Machine learning-based prediction of accelerated corneal collagen cross-linking surgery outcomes

Objective  To use machine learning to predict the efficacy of accelerated corneal collagen cross-linking (A-CXL) surgery, identify prognostic factors, and construct models to predict postoperative disease progression.

Methods  A single-center retrospective study was conducted.A total of 82 keratoconus patients (112 eyes) who underwent A-CXL surgery at the West China Hospital of Sichuan University between March and December 2021 were enrolled.Preoperative and follow-up examinations included anterior segment evaluation by slit-lamp microscopy, corneal topography using Pentacam, and corneal biomechanical indices using Corvis ST.Disease progression was defined as an increase in maximum keratometry (Kmax) of ≥1 D from the preoperative level at the last follow-up.Various machine learning algorithms were employed to analyze corneal topography, biomechanical parameters and corneal densitometry values to identify prognostic factors and construct models for predicting postoperative disease progression.This study adhered to the Declaration of Helsinki.The study protocol was approved by the Ethics Committee of West China Hospital, Sichuan University (No.2023496).Written informed consent was obtained from each subject.

Results  During follow-up, 15.1% (17/112) of the eyes showed progression after A-CXL.The preoperative astigmatism and stress-strain index (SSI) in the progression group were (-5.41±2.72)D and 1.41±0.78, respectively, which were significantly higher than (-3.30±2.54)D and 0.95±0.98 in the non-progression group ( t=2.80, 2.03; both P<0.05).Cox regression analysis identified preoperative astigmatism (hazard ratio [HR]=1.20), SSI (HR=1.10), and anterior corneal densitometry of 2-6 mm (CDA6) (HR=2.10) as significant risk factors for post-A-CXL progression.Among various machine learning models developed and validated, the area under the curve (AUC) values for logistic regression, multilayer perceptron (MLP) model, and random forest (RF) exceeded 0.700.For F1-score, the AUC values for logistic regression, MLP, and RF were 0.870, 0.880, and 0.880, respectively.The network structure of the visualized MLP was a single-layer, 24-neurons neural network with 80% accuracy in predicting whether progression occurred after A-CXL.The clinical nomogram developed in conjunction with astigmatism, SSI, and CDA6 predicted the cumulative probability of progression at 0.5, 1, and 2 years postoperatively based on the sum of the specified values for each variable, and based on the optimal cutoff value, keratoconus corneas could be classified into high-, intermediate-, and low-risk groups, respectively.The time-dependent subject operating characteristic curves of the nomogram showed AUCs of 0.734, 0.685, and 0.935 at 0.5, 1, and 2 years postoperatively, respectively, all of which performed well in predicting progression.

Conclusions  Preoperative astigmatism, SSI, and CDA6 are significant risk factors for post-A-CXL progression in keratoconus.The MLP model can accurately predict postoperative disease progression, and the clinical nomogram combining preoperative astigmatism, SSI, and CDA6 can effectively differentiate between low-, medium-, and high-risk postoperative progression outcomes.

KEYWORDS：