Exploring Machine Learning Models for Vault Safety in ICL Implantation: A Comparative Analysis of Regression and...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accurate prediction of postoperative vault height after implantable collamer lens (ICL) V4c implantation is crucial for preventing complications and ensuring optimal surgical results.
• This study evaluates machine learning (ML) models for predicting ICL vault height, comparing regression and classification approaches.

Purpose of the Study:

• To assess the performance of various ML algorithms in predicting postoperative vault height following ICL V4c implantation.
• To determine the efficacy of regression versus classification ML models for clinical application in ICL surgery.

Main Methods:

• Retrospective analysis of biometric and demographic data relevant to ICL implantation.
• Development and evaluation of regression and classification models using gradient boosting, random forest, and CatBoost algorithms.
• Performance assessment using metrics like MAE, RMSE, accuracy, F1-score, and AUC.

Main Results:

• Regression models showed moderate predictive performance (Random Forest: MAE 134.0 µm, RMSE 171.3 µm).
• Classification models demonstrated superior clinical applicability, with Gradient Boosting achieving 89% accuracy (AUC 0.89) for binary vault prediction (<250 µm vs. ≥250 µm).
• Random Forest excelled in binary classification (≥750 µm vs. ≤750 µm) with 86% accuracy (AUC 0.88) and multi-class prediction of intermediate vault heights (94.6% accuracy). Models struggled with extreme vault categories.

Conclusions:

• Classification models, especially Gradient Boosting and Random Forest, show significant potential for predicting clinically relevant ICL vault height categories.
• Classification approaches are better suited for clinical applications compared to regression models due to their higher predictive accuracy in defined vault ranges.
• Future research should aim to improve prediction accuracy for extreme vault heights and explore advanced ML techniques like ensemble deep learning.