电化学阻抗谱与人工智能非破坏性预测镍钛旋转锉剩余使用寿命
该研究对94根镍钛旋转锉从初始状态至疲劳断裂进行电化学阻抗谱监测,构建864条有效测量数据集,采用频率点与滑动窗口联合优化策略筛选特征,评估18种模型并建立三级临床风险预警系统。CatBoost取得最低交叉验证平均绝对误差(0.133 ± 0.015)和平均三级风险分类准确率(71.1% ± 2.9%),联合优化策略降低特征维度83.0%且性能未显著下降。研究表明基于EIS与AI的非破坏性评估方法可在断裂前定量评价镍钛锉疲劳状态,为个性化器械更换决策提供临床依据。
要点
- CatBoost模型交叉验证MAE为0.133 ± 0.015,三级风险分类准确率71.1% ± 2.9%。
- 联合优化策略降低特征维度83.0%且未显著影响模型性能。
- 传统机器学习模型在小表格数据集上优于常规训练深度学习架构和TabPFN。
核心结论
- 基于电化学阻抗谱与人工智能的非破坏性评估方法可在镍钛旋转锉断裂前定量评价其疲劳状态。
- 该框架为个性化器械更换决策提供具有临床意义的依据。
研究设计与样本
- 对94根镍钛旋转锉从初始状态至疲劳断裂进行电化学阻抗谱监测,构建含864条有效测量的数据集。
- 设计频率点与滑动窗口联合优化策略筛选关键特征并消除特征维度冗余。
- 采用五折交叉验证评估18种模型:9种传统机器学习模型、8种常规训练深度学习架构和预训练表格基础模型TabPFN。
- 建立三级临床风险预警系统。
主要发现
- CatBoost取得最低交叉验证平均绝对误差(0.133 ± 0.015)。
- CatBoost平均三级风险分类准确率为71.1% ± 2.9%。
- 联合优化策略降低特征维度83.0%且未显著损害性能。
- 最佳传统机器学习模型在小表格数据集上优于常规训练深度学习架构和TabPFN。
关键数值与推荐
- CatBoost交叉验证平均绝对误差为0.133 ± 0.015。
- CatBoost平均三级风险分类准确率为71.1% ± 2.9%。
- 联合优化策略降低特征维度83.0%。
临床怎么用
- 可利用EIS与AI框架在镍钛旋转锉断裂前定量评估疲劳状态,辅助个性化器械更换决策。
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
INTRODUCTION: Nickel-titanium (NiTi) rotary files are widely used in modern root canal therapy. Although instrument separation is relatively uncommon, sudden fatigue fracture may lead to serious clinical consequences. Current methods are limited to detecting macroscopic defects and cannot quantitatively assess early-stage microscopic fatigue damage. To address this issue, we combined electrochemical impedance spectroscopy (EIS) with artificial intelligence to develop a framework for the non-destructive prediction of the remaining useful life of NiTi rotary files. METHODS: We conducted EIS monitoring on 94 NiTi rotary files from their pristine state until fatigue fracture, and built a dataset containing 864 valid measurements. We designed a frequency point and sliding window joint optimization strategy to select key features and eliminate feature dimensionality redundancy. Using five-fold cross-validation, we evaluated 18 models: nine traditional machine-learning models, eight conventionally trained deep-learning architectures, and the pretrained tabular foundation model TabPFN; we established a three-level clinical risk-warning system. RESULTS: Among the evaluated models, CatBoost, a gradient-boosting decision-tree algorithm, achieved the lowest cross-validated mean absolute error (0.133 ± 0.015) and a mean three-level risk-classification accuracy of 71.1% ± 2.9%. Furthermore, the joint optimization strategy reduced feature dimensionality by 83.0% without significantly compromising the performance. The best-performing traditional machine-learning models outperformed conventionally trained deep-learning architectures and TabPFN on this small tabular dataset. CONCLUSIONS: This study confirms that an EIS- and AI-based non-destructive assessment method enables the quantitative evaluation of the fatigue state of NiTi rotary files prior to fracture, providing a clinically meaningful basis for personalized instrument replacement decisions.
引用
- Ma Y, Chen K, Xu Y, et al. Non-destructive Prediction of Remaining Useful Life for Nickel-Titanium Rotary Files Using Electrochemical Impedance Spectroscopy and Artificial Intelligence. J Endod. 2026. PMID: 42764067. doi:10.1016/j.joen.2026.09.004
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。