Abstract: Objective    To evaluate the raionality of the machine learning-based model in the selection of abutment teeth in the design of removable partial dentures. Methods    The clinical and imaging data of 50 patients with dentition defect treated in department of VIP Dental Service of Peking University School of Stomatology from October 1，2019 to May 1，2020 were retrospectively analyzed，with a total of 74 models of dentition defect in one jaw. The Kennedy classification of each case was recorded. The panoramic images and oral condition information of all patients were entered into the machine learning-based model. An expert group evaluated the rationality of the results according to the uniform evaluation criteria in terms of the periodontal condition，pulp condition，number，position，and distribution of the abutments. Results    Kennedy class Ⅰ，Ⅱ，Ⅲ and Ⅳ accounted for 37.8%（28/74），44.6%（33/74），12.2%（9/74）and 5.4%（4/74），respectively. In terms of the periodontal condition，pulp condition，number，position，and distribution of the abutments，the rationality was evaluated as reasonable in 87.8%（65/74），90.5%（67/74），91.9%（68/74），78.3%（58/74），and 82.4%（61/74） of the cases，basically reasonable in 10.8%（8/74），6.8%（5/74），5.4%（4/74），14.9%（11/74），and 8.1%（6/74）of the cases，and unreasonable in < 10% of the cases. Only in terms of the position and distribution of the abutments，the evaluation results of the rationality of abutment selection for dentition defect models of different Kennedy classifications were statistically different （H values were 10.191 and 11.554，respectively，both P < 0.05）. The abutment selection results showed more unreasonable cases for Kennedy class Ⅳ dentition defect models. Conclusion    The machine learning-based removable partial denture abutment selection model performs well in dentition defect models of different Kennedy classifications in clinical practice，which can be used as a reference for clinical decision.

Key words: artificial intelligence, machine learning, removable partial denture, abutment teeth, prothodontics

摘要： 目的    评价基于机器学习的可摘局部义齿基牙选择模型的合理性。方法    选取2019-10-01至2020-05-01于北京大学口腔医院特诊科就诊的50例牙列缺损患者的临床和影像学资料进行回顾性分析，共计单颌牙列缺损模型74个，记录其Kennedy分类情况。将所有患者的曲面体层片和口腔状况信息录入至基于机器学习的可摘局部义齿基牙选择模型中，由专家小组按照统一评价标准对其结果进行基牙牙周状况、牙体牙髓状况、数目、位置、分布5个项目的合理性评价。结果    Kennedy第一类、第二类、第三类、第四类牙列缺损模型分别占37.8%（28/74）、44.6%（33/74）、12.2%（9/74）、5.4%（4/74）。在基牙牙周状况、牙体牙髓状况、数目、位置、分布5个项目中，评价为合理的依次占87.8%（65/74）、90.5%（67/74）、91.9%（68/74）、78.3%（58/74）、82.4%（61/74），评价为基本合理的依次占10.8%（8/74）、6.8%（5/74）、5.4%（4/74）、14.9%（11/74）、8.1%（6/74），评价为不合理的占比均< 10%。在基牙位置和基牙分布2个项目中，不同Kennedy分类的牙列缺损模型基牙选择的合理性评价的结果比较，差异有统计学意义（H值分别为10.191、11.554，均P < 0.05），其中Kennedy第四类牙列缺损模型基牙选择结果为不合理的占比较高。结论    基于机器学习的可摘局部义齿基牙选择模型在临床应用于不同Kennedy分类的牙列缺损模型中具有较好的合理性，可作为临床决策的参考。

关键词: 人工智能, 机器学习, 可摘局部义齿, 基牙, 修复