Furquim Camila Pinheiro, Caruth Lannawill, Chandrasekaran Ganesh, Cucchiara Andrew, Kallan Michael J, Martin Lynn, Feres Magda, Bittinger Kyle, Divaris Kimon, Glessner Joseph, Kantarci Alpdogan, Giannobile William, Verma Shefali Setia, Teles Flavia
Department of Basic & Translational Sciences, School of Dental Medicine, University of Pennsylvania, Philadelphia, Pennsylvania, USA.
Department of Periodontology and Oral Implantology, Dental Research Division, University of Guarulhos, Guarulhos, São Paulo, Brazil.
J Clin Periodontol. 2025 Oct;52(10):1478-1490. doi: 10.1111/jcpe.14194. Epub 2025 Aug 19.
To construct predictive models of periodontitis progression by applying Machine Learning (ML) to baseline data from a study of periodontitis progression.
Logistic regression (LR), multi-layer perceptron (MLP) and probabilistic graphic model (PGM) were utilised on data from a multi-centre longitudinal study in which periodontally healthy (n = 113) and periodontitis participants (n = 302) were examined bi-monthly for 12 months without treatment. Periodontal examination was performed, and salivary levels of 10 analytes were determined. Clinical and demographic parameters and analytes levels were input into the model. The performance of 14 models was compared using the area under the receiver operating characteristic curve (AUROC), and feature importance was assessed using SHapley Additive exPlanations (SHAP).
The PGM model (Clinical measures, saliva IL-1β, age, sex) demonstrated the best overall performance (AUROC = 0.88), compared to LR (AUROC = 0.72) and MLP (AUROC = 0.58). Although MLP had a lower Brier score (0.12), its sensitivity was 0, limiting its clinical utility. In contrast, PGM achieved a balanced sensitivity (0.55) and specificity (0.81). Feature importance analyses highlighted the number of deep periodontal pockets as a key driver of model predictions in both PGM and MLP.
ML models can predict periodontitis progression, supporting early detection strategies. Our integrative approach, combining clinical data with salivary biomarkers such as IL-1β, improved predictive accuracy.
通过将机器学习（ML）应用于牙周炎进展研究的基线数据，构建牙周炎进展的预测模型。
对一项多中心纵向研究的数据使用逻辑回归（LR）、多层感知器（MLP）和概率图模型（PGM），该研究中对牙周健康者（n = 113）和牙周炎患者（n = 302）每两个月进行一次检查，为期12个月且不进行治疗。进行了牙周检查，并测定了10种分析物的唾液水平。将临床和人口统计学参数以及分析物水平输入模型。使用受试者操作特征曲线下面积（AUROC）比较14种模型的性能，并使用SHapley加性解释（SHAP）评估特征重要性。
与LR（AUROC = 0.72）和MLP（AUROC = 0.58）相比，PGM模型（临床指标、唾液白细胞介素-1β、年龄、性别）表现出最佳的总体性能（AUROC = 0.88）。尽管MLP的布里尔评分较低（0.12），但其灵敏度为0，限制了其临床应用。相比之下，PGM实现了平衡的灵敏度（0.55）和特异性（0.81）。特征重要性分析突出了深牙周袋的数量是PGM和MLP中模型预测的关键驱动因素。
ML模型可以预测牙周炎进展，支持早期检测策略。我们将临床数据与白细胞介素-1β等唾液生物标志物相结合的综合方法提高了预测准确性。
