文献速递眼底病与葡萄膜炎

机器学习预测原发性孔源性视网膜脱离术后视力预后不良

Identifying patients with poor visual outcomes after primary rhegmatogenous retinal detachment surgery using machine learning.
PubMed 文献 · Br J Ophthalmol · 2026-07-03

本研究基于日本视网膜脱离登记处数据,纳入2658例原发性孔源性视网膜脱离术后6个月有视力记录的患者,从288个特征中筛选47个临床特征,构建LightGBM、XGBoost、随机森林和逻辑回归模型预测术后6个月视力预后不良(logMAR视力≥1.0)。所有机器学习模型曲线下面积(AUROC)为0.876-0.901,其中随机森林模型最高(AUROC 0.901,95%CI 0.881-0.921),优于PRO评分(AUROC 0.794,95%CI 0.744-0.839)。

要点

  • 机器学习模型可高精度预测原发性孔源性视网膜脱离术后6个月视力预后不良。
  • 随机森林模型区分度最高(AUROC 0.901),优于现有PRO评分(AUROC 0.794)。
  • 模型基于47个临床特征构建,数据来源于日本视网膜脱离登记处。

核心结论

  • 机器学习模型对原发性孔源性视网膜脱离术后视力预后不良具有高区分度预测能力。

研究设计与样本

  • 回顾性队列研究,数据来自日本视网膜脱离登记处,纳入2658例术后6个月有视力记录的原发性孔源性视网膜脱离患者。
  • 从288个特征中筛选47个临床特征,使用LightGBM、XGBoost、随机森林和逻辑回归构建预测模型。

主要发现

  • 所有机器学习模型预测术后6个月logMAR视力≥1.0的AUROC为0.876-0.901。
  • 随机森林模型区分度最高,AUROC为0.901(95%CI 0.881-0.921)。
  • 随机森林模型性能优于PRO评分(AUROC 0.794,95%CI 0.744-0.839)。

关键数值与推荐

  • 视力预后不良定义为术后6个月logMAR视力≥1.0。
  • 随机森林模型AUROC 0.901(95%CI 0.881-0.921),PRO评分AUROC 0.794(95%CI 0.744-0.839)。

临床怎么用

  • 可辅助识别原发性孔源性视网膜脱离术后视力预后不良的高风险患者。
  • 随机森林模型区分度优于PRO评分,但需外部验证后用于临床决策。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

原文

AIMS: To develop prediction models for identifying cases with poor visual outcomes after surgery for primary rhegmatogenous retinal detachment (RRD). METHODS: All data were obtained from the Japan-Retinal Detachment Registry. Patients with visual acuity (VA) recorded at 6 months after the surgery for primary RRD were included. For identifying the cases with poor VA defined as the logarithm of the minimum angle of resolution VA≥1.0 at 6 months post-surgery, 47 clinical features were selected out of 288 features, then three different machine learning (ML) algorithms (LightGBM, XGBoost and Random Forest) and logistic regression were used to construct models. Discriminative accuracies of the constructed models were compared with the Primary Retinal detachment Outcomes (PRO) score, a previously proposed scoring system for the same purpose. We also estimated the continuous VA using the same three ML algorithms and linear regression model to detect individuals with poor postoperative VA. RESULTS: A total of 2658 patients were included. We observed that all the constructed ML models showed high discriminative accuracies (area under the receiver operating characteristic curve; AUROC: 0.876-0.901). Of the constructed models, the Random Forest model achieved the highest discriminative accuracy (AUROC: 0.901; 95% CI 0.881 to 0.921). The performance of this model appeared to be superior to the PRO score (AUROC: 0.794, 95% CI 0.744 to 0.839). CONCLUSION: ML models predict poor visual prognosis after primary RRD surgery with high discriminative accuracy.

来源:Br J Ophthalmol,查看原文。版权归原作者所有。

引用

  1. Sajiki AF, Fukutsu K, Matsumoto A, et al. Identifying patients with poor visual outcomes after primary rhegmatogenous retinal detachment surgery using machine learning. Br J Ophthalmol. 2026. PMID: 42399073. doi:10.1136/bjo-2025-328480

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。