RT2C:利用结构化牙科数据和RNN预测新发龋的时间
本研究基于4个牙科机构466,782名患者和1,062,553次检查的结构化纵向电子牙科记录,开发了基于双向门控循环单元的生存模型RT2C,用于预测患者下次就诊前新发龋的概率和时间。在独立测试集上C-index达0.883±0.08,优于传统生存模型和仅用龋风险评分的基线,且在不同保险、年龄、种族和风险组中表现一致。
要点
- RT2C利用纵向电子牙科记录预测下次就诊前新发龋,C-index达0.883±0.08。
- 模型同时输出新发龋概率和时间,与CAMBRA风险评估框架对接。
- 在不同人口学和临床亚组中表现稳定,具有稳健性和公平性。
核心结论
- RT2C模型能基于结构化纵向牙科数据预测下次就诊前新发龋的概率和时间,测试集C-index为0.883±0.08,优于仅用龋风险评分和传统生存模型基线。
研究设计与样本
- 回顾性队列研究,数据来自4个牙科机构2019至2023年的466,782名患者和1,062,553次检查(D0120、D0150、D0145)。
- 患者分为开发集(326,748人;744,624次就诊)、验证集(46,678人;106,616次)和测试集(93,356人;211,313次),事件率约20%。
- RT2C采用双向门控循环单元网络捕捉就诊间时间依赖,估计个体化生存概率。
- 性能与Cox比例风险模型和随机生存森林比较,指标包括C-index、积分Brier评分和校准分析。
- 亚组分析评估保险类型、年龄组、种族组和基于CRA的风险组的泛化性。
主要发现
- RT2C在独立测试集上C-index最高(0.883±0.08),优于仅用CRA和传统生存基线。
- 性能在人口学和临床亚组中保持一致,显示稳健性和公平性。
临床怎么用
- 模型可整合纵向电子牙科记录,在常规检查中评估患者下次就诊前新发龋的风险和时间。
- 输出可与CAMBRA框架对接,辅助个性化预防性龋病管理。
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
Dental caries is a highly prevalent chronic condition requiring accurate tools to identify at-risk patients and guide preventive care. Most existing caries risk models provide only binary predictions and overlook time-to-event information. Considering that modern caries management relies on risk assessment during routine dental examinations, we developed RT2C, a survival model based on a recurrent neural network (RNN). This model supports caries risk management by assessing the likelihood (Caries Management by Risk Assessment [CAMBRA]) that a patient will be diagnosed with new caries by the next visit. It also estimates the time until this event occurs, using structured longitudinal dental data. Data were extracted from 4 dental organizations, comprising 466,782 patients and 1,062,553 examination visits (D0120, D0150, D0145) between 2019 and 2023. Structured electronic dental record (EDR) data included demographics, dental care history, clinical findings, and the CAMBRA-caries risk assessment (CRA). Patients were split into development (326,748 patients; 744,624 visits), validation (46,678; 106,616), and test (93,356; 211,313) sets, with similar event rates (~20%). The RT2C model used a bidirectional gated recurrent unit network to capture temporal dependencies across visits and estimate individualized survival probabilities. Model performance was compared against a Cox proportional hazards model and a random survival forest using the concordance index (C-index), integrated Brier score, and calibration analysis. Subgroup analyses evaluated generalizability across insurance types, age groups, racial groups, and CRA-based risk groups. The RT2C model achieved the highest concordance index on the held-out test set (0.883 ± 0.08), outperforming the CRA-only and traditional survival baselines. Performance remained consistent across demographic and clinical subgroups, demonstrating robustness and fairness across diverse populations. RT2C models visit-indexed risk trajectories and predicted new caries occurrence by the subsequent visit, aligning with the preventive framework of CAMBRA. By integrating longitudinal EDR data, it adds value to existing CRA tools and supports proactive, personalized caries management.
引用
- Liu X, Kookal KK, Brandon R, et al. RT2C: Predicting Time-to-New Caries with Structured Dental Data Using RNN. J Dent Res. 2026. PMID: 42470279. doi:10.1177/00220345261462463
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。