Scholarlink

Purpose <p>To develop and validate machine learning (ML) models for predicting early postoperative corneal edema (CE) after phacoemulsification in patients with normal preoperative corneal endothelium.</p> Methods <p>A retrospective cohort study analyzed data from 1128 eyes undergoing uncomplicated phacoemulsification at Tianjin Medical University Eye Hospital (May 2024-May 2025). CE was diagnosed on postoperative day 1 based on central corneal thickness increase &gt; 50&#xa0;μm, central stromal opacification, and blurred iris texture. Lasso regression preliminarily screened 26 clinical features to 9 key variables. Eleven ML models and a logistic regression model were developed. A hybrid Recursive Feature Elimination and Exhaustive Search strategy identified optimal feature subsets. Model performance is evaluated using metrics such as Area Under the Curve (AUC), accuracy, sensitivity, specificity, and Brier Score.</p> Results <p>Key predictors consistently identified included cumulative dissipated energy (CDE), age, and lens nucleus hardness. The optimized Support Vector Machine (SVM) model demonstrated superior performance during internal validation (AUC = 0.92, accuracy = 91.56%) and maintained strong generalizability on the independent test set (AUC = 0.89, accuracy = 84.50%). Other models like GradientBoost and PLS-DA also showed good performance post-optimization. The traditional LR model underperformed (optimized AUC = 0.81).</p> Conclusions <p>ML models, particularly SVM, effectively predict postoperative CE risk in cataract patients with normal corneal endothelium. CDE, age, and nucleus hardness are critical predictors. The SVM model provides a robust, preoperative tool for predicting individual risk of postoperative CE. By identifying high-risk patients, it has the potential to guide surgical technique selection, optimize patient counseling, and streamline postoperative care pathways, thereby enhancing overall surgical quality and patient satisfaction.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Machine learning for risk stratification of postoperative corneal edema in East Asian cataract patients: a model for precision ophthalmology

  • 期刊论文

  • 发表日期：

    2026年5月13日

查看全文
  • Shuaixin Lu,
  • Min Zhou,
  • Xiaohua Zhang,
  • Peiyao Huang,
  • Simo Pan,
  • Lingling Ba,
  • Jie Wang,
  • Xinmiao Zhou,
  • Kai Wen,
  • Jing Sun

摘要

Purpose

To develop and validate machine learning (ML) models for predicting early postoperative corneal edema (CE) after phacoemulsification in patients with normal preoperative corneal endothelium.

Methods

A retrospective cohort study analyzed data from 1128 eyes undergoing uncomplicated phacoemulsification at Tianjin Medical University Eye Hospital (May 2024-May 2025). CE was diagnosed on postoperative day 1 based on central corneal thickness increase > 50 μm, central stromal opacification, and blurred iris texture. Lasso regression preliminarily screened 26 clinical features to 9 key variables. Eleven ML models and a logistic regression model were developed. A hybrid Recursive Feature Elimination and Exhaustive Search strategy identified optimal feature subsets. Model performance is evaluated using metrics such as Area Under the Curve (AUC), accuracy, sensitivity, specificity, and Brier Score.

Results

Key predictors consistently identified included cumulative dissipated energy (CDE), age, and lens nucleus hardness. The optimized Support Vector Machine (SVM) model demonstrated superior performance during internal validation (AUC = 0.92, accuracy = 91.56%) and maintained strong generalizability on the independent test set (AUC = 0.89, accuracy = 84.50%). Other models like GradientBoost and PLS-DA also showed good performance post-optimization. The traditional LR model underperformed (optimized AUC = 0.81).

Conclusions

ML models, particularly SVM, effectively predict postoperative CE risk in cataract patients with normal corneal endothelium. CDE, age, and nucleus hardness are critical predictors. The SVM model provides a robust, preoperative tool for predicting individual risk of postoperative CE. By identifying high-risk patients, it has the potential to guide surgical technique selection, optimize patient counseling, and streamline postoperative care pathways, thereby enhancing overall surgical quality and patient satisfaction.