文献速递牙体牙髓病口腔修复与种植

评估牙科电子健康记录用于机器学习预测操作结局的准备度:来自BigMouth数据库关于复合树脂与银汞合金修复体存留率的见解

Assessing the readiness of dental electronic health records for machine learning prediction of procedure outcomes: Insights from the bigmouth repository on composite and amalgam restoration survival rates.
PubMed 文献 · J Dent · 2025-06-06

本研究分析BigMouth牙科数据仓库中2011至2020年九所大学21,510例活髓恒磨牙后牙修复体,以五年内根管治疗或拔牙为失败指标,比较银汞合金与复合树脂的存留率。经典生存分析和机器学习模型均显示复合树脂五年存留率更优,银汞合金失败率为8.84%,复合树脂为4.36%。机器学习模型预测性能仅为中等,Cox pH的C-Index最高为0.64,结论认为当前牙科电子健康记录尚不足以支持准确的修复失败预测。

要点

  • 银汞合金修复体失败率为8.84%,高于复合树脂的4.36%。
  • Cox pH的C-Index为0.64,随机生存森林和DeepSurv均为0.63。
  • 年龄、修复类型和充填面数是修复体存留的最显著预测因素。

核心结论

  • 当前牙科电子健康记录尚不具备准确评估和预测修复失败的能力,机器学习模型仅达到中等预测性能,主要因缺乏精确评估所需的关键临床变量。

研究设计与样本

  • 分析BigMouth牙科数据仓库中2011至2020年九所大学21,510例活髓恒磨牙后牙修复体,失败定义为五年内根管治疗或拔牙,采用Cox比例风险分析和机器学习模型比较银汞合金与复合树脂修复体。

主要发现

  • 经典生存分析显示银汞合金失败率为8.84%,复合树脂为4.36%。
  • 机器学习生存模型中Cox pH的C-Index最高为0.64,随机生存森林和DeepSurv均为0.63。
  • 随机生存森林的时间依赖曲线下面积最高为0.620。
  • 年龄、修复类型和充填面数是修复体存留的最显著预测因素。

临床怎么用

  • 作者指出牙科电子健康记录的数据缺口限制了基于机器学习的风险预测准确性。
  • 作者认为加强诊断编码和数据标准化可改善预测建模,支持循证决策和优化治疗计划。
  • 作者认为强化电子健康记录文档将使人机智能驱动工具提升临床结局并推进精准牙科。

局限性

  • 原文未报告局限性

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: Dental electronic health records (EHRs) often lack comprehensive data for evaluating procedure outcomes. Machine learning (ML) enables predictive modeling but its applicability to dental EHR data remains unclear. This study assessed the readiness of dental EHRs for predicting restoration failure using classical and ML models. METHODS: Data from the BigMouth Dental Data Repository was analyzed, focusing on posterior restorations in permanent molars of vital teeth. Failure was defined as root canal treatment or extraction within five years. Predictors included restoration type, caries depth, number of surfaces filled, time to failure, age, and gender. Cox proportional hazards analysis and ML models compared amalgam and composite restorations. RESULTS: A total of 21,510 restorations placed between 2011 and 2020 across nine universities were identified. Both classical and ML analyses indicated superior five-year survival for composite restorations. Classic survival analysis showed a higher failure rate for amalgam (8.84 %) compared to composite (4.36 %). Among ML survival models, Cox Proportional Hazards (Cox pH), Random Survival Forest (RSF), and DeepSurv, Cox pH had the highest Concordance Index (C-Index) at 0.64, while RSF and DeepSurv both achieved 0.63. RSF recorded the highest Time-Dependent Area Under Curve (0.620). Age, restoration type, and number of surfaces filled were the most significant predictors of restoration survival. CONCLUSION: Our findings demonstrate that dental EHRs, in their current state, are not yet equipped to support accurate assessment and prediction of restoration failure. ML models achieved only moderate predictive performance, largely due to the absence of key clinical variables necessary for precise evaluation. CLINICAL SIGNIFICANCE: Gaps in dental EHR data limit the accuracy of ML-based risk prediction. Enhancing diagnostic coding and data standardization can improve predictive modeling, support evidence-based decision-making, and optimize treatment planning. Strengthening EHR documentation will enable Artificial Intelligence-driven tools to enhance clinical outcomes and advance precision dentistry.

来源:J Dent,查看原文。版权归原作者所有。

引用

  1. Alqaderi H, Alhazmi H, Gritzer L, et al. Assessing the readiness of dental electronic health records for machine learning prediction of procedure outcomes: Insights from the bigmouth repository on composite and amalgam restoration survival rates. J Dent. 2025. PMID: 40482587. doi:10.1016/j.jdent.2025.105865

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。