基于机器学习构建并验证种植体周围黏膜炎1年预测模型
本研究回顾性纳入2011年10月至2023年6月接受种植牙的患者构建可解释机器学习模型,外部测试使用另外两家医院数据,采用LASSO筛选特征并比较六种算法。支持向量机(SVM)模型在验证集AUC为0.902、Brier分数0.129,外部测试AUC为0.797、Brier分数0.256,区分度与校准度优于其他算法,SHAP解释显示角化黏膜宽度不足、口腔卫生差、联冠/桥修复、游离端缺牙、牙龈炎、吸烟、薄龈生物型及粘接固位修复会升高种植体周围黏膜炎(PIM)发生风险,而使用牙间刷/牙线及冲牙器为保护因素。
要点
- SVM模型基于10个特征预测种植术后1年PIM,验证集AUC 0.902,外部测试AUC 0.797。
- 角化黏膜宽度不足、口腔卫生差、联冠/桥修复、游离端缺牙、牙龈炎、吸烟、薄龈生物型及粘接固位修复升高PIM风险。
- 使用牙间刷/牙线及冲牙器为PIM保护因素,SHAP支持个体化风险评估。
核心结论
- 研究开发了首个可解释的机器学习工具,用于预测种植牙患者1年后PIM,SVM模型区分度与临床效用较高。
研究设计与样本
- 回顾性研究纳入2011年10月至2023年6月接受种植牙的患者构建模型,并用另外两家医院数据做外部测试。
- 开发队列共341例患者、605枚种植体;外部测试队列共101例患者、170枚种植体。
- 采用LASSO进行特征筛选,比较六种机器学习算法以确定最优模型,并用SHAP解释模型与个体化风险评估。
主要发现
- 含10个特征的SVM模型在验证集AUC为0.902、Brier分数0.129,外部测试队列AUC为0.797、Brier分数0.256,区分度与校准度优于其他算法。
- SHAP解释显示角化黏膜宽度不足、口腔卫生差、修复类型为联冠/桥、游离端缺牙、牙龈炎、吸烟习惯、薄龈生物型及粘接固位修复会升高PIM发生。
- 使用牙间刷/牙线及使用冲牙器为保护因素。
关键数值与推荐
- SVM模型验证集AUC 0.902、Brier分数0.129;外部测试队列AUC 0.797、Brier分数0.256。
- 模型纳入10个特征,采用LASSO筛选,SHAP用于解释和个体化风险评估。
临床怎么用
- 可用于种植术后1年PIM风险分层,识别高风险患者并加强口腔卫生宣教与随访。
- 对存在角化黏膜宽度不足、薄龈生物型、吸烟、联冠/桥或粘接固位修复等风险因素者,可提前采取针对性干预。
- 鼓励使用牙间刷/牙线及冲牙器作为保护性口腔卫生措施。
局限性
- 原文未报告局限性相关内容。
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: To develop and validate an interpretable machine learning (ML) model for early prediction of peri-implant mucositis (PIM). MATERIAL AND METHODS: This retrospective study enrolled patients receiving dental implants between October 2011 to June 2023 in our cohort to develop a ML-based model. External testing was subsequently performed utilizing datasets from two other hospitals. The least absolute shrinkage and selection operator (LASSO) was utilized for feature selection. Six ML algorithms were applied to analyze and identify the optimal model. Shapley Additive exPlanations (SHAP) interpretation was developed for model interpretation and personalized risk assessment. RESULTS: A total of 341 patients with 605 implants were included in our cohort and 101 patients with 170 implants were included in the external testing cohort. The Area Under Curves (AUCs) and Brier scores of the Support Vector Machine (SVM) model with 10 features were 0.902 and 0.129 in the validation set, while 0.797 and 0.256 in the external testing cohort, demonstrating superior discrimination and calibration compared to other algorithms. The SHAP algorithm interpreted that inadequate keratinized mucosa width, poor oral hygiene, restoration type (splinted crown/bridge), free-end edentulism, gingivitis, smoking habit, thin gingival biotype, and cement-retained restoration would elevate the occurrence of PIM. On the contrary, use of interproximal brushing/flossing and use of oral irrigator served as protective factors. CONCLUSION: This study developed the first interpretable ML-based tool for predicting PIM in patients with dental implant placement after 1-year, with the SVM model showing high discriminative ability and clinical utility for risk stratification.
引用
- Liu L, Li B, Qian Y, et al. Development and Validation of a Machine Learning-Based Model to Predict Peri-Implant Mucositis in Patients With Dental Implant Placement After 1-Year. Clin Oral Implants Res. 2026. PMID: 41937253. doi:10.1111/clr.70125
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。