Low Vault Prediction After Implantable Collamer Lens Implantation – From Logistic Regression To Machine Learning Models

The posterior-chamber phakic Implantable Collamer Lens (ICL) is usually the first choice for the surgical correction of moderate-to-high myopia due to its excellent visual and refractive results. Although it presents a good safety profile, most complications after ICL implantation are related to the postoperative vault. Low vault has been associated with cataract formation and Toric ICL misalignment. The aims of this study are to analyze the influencing factors and to predict the postoperative Vault.

Cornea and Anterior Segment Department, in Hospital de Braga, Braga, Portugal 

A total of 539 eyes submitted to ICL implantation between 2012 and 2022 were included in the study. The eyes were divided into two groups based on their immediate postoperative vault: Group 1 (0 to 250 μm) and Group 2 (>250 μm). Binary Logistic Regression and Machine Learning classification algorithms (Random Forest and Gradient Boosting) were applied to predict the vault. Test accuracy, precision, and mean area under the curve (AUC) were calculated for each model. Permutation importance and Impurity-based feature importance were used to investigate the importance of each input features. 

Group 1 (155 eyes) and Group 2 (384 eyes) had a mean vault of 161.2 ± 65.69μm and 480.43 ± 193.92μm, respectively. The groups differed significantly in age, gender, manifest cylinder, anterior chamber depth (ACD), white-to-white distance, anterior chamber angle (ACA) and number of Toric ICL. In vault prediction, Random Forest yielded 80.0% of Test Accuracy and 0.78 of AUC; followed by Gradient Boosting (76.8% and 0.77) and Logistic Regression (76.7%, 0.80). Random Forest had the best result in precision (69.2%). Logistic Regression demonstrated that increasing of ACD, mean Keratometry and ICL size was associated with a lower probability of low Vault. ACD and ICL size were the most important input features in the classification models.

Deeper ACD and increase of ICL size are associated to less likelihood of obtaining a Low postoperative Vault. Machine learning models may, in the future, assist ophthalmologists in enhancing surgery safety, designing surgical strategies, and predicting clinical outcomes.