PDF(Pubmed)
Abstract：
BACKGROUND: This study leverages a large retrospective cohort of head and neck cancer patients in order to develop machine learning models to predict radiation induced hyposalivation from dose-volume histograms of the parotid glands.
METHODS: The pre and post-radiotherapy salivary flow rates of 510 head and neck cancer patients were used to fit three predictive models of salivary hypofunction, (1) the Lyman-Kutcher-Burman (LKB) model, (2) a spline-based model, (3) a neural network. A fourth LKB-type model using literature reported parameter values was included for reference. Predictive performance was evaluated using a cut-off dependent AUC analysis.
RESULTS: The neural network model dominated the LKB models demonstrating better predictive performance at every cutoff with AUCs ranging from 0.75 to 0.83 depending on the cutoff selected. The spline-based model nearly dominated the LKB models with the fitted LKB model only performing better at the 0.55 cutoff. The AUCs for the spline model ranged from 0.75 to 0.84 depending on the cutoff chosen. The LKB models had the lowest predictive ability with AUCs ranging from 0.70 to 0.80 (fitted) and 0.67 to 0.77 (literature reported).
CONCLUSIONS: Our neural network model showed improved performance over the LKB and alternative machine learning approaches and provided clinically useful predictions of salivary hypofunction without relying on summary measures.
METHODS: The pre and post-radiotherapy salivary flow rates of 510 head and neck cancer patients were used to fit three predictive models of salivary hypofunction, (1) the Lyman-Kutcher-Burman (LKB) model, (2) a spline-based model, (3) a neural network. A fourth LKB-type model using literature reported parameter values was included for reference. Predictive performance was evaluated using a cut-off dependent AUC analysis.
RESULTS: The neural network model dominated the LKB models demonstrating better predictive performance at every cutoff with AUCs ranging from 0.75 to 0.83 depending on the cutoff selected. The spline-based model nearly dominated the LKB models with the fitted LKB model only performing better at the 0.55 cutoff. The AUCs for the spline model ranged from 0.75 to 0.84 depending on the cutoff chosen. The LKB models had the lowest predictive ability with AUCs ranging from 0.70 to 0.80 (fitted) and 0.67 to 0.77 (literature reported).
CONCLUSIONS: Our neural network model showed improved performance over the LKB and alternative machine learning approaches and provided clinically useful predictions of salivary hypofunction without relying on summary measures.
摘要：
背景：这项研究利用了大量头颈部癌症患者的回顾性队列研究，以开发机器学习模型来从腮腺的剂量-体积直方图中预测辐射引起的唾液分泌不足。
方法：用510例头颈部肿瘤患者放疗前后唾液流速拟合唾液功能减退的三种预测模型，(1)Lyman-Kutcher-Burman(LKB)模型,(2)基于样条模型,(3)神经网络。包括使用文献报道的参数值的第四LKB类型模型作为参考。使用截止依赖性AUC分析评估预测性能。
结果：神经网络模型在LKB模型中占主导地位，在每个截止值都表现出更好的预测性能，AUC范围为0.75至0.83，具体取决于所选择的截止值。基于样条的模型几乎主导了LKB模型，而拟合的LKB模型仅在0.55截止处表现更好。样条模型的AUC范围为0.75至0.84,取决于所选择的截止值。LKB模型的预测能力最低，AUC范围为0.70至0.80（拟合）和0.67至0.77（文献报道）。
结论：我们的神经网络模型显示出优于LKB和替代机器学习方法的性能，并提供了唾液功能减退的临床有用预测，而不依赖于总结措施。
方法：用510例头颈部肿瘤患者放疗前后唾液流速拟合唾液功能减退的三种预测模型，(1)Lyman-Kutcher-Burman(LKB)模型,(2)基于样条模型,(3)神经网络。包括使用文献报道的参数值的第四LKB类型模型作为参考。使用截止依赖性AUC分析评估预测性能。
结果：神经网络模型在LKB模型中占主导地位，在每个截止值都表现出更好的预测性能，AUC范围为0.75至0.83，具体取决于所选择的截止值。基于样条的模型几乎主导了LKB模型，而拟合的LKB模型仅在0.55截止处表现更好。样条模型的AUC范围为0.75至0.84,取决于所选择的截止值。LKB模型的预测能力最低，AUC范围为0.70至0.80（拟合）和0.67至0.77（文献报道）。
结论：我们的神经网络模型显示出优于LKB和替代机器学习方法的性能，并提供了唾液功能减退的临床有用预测，而不依赖于总结措施。
