Modeling of Xerostomia After Radiotherapy for Head and Neck Cancer: A Registry Study

Abstract

Aim: Data from a local quality registry are used to model the risk of late xerostomia after radiotherapy for head and neck cancer (HNC), based on dosimetric- and clinical variables. Strengths and weaknesses of using quality registry data are explored.

Methods: HNC patients treated with radiotherapy at the Karolinska University hospital are entered into a quality registry at routine follow up, recording morbidity according to a modified RTOG/LENT-SOMA scale. Other recorded parameters are performance status, age, gender, tumor location, tumor stage, smoking status, chemotherapy and radiotherapy data, including prescribed dose and organ-at-risk (OAR) dose. Most patients are entered at several time points, but at variable times after treatment. Xerostomia was modeled based on follow-up data from January 2014 to October 2018, resulting in 753 patients. Two endpoints were considered: maximum grade ≥2 (XER_G≥2) or grade ≥3 (XER_G≥3) late xerostomia. Univariate Cox regression was used to select variables for two multivariate models for each endpoint, one based on the mean dose to the total parotid volume (D_tot) and one based on the mean dose to the contralateral parotid (D_contra). Cox regression allows the estimation of the risk of xerostomia at different time points; models were presented visually as nomograms estimating the risk at 9, 12, and 24 months respectively.

Results: The toxicity rates were 366/753 (49%) for XER_G≥2 and 40/753 (5.3%) for XER_G≥3. The multivariate models included several variables for XER_G≥2, and dose, concomitant chemotherapy and age were included for XER_G≥3. Induction chemotherapy and an increased number of fractions per week were associated with a lower risk of XER_G≥2. However, since the causality of these relationships have limited support from previous studies, alternative models without these variables were also presented. The models based on the mean dose to the total parotid volume and the contralateral parotid alone were very similar.

Conclusion: Late xerostomia after radiotherapy can be modeled with reasonable predictive power based on registry data; models are presented for different endpoints highly relevant in clinical practice. However, the risk of modeling indirect relationships, given the unavoidably heterogeneous registry data, needs to be carefully considered in the interpretation of the results.

Keywords: cox regression; head and neck cancer; nomogram; registry analysis; xerostomia.

FIGURE 1

Calibration plots for the models in Table 2 at 1 year: (A) the D_tot model for XER_G≥2, (B) the D_contra model for XER_G≥2, (C) the D_tot model for XER_G≥3, (D) the D_contra model for XER_G≥3. The histogram on the upper × axis represents the frequency distribution of 1 minus the predicted probabilities (c.f. a survival analysis).

FIGURE 2

Calibration plots for the models in Table 3 at 1 year: (A) the D_tot model for XER_G≥2, (B) the D_contra model for XER_G≥2, The histogram on the upper × axis represents the frequency distribution of 1 minus the predicted probabilities (c.f. a survival analysis).

FIGURE 3

The predicted risk of grade ≥2 xerostomia at 9, 12, and 24 months after radiotherapy, not including dose, due to the inverse dose-response relationship. Model specifics are listed in Supplementary Table A2 in the Supplementary Material. *Note that the model does not imply significant differences between all categories with respect to the endpoint – the p-value only refers to the difference from the reference category (see Supplementary Table A2 for p-values). “No” was reference for concomitant chemotherapy and Erbitux was not significantly different from the reference. “N2c” was reference for N stage and N3 was not statistically different. Note that the model does not establish whether there is a difference between N0 and N1. “Oral cavity” was reference for tumor location and oropharynx was not statistically different.

FIGURE 4

The predicted risk of grade ≥3 xerostomia at 9, 12, and 24 months after radiotherapy, using the D_tot dose variable. ^∗Note that not all categories are significantly different from each other with respect to the endpoint – the p-value only refers to the difference from the reference category (see Table 2 for p-values). “No” was reference for concomitant chemotherapy and Erbitux was not significantly different.