Abstract The measurement of teeth in vivo (i.e., in the mouth, without extraction) with EPR spectroscopy in the L-band would allow to screen large groups of population in an event of an acute radiation exposure and in routine epidemiological studies. The radiation dose is proportional to intensity of the radiation-induced signal amplitude determined after subtraction of both native and solar light induced signal amplitudes from the total signal amplitude measured in L-band. Therefore, to improve the dose assessments of in vivo tooth dosimetry a better accuracy of native background signal is necessary. In this work, we present a search for the optimal machine learning approach for predicting of intensity of the native signal amplitude. The study used the dataset from two institutes composed of 1800 EPR spectra which were recorded in the X-band at a large-scale examination of the population of the Central Russia and North Kazakhstan. To determine the relevance of 12 various features a preliminary statistical significance analysis was used. Predictive models for native signal amplitude determination were developed and trained using standard Python frameworks for machine learning and data processing. The employed algorithms included 8 most popular machine learning regressors. To tune the performance of each algorithm a common evaluation technique, namely cross-validation, was used. Finally, mean squared error and coefficient of determination were calculated for performance analysis of the employed models. Comparison among the performance of all established prediction models revealed that Random Forest and Gradient Boosting had most superior performance. Overall, the application of machine learning methods was shown to provide a minor (5–11% in terms of R2) but non negligible improvement to the accuracy of native signal amplitude prediction. Using the technique of adding synthetic noise variables, the most significant factor regarding the prediction was position of tooth in the quadrant.

中文翻译：

---

用于确定牙釉质中原生背景 EPR 信号幅度的机器学习

摘要 使用 L 波段 EPR 光谱在体内（即在口腔中，无需拔牙）测量牙齿将允许在急性辐射暴露事件和常规流行病学研究中筛查大量人群。辐射剂量与从 L 波段测量的总信号幅度中减去自然光和太阳光诱导信号幅度后确定的辐射诱导信号幅度的强度成正比。因此，为了改进体内牙齿剂量测定的剂量评估，需要更好的原生背景信号准确度。在这项工作中，我们提出了一种用于预测本机信号幅度强度的最佳机器学习方法的搜索。该研究使用了来自两个研究所的数据集，该数据集由 1800 个 EPR 光谱组成，这些光谱是在对俄罗斯中部和北哈萨克斯坦人口进行大规模检查时记录在 X 波段的。为了确定 12 个不同特征的相关性，使用了初步的统计显着性分析。使用用于机器学习和数据处理的标准 Python 框架开发和训练用于确定本地信号幅度的预测模型。所采用的算法包括 8 个最流行的机器学习回归器。为了调整每个算法的性能，使用了一种通用的评估技术，即交叉验证。最后，计算均方误差和决定系数，用于对所采用模型的性能分析。比较所有已建立的预测模型的性能，发现随机森林和梯度提升的性能最为优越。总体而言，机器学习方法的应用被证明对原始信号幅度预测的准确性有轻微的提高（就 R2 而言为 5-11%）但不可忽略不计。使用添加合成噪声变量的技术，预测的最重要因素是象限中牙齿的位置。