Random Forest-Driven Multi-Feature Machine Learning Model For Predicting Postoperative Vault After Icl Implantation

Postoperative vault is a critical indicator for evaluating the safety and efficacy of Implantable Collamer Lens (ICL) implantation in the posterior chamber of the phakic eye. This study, based on an artificial intelligence model, combines random forest feature selection with machine learning algorithms to construct a multi-time-point postoperative vault prediction system.

In recent years, the random forest algorithm, whose ensemble learning characteristics can efficiently handle high-dimensional data and reveal key influencing factors through feature importance ranking. This study aims to construct a random forest-driven multi-feature model that integrates preoperative ocular parameters,providing a reliable tool for personalized surgical planning and complication prevention.

The random forest model was used to measure and rank the importance of preoperative ocular biological parameters, such as anterior chamber angle A (Pentacam), anterior chamber angle B (UBM), anterior chamber depth (ACD), corneal diameter (White to White, WTW), horizontal sulcus-to-sulcus diameter (STS-H), vertical sulcus-to-sulcus diameter (STS-V), horizontal lens vault (STSL-H), and vertical lens vault (STSL-V). Significant variables affecting the vault were selected, and further prediction models for postoperative vault at 1 week, 1 month, 3 months, and 6 months were established using Multilayer Perceptron (MLP) and Support Vector Regression (SVR). Hyperparameters were optimized through cross-validation.

Random forest feature selection identified STS-H, STS-V,STSL-V as key predictors. The importance ranking from high to low is as follows: STS-V 0.17286488, STS-H 0.1714251, STS-H-WTW 0.13269247, STSL-H 0.11854116, STSL-V 0.1165233, anterior chamber angle (Pentacam) 0.11530968, STSV-STSH 0.11048173, anterior chamber angle (UBM) 0.02571944, WTW 0.02549202, ACD 0.01979203. Among this three models, the random forest had the lowest root mean square error (MSE) (0.040, 0.024, 0.037, 0.043, 0.035,respectively)(P < 0.05). MLP showed lower prediction errors in the early postoperative period (1 w,1 m) (37.02±20.27 μm), while SVR performed better in the long term (3 m, 6 m) (42.44±21.68 μm).

Random forest feature selection improved the accuracy of the prediction model, confirming the significant impact of preoperative parameter importance ranking on prediction performance. By comparing the three artificial intelligence learning models—Support Vector Machine, Multilayer Perceptron, and Random Forest—this study found that the random forest performed best in predicting postoperative vault after ICL implantation. Its advantages in feature importance analysis and prediction accuracy provide a reliable intelligent analysis tool for personalized ICL surgical planning and dynamic monitoring of postoperative vault, demonstrating potential clinical translation value.