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

introductionAccurate prediction of postoperative vault height following implantable collamer lens (ICL) V4c implantation is critical for minimizing complications and achieving optimal surgical outcomes. This study aims to evaluate the performance of machine learning models in predicting postoperative vault height, focusing on both regression and classification approaches.

methodsThis retrospective study analyzed biometric and demographic data, including anterior chamber depth, white-to-white distance, and ICL size, among other variables. Regression and classification models were developed using gradient boosting, random forest, and CatBoost algorithms. Regression models predicted vault height as a continuous variable, while classification models categorized vault heights into binary and multi-class tasks. Model performance was evaluated using metrics including the mean absolute error (MAE) and root mean squared error (RMSE) for regression models, while accuracy, F1-score, and area under the curve (AUC) were used for classification models.

resultsRegression models demonstrated moderate predictive performance, with random forest delivering the best performance (MAE: 134.0 µm, RMSE: 171.3 µm, Pearson's correlation coefficient: 0.45). On the other hand, classification models exhibited greater clinical applicability than regression approaches. For the binary classification task (vault < 250 µm vs. ≥ 250 µm), gradient boosting achieved the highest overall performance, with accuracy of 89 ± 12% and an AUC of 0.89. In the task of predicting vault > 750 µm vs. ≤ 750 µm, random forest emerged as the most effective classifier, achieving accuracy of 86 ± 9% and an AUC of 0.88. In multi-class classification, models demonstrated superior performance in predicting intermediate vault heights (250 µm ≤ vault ≤ 750 µm). Random forest achieved the highest accuracy of 94.6% in this category. However, all models faced challenges in accurately classifying extreme vault categories.

conclusionsClassification models, particularly gradient boosting and random forest, demonstrated strong potential for predicting clinically significant vault categories, enabling personalized surgical planning and improved risk management. While regression models offered moderate insights, their limitations suggest that classification approaches are better suited for clinical applications. Future research should focus on enhancing model accuracy for extreme vault prediction and integrating advanced techniques, such as ensemble deep learning, to further refine outcomes.