A machine learning model can predict the risk of hyperopic shift (HS) after combining phacoemulsification with an intraocular lens (PE+IOL) in patients with primary angle-closure glaucoma (PACG), a paper says. glaucoma journal.
This retrospective cohort study included 423 eyes of PACG patients who underwent PE+IOL surgery between June 2019 and June 2024 based on predefined inclusion and exclusion criteria. Of these, 267 were in the non-HS group and 156 were in the HS group.
The researchers collected patient demographic information, preoperative eye examination findings, and refractive changes 3 to 6 months after surgery.
The Boruta algorithm identified key predictive variables including target refraction, preoperative best-corrected visual acuity (BCVA), axial length (AL), central corneal thickness (CCT), anterior chamber depth (ACD), lens thickness (LT), white-to-white distance (W2W), and pupil diameter (P). Both the SVM and LR models showed moderate classification performance with the highest predictive accuracy with AUC of 0.704 and 0.696, respectively, the report shows.
SVM and LR models demonstrate the highest predictive accuracy and classification performance, providing valuable insights for early identification of HS risk and development of personalized surgical plans.
“Based on retrospective data analysis, this study successfully developed a predictive model for HS risk after PE+IOL surgery in PACG patients,” the researchers said. “SVM and LR models demonstrated the highest predictive accuracy and classification performance, providing valuable insights for early identification of HS risk and development of personalized surgical plans.”
Limitations of the study include its retrospective design, relatively limited sample size, and although machine learning models can improve predictive accuracy, their “black box” nature poses interpretability challenges. Furthermore, although multiple key features were selected using the Boruta algorithm, due to model complexity and processing of high-dimensional data, some potentially relevant features may be overlooked, which may affect the comprehensiveness and accuracy of the predictions.
