Design and Selection of Machine Learning Methods Using Radiomics and Dosiomics for Normal Tissue Complication Probability Modeling of Xerostomia

Abstract

Purpose: The purpose of this study is to investigate whether machine learning with dosiomic, radiomic, and demographic features allows for xerostomia risk assessment more precise than normal tissue complication probability (NTCP) models based on the mean radiation dose to parotid glands.

Material and methods: A cohort of 153 head-and-neck cancer patients was used to model xerostomia at 0-6 months (early), 6-15 months (late), 15-24 months (long-term), and at any time (a longitudinal model) after radiotherapy. Predictive power of the features was evaluated by the area under the receiver operating characteristic curve (AUC) of univariate logistic regression models. The multivariate NTCP models were tuned and tested with single and nested cross-validation, respectively. We compared predictive performance of seven classification algorithms, six feature selection methods, and ten data cleaning/class balancing techniques using the Friedman test and the Nemenyi post hoc analysis.

Results: NTCP models based on the parotid mean dose failed to predict xerostomia (AUCs < 0.60). The most informative predictors were found for late and long-term xerostomia. Late xerostomia correlated with the contralateral dose gradient in the anterior-posterior (AUC = 0.72) and the right-left (AUC = 0.68) direction, whereas long-term xerostomia was associated with parotid volumes (AUCs > 0.85), dose gradients in the right-left (AUCs > 0.78), and the anterior-posterior (AUCs > 0.72) direction. Multivariate models of long-term xerostomia were typically based on the parotid volume, the parotid eccentricity, and the dose-volume histogram (DVH) spread with the generalization AUCs ranging from 0.74 to 0.88. On average, support vector machines and extra-trees were the top performing classifiers, whereas the algorithms based on logistic regression were the best choice for feature selection. We found no advantage in using data cleaning or class balancing methods.

Conclusion: We demonstrated that incorporation of organ- and dose-shape descriptors is beneficial for xerostomia prediction in highly conformal radiotherapy treatments. Due to strong reliance on patient-specific, dose-independent factors, our results underscore the need for development of personalized data-driven risk profiles for NTCP models of xerostomia. The facilitated machine learning pipeline is described in detail and can serve as a valuable reference for future work in radiomic and dosiomic NTCP modeling.

Keywords: IMRT; NTCP; dosiomics; head and neck; machine learning; radiomics; radiotherapy; xerostomia.

Figures

Figure 1

Frequency of the follow-up reports collection.

Figure 2

The workflow of a multivariate five-step model building comprising, in this order, feature-group selection, feature scaling, sampling, feature selection, and classification.

Figure 3

Predictive power of individual features in the time-specific models measured with the area under the receiver operating characteristic curve (AUC). The left-hand side vertical axis lists the features, the right-hand side vertical axis lists the feature groups. The AUCs were calculated from the corresponding Mann–Whitney U statistic. Bars marked with * are significant at the false discovery rate (FDR) ≤ 0.05.

Figure 4

The mean dose and the absolute right–left dose gradient distribution in our patient cohort.

Figure 5

A comparison of classification, feature selection, and sampling algorithms in terms of their predictive performance in model tuning. All heat maps in a given column belong to a single end point, whereas all heat maps in a given row correspond to a single classifier. In each heat map, rows represent feature selection algorithms and columns correspond to sampling methods. The color maps are normalized per end point. The color bar ticks correspond to the worst, average, and the best model performance.

Figure 6

Heat maps showing a proportion of times a given algorithm on the vertical axis outperformed another algorithm on the horizontal axis in terms of the best AUC in model tuning. For example, support vector machines (SVM) performed better than extra-trees (ET) in 73% of the time-specific models.

Figure 7

A comparison of classification, feature selection, and sampling methods against one another with the Nemenyi test. Lower ranks correspond to better performance of the algorithm, that is rank 1 is the best. Algorithms which ranks differ by less than the critical difference (CD) are not significantly different at 0.05 significance level and are connected by the black bars.

Figure 8

Features underlying the multivariate models of long-term xerostomia. i, ipsilateral gland; c, contralateral gland.