# Digital transformation of care for keratoconus patients: ML modeling structural outcomes of corneal collagen cross-linking

## TL;DR

This study uses machine learning to predict structural outcomes of corneal collagen cross-linking surgery for keratoconus patients.

## Contribution

A novel ML-based approach is introduced to assess and predict corneal thickness changes after CXL using multimodal preoperative diagnostics.

## Key findings

- Postoperative minimal corneal thickness (MCT) is a more meaningful marker of corneal thinning than central corneal thickness (CCT).
- Baseline pachymetry data and topographic indices strongly correlate with postoperative outcomes.
- Multimodal preoperative diagnostics improve risk assessment and prognosis accuracy.

## Abstract

Structural outcomes of corneal collagen cross-linking (CXL) have not been thoroughly investigated. Clinical risk assessment would benefit from a reliable prognosis of postoperative minimal (MCT) and central corneal thickness (CCT).

The objective of this study was to find a combination of diagnostic modalities and measurements that reliably reflect CXL efficiency in terms of corneal thickness.

We retrospectively reviewed the medical histories of 107 patients (131 eyes) who underwent CXL. The dataset included preoperative examinations and follow-up results, which totalled 796 observations.

The postoperative changes in MCT are more pronounced, clinically relevant, and meaningful than in CCT. MCT should serve as the major clinical marker of corneal thinning after CXL. The cornea's potential to recover reduces in advanced keratoconus. A polynomial curve demonstrates the natural course of corneal remodeling. It includes thinning immediately after CXL and stabilization with partial recovery of corneal thickness over time. Baseline pachymetry data can adequately reflect the outcomes. Preoperative BAD and topographic indices strongly correlate with the outcomes. Keratometry and refractometry data exhibit moderate associations with postoperative corneal thickness. The models trained on a combination of top correlating features, clinical data, and time after intervention provide the most reliable prognosis.

Risk assessment is accurate with multimodal preoperative diagnostics. A stratification system should take into account findings in different diagnostic modalities.