Development of multivariate NTCP models for radiationinduced hypothyroidism: a comparative analysis
 Laura Cella†^{1, 2}Email author,
 Raffaele Liuzzi†^{1, 2},
 Manuel Conson^{2},
 Vittoria D’Avino^{1},
 Marco Salvatore^{2} and
 Roberto Pacelli^{1, 2}
DOI: 10.1186/1748717X7224
© Cella et al.; licensee BioMed Central Ltd. 2012
Received: 25 October 2012
Accepted: 20 December 2012
Published: 27 December 2012
Abstract
Background
Hypothyroidism is a frequent late side effect of radiation therapy of the cervical region. Purpose of this work is to develop multivariate normal tissue complication probability (NTCP) models for radiationinduced hypothyroidism (RHT) and to compare them with already existing NTCP models for RHT.
Methods
Fiftythree patients treated with sequential chemoradiotherapy for Hodgkin’s lymphoma (HL) were retrospectively reviewed for RHT events. Clinical information along with thyroid gland dose distribution parameters were collected and their correlation to RHT was analyzed by Spearman’s rank correlation coefficient (Rs). Multivariate logistic regression method using resampling methods (bootstrapping) was applied to select model order and parameters for NTCP modeling. Model performance was evaluated through the area under the receiver operating characteristic curve (AUC). Models were tested against external published data on RHT and compared with other published NTCP models.
Results
If we express the thyroid volume exceeding X Gy as a percentage (V_{x}(%)), a twovariable NTCP model including V_{30}(%) and gender resulted to be the optimal predictive model for RHT (Rs = 0.615, p < 0.001. AUC = 0.87). Conversely, if absolute thyroid volume exceeding X Gy (V_{x}(cc)) was analyzed, an NTCP model based on 3 variables including V_{30}(cc), thyroid gland volume and gender was selected as the most predictive model (Rs = 0.630, p < 0.001. AUC = 0.85). The threevariable model performs better when tested on an external cohort characterized by large interindividuals variation in thyroid volumes (AUC = 0.914, 95% CI 0.760–0.984). A comparable performance was found between our model and that proposed in the literature based on thyroid gland mean dose and volume (p = 0.264).
Conclusions
The absolute volume of thyroid gland exceeding 30 Gy in combination with thyroid gland volume and gender provide an NTCP model for RHT with improved prediction capability not only within our patient population but also in an external cohort.
Keywords
NTCP modeling Radiotherapy Hypothyroidism BootstrappingBackground
Radiationinduced hypothyroidism (RHT) is a frequent side effect after therapeutic irradiation of the cervical region and it has been described in patients undergoing radiation therapy (RT) for different neoplasms such as lymphoma, headandneck cancer and breast cancer [1–3].
The amelioration of life span expectations of cancer patients requires the maximum possible effort to reduce iatrogenic diseases like RHT. The evolution of radiation therapy technology has enhanced the ability to adapt RT techniques to the individual patient. However, in order to establish tailored strategies for a riskadapted RT, it is essential to identify specific clinical and dosimetric parameters that are involved in the process of modeling normal tissue complication probability (NTCP). Input parameters have been recognized to be among the most critical features of an effective NTCP model [4]. Models that take into account relationships among different patientrelated and dosimetric factors may offer a powerful approach to the optimization of risk ascertainment for many different endpoints [5]. As a consequence, datadriven multivariate modeling of NTCP [6] is increasingly being used unlike traditional NTCP models that only involve dose distribution parameters of a specific organ at risk like the LymanKutcherBurman model.
Recently, a multivariate NTCP model for RHT based on mean thyroid dose and thyroid volume was developed by Boomsma et al [7] in patients treated for headandneck cancer. A thyroid volume effect in RHT development, following RT of breast cancer, was also emphasized in a casecontrol study where the absolute volume receiving more than 30 Gy was recognized as a critical factor for hypothyroidism development [8]. In a previous work [9] on RHT in Hodgkin’s lymphoma (HL) patients, after conventional multivariate analysis method, the percentage of thyroid volume exceeding 30 Gy (V_{30}(%)) was found to be the only predictor of RHT. All the above mentioned results, although similar, are not coincident and seem to suggest different prognostic variables for RHT among patients from different populations.
In this framework, the present report expands on the potential of building an effective multivariate NTCP model for RHT and extends the complexity of the analysis in order to evaluate if general information on RHT risk assessment may be extrapolated regardless of the cohort of patients on which the model is built on. To this end, NTCP modelling exercises were performed using bootstrapping together with validation and performance comparisons on different patients cohorts evaluated for RHT using data from the literature [7–9].
Methods
Patient dataset
Patient, disease and treatment characteristics
Median age (years)  27.5 (14–70)  
Median thyroid volume (cc)  13.7 (6.7–44.0)  
Gender  N  % 
Male  25  47.2 
Female  28  52.8 
Histology  
Nodular sclerosis  38  71.7 
Mixed cellularity  10  18.9 
Lymphocyterichclassical  5  9.4 
Stage  
III  42  79.2 
IIIIV  11  20.8 
Radiotherapy dose delivered  
30 Gy  23  43.4 
32 Gy  25  47.2 
36 Gy  5  9.4 
Chemotherapy regimen  
ABVD  15  28.3 
VEBEP  38  71.7 
All patients were treated with full 3D CT based radiation treatment planning as described in detail in a previous publication [10]. In short, threedimensional conformal plans were generated using a commercial treatment planning system (XiO, Elekta CMS. St Louis. MO) and the convolution dose calculation algorithm, appropriate in the presence of heterogeneous tissues, was applied. RT was administered using 6 20 MV photon beams from a linear accelerator with anteroposteriorposteroanterior fields. A total median dose of 32 Gy (range 30–36) in 20 daily fractions of 1.5–1.8 Gy was planned. For all patients, the thyroid gland was retrospectively delineated on purpose on the planning CTimages by the same radiation oncologist (M.C.). The thyroid gland volume, the minimum (D_{min}), maximum (D_{max}) and mean doses (D_{mean}), the absolute volume of thyroid and the percentage of thyroid volume exceeding 10, 20 and 30 Gy (V_{x}(cc) and V_{x}(%), respectively) were calculated from the dose volume histograms. In addition, the “residual X Gy thyroid volume”, defined as the difference between the thyroid gland volume and Vx (cc), was calculated.
Statistical modeling
Dosimetric parameters of the thyroid gland along with patient clinical information (thyroid gland volume, age, gender, chemotherapy, and clinical stage) were included in the analysis. Univariate logistic analysis for each variable was performed using the Spearman’s rank correlation (Rs) coefficient to assess intervariable correlation and correlation with RHT risk.
We separately analyzed two sets of candidate predictors: set 1 includes the clinical variables, plus D_{min,} D_{max,} D_{mean} and V_{x}(%), and set 2 includes the same variables as set 1 but V_{x} was expressed as absolute volume, V_{x} (cc).
Where x_{ 1 }, x_{ 2 }.…. x_{ n } represent different input variables and β_{ 0 }, β_{ 1 }.…. β_{ n } are the corresponding regression coefficients.
In order to avoid overfitting, when the Rs coefficient between two variables was greater than 0.85 we excluded the one with the lowest correlation with RHT [11] from the subsequent multivariate analysis.
Data analysis was performed by an open source available package (Dose Response Explorer System [12]) for combined modeling of multiple dosimetric parameters and clinical factors using multiterm regression modeling. In summary, the modeling process consists of a twostep process. In a first step, the model size (number of variables significantly predictive) is estimated by bootstrapping and in the second step regression coefficients are estimated using forward selection on multiple bootstrap samples, the most frequent model being the optimal one. Model predictive power is quantified using Rs correlation coefficient while the area under the receiver operating characteristic curve (AUC) was used to evaluate the discriminating ability of model fits.
Subsequently, the obtained NTCP models were validated against an independent external cohort. To this end, data on RHT in breast cancer patients with irradiated supraclavicular lymph nodes were taken from the literature [8]. For comparison purpose, we also evaluated the NTCP model for RHT proposed by Boomsma et al [7] that is based on thyroid gland mean dose and volume. Model comparison was performed using a z test on the AUC of receiver operating characteristic (ROC) curves. A p value less than 0.05 was considered statistically significant. Statistics was performed using MedCalc (MedCalc, Mariakerke, Belgium).
Results and discussion
Models
Bestfitted regression coefficients and 95% confidence intervals for model 1 and model 2
Parameter  Estimated coefficient  StdError  pvalue 

Model 1  
gender  −2.32  0.83  0.0062 
V_{30}(%)  0.038  0.01  0.0009 
constant  −1.83  
Model 2  
gender  −2.21  0.85  0.0110 
V_{30}(cc)  0.26  0.09  0.0021 
thyroid volume (cc)  −0.27  0.11  0.0140 
constant  1.94 
Actually, we can consider the above model 1 and model 2 as equivalent models being V_{30}(%) the ratio of V_{30}(cc) to thyroid gland volume. In our previous work [9] we already found that thyroid V_{30}(%) predicts the risk of developing RHT. However other groups [7, 8] have shown a thyroid gland volume effect in RHT development: the risk increases with smaller thyroid gland volume. For this reason, in this work we separately analyze the V_{x} parameters as percentages and as absolute volumes.
It is interesting to note that both our NTCP models include gender. This result is in agreement with the metaanalysis by Vogelius et al [13] who identified gender, together with race and surgery of the neck, to be as a significant prognostic clinical variable in RHT development.
Models’ comparison and validation
Area under the receiver operating characteristic curve (AUC) and 95% confidence intervals for all the models applied on our Hodgkin’s lymphoma (HL) dataset and on an external breast cancer dataset[8]
Applying model 1 and model 2 to the external casecontrol cohort of breast cancer patients, we have obtained the ROC curves showed in Figure 4b. In this case, model 1 fails to predict RHT (AUC = 0.568, 95% CI 0.3280.741) while model 2 has a high performance (AUC = 0.914, 95% CI 0.768–0.984). This result can be ascribed to the fact that, unlike our patients, the external cohort is characterized by large interindividual and intergroup variations in thyroid volumes. Therefore model 2, where V_{30} is expressed as absolute volume coupled with the thyroid volume, results to be more effective in RHT prediction.
Subsequently, we have analyzed the Boomsma NTCP model for RHT. It should be noted that these authors reported an AUC of 0.85 (95% CI 0.78–0.92) on their headandneck cancer patient dataset. The Boomsma model and model 1 and model 2 performances are not statistically different (p = 0.67) when each is evaluated on its own internal data set.
The ROC curves generated applying the Boomsma model on our HL dataset and on the breast cancer dataset are shown in Figure 4a and 4b, respectively.
On our cohort of patients, the performance of Boomsma NTCP model resulted statistically lower than that of model 1 or model 2 (p < 0.05). Conversely, on validation breast cancer cohort model 2 and Boomsma model have comparably high performance (p = 0.26).
Based on the AUC analysis, both model 2 and Boomsma model seem to be successfully applicable to predict RHT also on a different population.
The difference between the above models relies on the use of V_{30} (cc) and gender for model 2 and on the use of D_{mean} for Boomsma model, while the thyroid gland volume is a common variable. The different selection of dosimetric variables may be ascribed to the relatively high uniform thyroid dose distribution in a headandneck cancer cohort (where up to 70 Gy are prescribed with a V30(cc) probably equal to the thyroid gland volume) compared with thyroid dose distribution in our Hodgkin lymphoma patients treated with a median dose of 32 Gy [14].
Besides the prediction performance, we believe that a model that also considers gender could be advantageous being the estimated rate of hypothyroidism in the general population higher in women than in men [15]. In addition, to explain higher susceptibility of women to RHT, it has been assumed that RT could work as a multiplicative factor that increases the baseline risk of the general population [13]. This could justify the comparable performance of model 2 and Boomsma model when applied on a uniform female cohort as the breast cancer patient dataset, while a lower performance of Boomsam model is observed when it is applied on HL patients where female and male are almost homogenously represented.
In treatment planning optimization procedures, the separate use of thyroid gland volume along with a dosimetric parameter (V_{30}(cc) or D_{mean}) is not easily tunable. In this framework, the “residual 30 Gy thyroid volume” defined as the difference between the thyroid gland volume and V_{30}(cc) may be easier to use. From our HL data, the “residual 30 Gy thyroid volume” was found to be a significant predictor of RHT as well (Rs = 0.56). The median “residual 30 Gy thyroid volume” of patients with RHT was 0.2 cc (range 0.015.6 cc) in contrast to a median value of 9.4 cc (range 0.031.2 cc) for those without RHT. From ROC analyses we have estimated a cutoff volume equal to 7 cc (AUC = 0.81, 95% CI 0.7200.904) for the “residual 30 Gy thyroid volume” as a critical value above which there is a high probability for the thyroid to maintain its functionality. This result is in agreement with the work by Johansen et al [8] where a median residual 30 Gy thyroid volume of 5 cc was found in patients who developed RHT in contrast to a median value of 11 cc in patients who did not develop RHT.
Conclusions
In this study we have developed a multivariate NTCP model for RHT based on dosimetric and clinical variables: the absolute volume of thyroid gland exceeding 30 Gy, thyroid gland volume and gender. This threevariable model provides an improved prediction capability not only within our patient population but also in an external validation cohort. In addition, we have found a cutoff “residual 30 Gy volume” for thyroid gland that should be considered in the treatment planning procedure in order to maintain the gland functionality.
Notes
Abbreviations
 ATG:

Thyroglobulin antibody
 AUC:

Area under the curve
 CI:

Confidence interval
 FT3:

Free triiodothyronine
 FT4:

Free thyroxine
 HL:

Hodgkin’s lymphoma
 NTCP:

Normal tissue complication probability
 RHT:

Radiationinduced hypothyroidism
 ROC:

Receiver operator characteristic
 Rs:

Spearman’s rank correlation
 RT:

Radiation therapy
 TSH:

Thyroid stimulating hormone.
Declarations
Acknowledgements
The authors acknowledge partial support from Italian Ministry for Education, University and Research (MIUR) in the framework of FIRB (RBFR10Q0PT_001 “DROPS” and RBNE08YFN3 “MERIT”).
Authors’ Affiliations
References
 Alterio D, JereczekFossa BA, Franchi B, et al.: Thyroid disorders in patients treated with radiotherapy for headandneck cancer: a retrospective analysis of seventythree patients. Int J Radiat Oncol Biol Phys 2007,67(1):144150. 10.1016/j.ijrobp.2006.08.051View ArticlePubMed
 Bethge W, Guggenberger D, Bamberg M, et al.: Thyroid toxicity of treatment for Hodgkin’s disease. Ann Hematol 2000,79(3):114118. 10.1007/s002770050565View ArticlePubMed
 Reinertsen KV, Cvancarova M, Wist E, et al.: Thyroid function in women after multimodal treatment for breast cancer stage II/III: comparison with controls from a population sample. Int J Radiat Oncol Biol Phys 2009,75(3):764770. 10.1016/j.ijrobp.2008.11.037View ArticlePubMed
 Trott KR, Doerr W, Facoetti A, et al.: Biological mechanisms of normal tissue damage: Importance for the design of NTCP models. Radiother Oncol 2012,105(1):7985. 10.1016/j.radonc.2012.05.008View ArticlePubMed
 Defraene G, Van den Bergh L, AlMamgani A, et al.: The benefits of including clinical factors in rectal normal tissue complication probability modeling after radiotherapy for prostate cancer. Int J Radiat Oncol Biol Phys 2012,82(3):12331242. 10.1016/j.ijrobp.2011.03.056View ArticlePubMed
 El Naqa I, Bradley J, Blanco AI, et al.: Multivariable modeling of radiotherapy outcomes, including dosevolume and clinical factors. Int J Radiat Oncol Biol Phys 2006,64(4):12751286. 10.1016/j.ijrobp.2005.11.022View ArticlePubMed
 Boomsma MJ, Bijl HP, Christianen ME, et al.: A prospective cohort study on radiationinduced hypothyroidism: development of an NTCP model. Int J Radiat Oncol Biol Phys 2012,84(3):e351e356. 10.1016/j.ijrobp.2012.05.020View ArticlePubMed
 Johansen S, Reinertsen KV, Knutstad K, et al.: Dose distribution in the thyroid gland following radiation therapy of breast cancer–a retrospective study. Radiat Oncol 2011, 6: 68. 10.1186/1748717X668PubMed CentralView ArticlePubMed
 Cella L, Conson M, Caterino M, et al.: Thyroid V30 predicts radiationinduced hypothyroidism in patients treated with sequential chemoradiotherapy for Hodgkin’s lymphoma. Int J Radiat Oncol Biol Phys 2012,82(5):18021808. 10.1016/j.ijrobp.2010.09.054View ArticlePubMed
 Cella L, Liuzzi R, Magliulo M, et al.: Radiotherapy of large target volumes in Hodgkin’s lymphoma: normal tissue sparing capability of forward IMRT versus conventional techniques. Radiat Oncol 2010,5(1):33. 10.1186/1748717X533PubMed CentralView ArticlePubMed
 Huang EX, Bradley JD, El Naqa I, et al.: Modeling the risk of radiationinduced acute esophagitis for combined Washington University and RTOG trial 9311 lung cancer patients. Int J Radiat Oncol Biol Phys 2012,82(5):16741679. 10.1016/j.ijrobp.2011.02.052PubMed CentralView ArticlePubMed
 El Naqa I, Suneja G, Lindsay PE, et al.: Dose response explorer: an integrated opensource tool for exploring and modelling radiotherapy dosevolume outcome relationships. Phys Med Biol 2006,51(22):57195735. 10.1088/00319155/51/22/001View ArticlePubMed
 Vogelius IR, Bentzen SM, Maraldo MV, et al.: Risk factors for radiationinduced hypothyroidism: a literaturebased metaanalysis. Cancer 2011,117(23):52505260. 10.1002/cncr.26186View ArticlePubMed
 Cella L, Conson M, Liuzzi R, et al.: In Regard to Boomsma et al. Int J Radiat Oncol Biol Phys 2013,85(1):11. 10.1016/j.ijrobp.2012.09.009View ArticlePubMed
 Gharib H, Tuttle RM, Baskin HJ, et al.: Subclinical thyroid dysfunction: a joint statement on management from the American association of clinical endocrinologists, the American thyroid association, and the endocrine society. J Clin Endocrinol Metab 2005,90(1):581585. discussion 586587View ArticlePubMed
Copyright
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.