混合无监督聚类方法预测种植牙失败风险
本研究回顾性分析8513例单颗种植患者临床特征,采用无监督学习与生存分析结合的混合方法聚类,识别出6个特征及预后不同的聚类,其中1个聚类种植存活率显著较低,并含6个高风险特征,据此构建时间分层列线图预测种植失败风险。
要点
- 基于聚类特征可构建时间分层列线图预测种植失败风险。
- 高风险特征包括年龄、吸烟史、种植体直径、长度、位置及手术方式。
- 列线图辅助术前评估与临床决策,模型需多中心前瞻性验证。
核心结论
- 混合无监督聚类方法可识别与种植失败相关的高风险临床特征并构建可视化预测模型。
研究设计与样本
- 回顾性分析8513例接受单颗种植患者的多个临床特征。
- 采用无监督学习算法与生存分析结合的混合方法进行数据挖掘。
主要发现
- 两步聚类与Kaplan-Meier分析识别出6个特征及预后不同的聚类。
- 仅一个聚类种植存活率显著低于其他聚类,该聚类内含6个高风险特征。
- 基于时间分层多变量Cox回归构建列线图,预测种植后0-120天、120-310天及超过310天的种植失败风险。
- 三个时间段模型的一致性指数分别为0.642、0.781和0.715。
关键数值与推荐
- 高风险特征包括年龄、吸烟史、种植体直径、种植体长度、种植体位置和手术方式。
临床怎么用
- 利用基于聚类特征的列线图可视化预测种植失败风险,辅助牙科医生术前评估和临床决策。
局限性
- 需进一步通过多模态、多中心、前瞻性队列进行验证。
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
OBJECTIVES: The aim of this study was to predict the risk of dental implant loss by clustering features associated with implant survival rates. MATERIALS AND METHODS: Multiple clinical features from 8513 patients who underwent single implant placement were retrospectively analysed. A hybrid method integrating unsupervised learning algorithms with survival analysis was employed for data mining. Two-step cluster, univariate Cox regression, and Kaplan‒Meier survival analyses were performed to identify the clustering features associated with implant survival rates. To predict the risk of dental implant loss, nomograms were constructed on the basis of time-stratified multivariate Cox regression. RESULTS: Six clusters with distinct features and prognoses were identified using two-step cluster analysis and Kaplan‒Meier survival analysis. Compared with the other clusters, only one cluster presented significantly lower implant survival rates, and six specific clustering features within this cluster were identified as high-risk factors, including age, smoking history, implant diameter, implant length, implant position, and surgical procedure. Nomograms were created to assess the impact of the six high-risk factors on implant loss for three periods: 1) 0-120 days, 2) 120-310 days, and 3) more than 310 days after implant placement. The concordance indices of the models were 0.642, 0.781, and 0.715, respectively. CONCLUSIONS: The hybrid unsupervised clustering method, which clusters and identifies high-risk clinical features associated with implant loss without relying on predefined labels or target variables, represents an effective approach for developing a visual model for predicting implant prognosis. However, further validation with a multimodal, multicentre, prospective cohort is needed. CLINICAL SIGNIFICANCE: Visual prognosis prediction utilizing this nomogram that predicts the risk of implant loss on the basis of clustering features can assist dentists in preoperative assessments and clinical decision-making, potentially improving dental implant prognosis.
引用
- Xie C, Li Y, Liu K, et al. A hybrid unsupervised clustering method for predicting the risk of dental implant loss. J Dent. 2024. PMID: 39096996. doi:10.1016/j.jdent.2024.105260
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。