Abstract：
OBJECTIVE: The aim of this analysis was to compare a clinical periodontal prognostic system and a developed and externally validated artificial intelligence (AI)-based model for the prediction of tooth loss in periodontitis patients under supportive periodontal care (SPC) for 10 years.
METHODS: Clinical and radiographic parameters were analysed to assign tooth prognosis with a tooth prognostic system (TPS) by two calibrated examiners from different clinical centres (London and Pittsburgh). The prediction model was developed on the London dataset. A logistic regression model (LR) and a neural network model (NN) were developed to analyse the data. These models were externally validated on the Pittsburgh dataset. The primary outcome was 10-year tooth loss in teeth assigned with \'unfavourable\' prognosis.
RESULTS: A total of 1626 teeth in 69 patients were included in the London cohort (development cohort), while 2792 teeth in 116 patients were included in the Pittsburgh cohort (external validated dataset). While the TPS in the validation cohort exhibited high specificity (99.96%), moderate positive predictive value (PPV = 50.0%) and very low sensitivity (0.85%), the AI-based model showed moderate specificity (NN = 52.26%, LR = 67.59%), high sensitivity (NN = 98.29%, LR = 91.45%), and high PPV (NN = 89.1%, LR = 88.6%).
CONCLUSIONS: AI-based models showed comparable results with the clinical prediction model, with a better performance in specific prognostic risk categories, confirming AI prediction model as a promising tool for the prediction of tooth loss.
METHODS: Clinical and radiographic parameters were analysed to assign tooth prognosis with a tooth prognostic system (TPS) by two calibrated examiners from different clinical centres (London and Pittsburgh). The prediction model was developed on the London dataset. A logistic regression model (LR) and a neural network model (NN) were developed to analyse the data. These models were externally validated on the Pittsburgh dataset. The primary outcome was 10-year tooth loss in teeth assigned with \'unfavourable\' prognosis.
RESULTS: A total of 1626 teeth in 69 patients were included in the London cohort (development cohort), while 2792 teeth in 116 patients were included in the Pittsburgh cohort (external validated dataset). While the TPS in the validation cohort exhibited high specificity (99.96%), moderate positive predictive value (PPV = 50.0%) and very low sensitivity (0.85%), the AI-based model showed moderate specificity (NN = 52.26%, LR = 67.59%), high sensitivity (NN = 98.29%, LR = 91.45%), and high PPV (NN = 89.1%, LR = 88.6%).
CONCLUSIONS: AI-based models showed comparable results with the clinical prediction model, with a better performance in specific prognostic risk categories, confirming AI prediction model as a promising tool for the prediction of tooth loss.
摘要：
目的：本分析的目的是比较临床牙周预后系统和已开发和外部验证的基于人工智能（AI）的模型，以预测在支持牙周护理（SPC）下牙周炎患者的牙齿脱落10年。
方法：由来自不同临床中心（伦敦和匹兹堡）的两名校准检查者分析了临床和影像学参数，以通过牙齿预后系统（TPS）分配牙齿预后。预测模型是在伦敦数据集上开发的。建立了逻辑回归模型（LR）和神经网络模型（NN）来分析数据。这些模型在匹兹堡数据集上进行了外部验证。主要结果是预后不良的牙齿中10年牙齿脱落。
结果：共有69例患者的1626颗牙齿被纳入伦敦队列（发展队列），而116例患者的2792颗牙齿被纳入匹兹堡队列（外部验证数据集）。虽然验证队列中的TPS表现出高特异性(99.96%)，中等阳性预测值(PPV=50.0%)和极低灵敏度(0.85%)，基于AI的模型显示出中等特异性(NN=52.26%，LR=67.59%)，高灵敏度(NN=98.29%，LR=91.45%)，和高PPV(NN=89.1%，LR=88.6%）。
结论：基于AI的模型显示出与临床预测模型具有可比性的结果，在特定的预后风险类别中表现更好，确认AI预测模型是预测牙齿脱落的有前途的工具。
方法：由来自不同临床中心（伦敦和匹兹堡）的两名校准检查者分析了临床和影像学参数，以通过牙齿预后系统（TPS）分配牙齿预后。预测模型是在伦敦数据集上开发的。建立了逻辑回归模型（LR）和神经网络模型（NN）来分析数据。这些模型在匹兹堡数据集上进行了外部验证。主要结果是预后不良的牙齿中10年牙齿脱落。
结果：共有69例患者的1626颗牙齿被纳入伦敦队列（发展队列），而116例患者的2792颗牙齿被纳入匹兹堡队列（外部验证数据集）。虽然验证队列中的TPS表现出高特异性(99.96%)，中等阳性预测值(PPV=50.0%)和极低灵敏度(0.85%)，基于AI的模型显示出中等特异性(NN=52.26%，LR=67.59%)，高灵敏度(NN=98.29%，LR=91.45%)，和高PPV(NN=89.1%，LR=88.6%）。
结论：基于AI的模型显示出与临床预测模型具有可比性的结果，在特定的预后风险类别中表现更好，确认AI预测模型是预测牙齿脱落的有前途的工具。
