similar performance across models.

## Abstract

To develop and validate various models to predict total keratometry (TK) power vector components TKC0 and TKC45 from classical keratometry (K) KC0 and KC45 based on a large dataset of pre cataract surgery IOLMaster 700 measurements.

Retrospective cross‐sectional multicentric study evaluating a dataset containing 13 6378 IOLMaster 700 measurements including K and TK. Left eyes were mirrored about the facial axis. Based on 80% training data, we developed a global and segmented constant model (CM and CMS), a global and segmented (according to the angle A1 of the flat keratometric meridian) linear model (LM and LMS), a harmonic model (HM) and compared these to a classical constant (CMR) and linear models (LMR) segmented into with‐the‐rule, against‐the‐rule and oblique astigmatism. The performance was cross‐validated using the root‐mean‐squared model fit error (RMSE).

In the 20% test data, RMSE was 0.173 D before correction and was reduced by 40%–42% to 0.100 and 0.104 D with the correction models. The segmented models performed slightly better than the global models, and the linear models performed slightly better than the constant models. With the individually adjusted changepoints, the CMS and LMS performed slightly better than the reference models CMR and LMR. There was no systematic difference between the RMSE with training and test data, indicating no overfit of the models.

As the performance is quite similar for all tested correction models, we recommend using a simple global constant model to predict TK vector components. This could easily be implemented in any consumer software.