文献速递瓣膜外科

人工智能与机器学习模型用于二尖瓣修复长期耐久性的个体化预测

An artificial intelligence and machine learning model for personalized prediction of long-term mitral valve repair durability.
PubMed 文献 · J Thorac Cardiovasc Surg · 2025-07-18

该研究纳入444例因退行性二尖瓣反流接受初次二尖瓣修复的患者(2008-2024年),比较随机生存森林与Cox比例风险模型预测修复失败(复发反流/狭窄或再干预)的效能。随机生存森林区分度更优(C指数0.874 vs 0.796),并识别对合长度和早期平均二尖瓣跨瓣压差为关键预测因素,可生成个体化长期耐久性估计。

要点

  • 随机生存森林预测二尖瓣修复失败的区分度优于惩罚Cox模型(C指数0.874 vs 0.796)。
  • 对合长度每增加1 mm,Cox模型显示修复失败风险约降低40%。
  • 对合长度6、9、12 mm时,随机生存森林预测15年免于修复失败率分别为51%、91%、96%。

核心结论

  • 机器学习通过建模复杂非线性关联并识别临床可操作预测因素,在预测二尖瓣修复长期耐久性方面优于传统方法。
  • 将机器学习整合进外科实践可能支持更个体化、数据驱动的二尖瓣修复策略并改善长期结局。

研究设计与样本

  • 回顾性分析444例因退行性二尖瓣反流接受初次二尖瓣修复的患者(2008-2024年)。
  • 主要结局为二尖瓣修复失败,定义为复发反流/狭窄或再干预。
  • 比较随机生存森林与惩罚Cox比例风险模型的预测准确性和可解释性,并创建网页应用展示随机生存森林模型。

主要发现

  • 研究期间13例(3%)发生二尖瓣修复失败。
  • 随机生存森林区分度更优(C指数0.874 vs 0.796),并识别对合长度和早期平均二尖瓣跨瓣压差为关键预测因素。
  • Cox模型仅识别对合长度,每增加1 mm修复失败风险约降低40%。
  • 随机生存森林预测对合长度6 mm时5、10、15年免于修复失败率分别为94%、74%、51%;9 mm时为98%、94%、91%;12 mm时为99%、98%、96%。
  • 平均压差2至5 mm Hg与5至10年耐久性90%或更高相关,而8 mm Hg预测较差结局(10年68%,15年64%)。

关键数值与推荐

  • 对合长度6 mm:随机生存森林预测5、10、15年免于修复失败率分别为94%、74%、51%。
  • 对合长度9 mm:预测5、10、15年免于修复失败率分别为98%、94%、91%。
  • 对合长度12 mm:预测5、10、15年免于修复失败率分别为99%、98%、96%。
  • 平均压差2至5 mm Hg与5至10年耐久性90%或更高相关;8 mm Hg预测10年68%、15年64%。
  • Cox模型中对合长度每增加1 mm,修复失败风险约降低40%。

临床怎么用

  • 可利用随机生存森林模型基于对合长度和早期平均跨瓣压差等术前/术后变量个体化预测二尖瓣修复长期耐久性。
  • 对合长度较短(如6 mm)或早期平均压差较高(如8 mm Hg)的患者可识别为修复失败高风险,指导更密切随访或修复策略调整。
  • 网页应用形式可辅助临床医生在决策时直观获取患者特异性生存估计。

局限性

  • 样本量较小(444例),失败事件仅13例,可能限制模型稳定性与泛化能力。
  • 为单中心回顾性研究(2008-2024年),结论外推需外部验证。

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

OBJECTIVE: The study objective was to compare Random Survival Forest, a machine learning method, with Cox proportional hazards models in predicting long-term mitral valve repair durability, focusing on clinical utility and personalized decision-making. METHODS: We analyzed 444 patients undergoing primary mitral valve repair for degenerative mitral regurgitation (2008-2024). The primary outcome was mitral repair failure, defined as recurrent regurgitation/stenosis or reintervention. Random Survival Forest and penalized Cox proportional hazards models were compared for predictive accuracy and interpretability. A web-based application was created to demonstrate the Random Survival Forest model. RESULTS: The failure end point, mitral repair failure, occurred in 13 individuals (3%) during the study period. Random Survival Forest showed superior discrimination (Concordance index: 0.874 vs 0.796) and identified both coaptation length and early mean mitral gradient as key predictors. Cox proportional hazards identified coaptation length alone, with each 1-mm increase reducing failure by approximately 40%. Random Survival Forest-predicted freedom from mitral repair failure at 5, 10, and 15 years was 94%, 74%, and 51% for coaptation length of 6 mm; 98%, 94%, and 91% for 9 mm; and 99%, 98%, and 96% for 12 mm, respectively. Mean gradients of 2 to 5 mm Hg were linked to 90% or greater durability at 5 to 10 years, whereas 8 mm Hg predicted worse outcomes (68% at 10 years, 64% at 15 years). Random Survival Forest further provided nuanced interpretation of temporal risk patterns and generated patient-specific survival estimates to improve repair durability forecasting. CONCLUSIONS: Machine learning outperforms traditional methods by modeling complex, nonlinear associations and identifying clinically actionable predictors. Integrating machine learning into surgical practice may support more personalized, data-driven mitral repair strategies and improve long-term outcomes.

来源:J Thorac Cardiovasc Surg,查看原文。版权归原作者所有。

引用

  1. Malik MI, Nedadur R, Chu MWA, et al. An artificial intelligence and machine learning model for personalized prediction of long-term mitral valve repair durability. J Thorac Cardiovasc Surg. 2025. PMID: 40684853. doi:10.1016/j.jtcvs.2025.07.017

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。