gms | German Medical Science

Introduction: Corneal cross-linking (CXL) is the primary treatment for halting the progression of keratoconus (KC). CXL efficiency can vary among individuals and for optimal risk management a personalized risk profile to be addressed. The aim of the study was to combine preoperative diagnostic findings to develop an Artificial Intelligence (AI) model that would accurately predict CXL outcomes in KC patients. To achieve a reliable risk assessment model; the following were performed:

1.

analyze the dataset and assess an impact of preoperative keratometry findings on corneal power after CXL,

2.

explore association of CXL outcomes with results of visiorefractometry, pachymetry and topography data,

3.

model changes in the flat, steep and maximal corneal power after CXL,

4.

find the most significant predictors of CXL effectiveness in KC patients.

Methods: This is a retrospective study of 107 patients (112 eyes; 796 observations) treated with CXL for KC (79 males, 28 females). The exclusion criteria were retreatment. The participants were followed up to 40 (12.93±4.75) months post CXL. Clinical assessment, pachymetry, refraction, keratometry, and topography were recorded. Data for analysis were derived from topographic, topometric, and Belin-Ambrósio enhanced ectasia display (BAD) maps. The following methods were applied: Kruskal Wallis for Ophtalmometric distribution; Pearson and Spearman for variables correlation and Bayesian Information Criterion to develop the AI learning algorithms to predict CXL outcomes.

Results: Postoperative maximal cornea power is positively associated with preoperative cornea eccentricity, central keratoconus index and BAD-D (r=0.65, 0.81 and 0.84). The Kmax negatively correlates with preoperative minimal corneal thickness and best-corrected visual acuity (r=-0.59 and -0.57). Linear models of postoperative changes in corneal power outperform the polynomial ones (4.05±6.88 vs 4.1±6.79% MAE/ROV, p=0.051). A quadratic function depicts a sustainable Kmax reduction for 2 years, a subsequent 4-month-long plateau, and a rise in the maximal corneal power afterward. The effect of CXL is lost beyond 5 years after the treatment. The AI-model accurately predict changes in corneal curvature with 6.3% ratio of mean absolute error to the range of values.

Conclusion: The findings advocate for personalized approach to candidate selection for CXL. AI models of CXL outcomes may quickly and effectively perform risk stratification for selection of optimal treatment.