Direct Medical Costs of Advanced Breast Cancer Treatment: A Real-World Study in the Southeast of The Netherlands

Objectives Policy makers increasingly seek to complement data from clinical trials with information from routine care. This study aims to provide a detailed account of the hospital resource use and associated costs of patients with advanced breast cancer in The Netherlands. Methods Data from 597 patients with advanced breast cancer, diagnosed between 2010 and 2014, were retrieved from the Southeast Netherlands Advanced Breast Cancer Registry. Database lock for this study was in October 2017. We report the observed hospital costs for different resource categories and the lifetime costs per patient, adjusted for censoring using Lin’s method. The relationship between patients’ characteristics and costs was studied using multivariable regression. Results The average (SE) lifetime hospital costs of patients with advanced breast cancer were €52 709 (405). Costs differed considerably between patient subgroups, ranging from €29 803 for patients with a triple-negative subtype to €92 272 for patients with hormone receptor positive and human epidermal growth factor receptor 2 positive cancer. Apart from the cancer subtype, several other factors, including age and survival time, were independently associated with patient lifetime costs. Overall, a large share of costs was attributed to systemic therapies (56%), predominantly to a few expensive agents, such as trastuzumab (15%), everolimus (10%), and bevacizumab (9%), as well as to inpatient hospital days (20%). Conclusions This real-world study shows the high degree of variability in hospital resource use and associated costs in advanced breast cancer care. The presented resource use and costs data provide researchers and policy makers with key figures for economic evaluations and budget impact analyses.


Introduction
Breast cancer is the most common type of cancer among women and a leading cause of death worldwide. 1 Despite significant improvements in the early diagnosis and treatment of breast cancer, about 5% of the patients present with advanced (ie, metastatic) disease at the time of diagnosis, and a further 20% of patients will experience progression to advanced disease later in life. 2 Even though some patients live with advanced breast cancer (ABC) for many years, the disease is considered incurable, and the main objective of care is to prolong survival and sustain quality of life. Due to its high prevalence and high individual treatment costs, the economic burden of ABC in The Netherlands, as well as in many other countries, is substantial. [3][4][5][6] Over the last 2 decades, several new drugs for the treatment of ABC have become available. Many are highly expensive, posing a substantial economic burden on society and raising questions regarding their value for money. 6,7 Health economic evaluations are essential to inform reimbursement decisions for these novel agents.
Although randomized controlled trials (RCT) are considered the gold standard for supporting preapproval efficacy, their value for health economic evaluations and real-world effectiveness is limited. Resource utilization is rarely being considered in RCTs, and even if relevant data are collected, treatments are often tested under artificial circumstances. In routine care, patients often have more comorbidities, lower compliance, are older, and/or treatment patterns are different from patients included in RCTs. 8 Furthermore, a significant proportion of the resource utilization may fall beyond the trial horizon. In health economic evaluations, healthcare costs frequently have to be based on expert opinions and/or taken from different patient populations and other settings. Therefore, policy makers increasingly seek to complement data collected in RCTs with data from routine care. 9 Under real-world conditions, ABC care is complex: patients are highly heterogeneous and treatment choices and pathways are individualized and depend on patient characteristics, treatment responses, and preferences. 10 Therefore, aggregate cost estimates computed from administrative data are of limited value, and decision makers should aim to take into account all the relevant factors to prescribe optimal policies. Up until today, however, little is known about the real-world hospital resource use and the associated costs in patients with ABC. 11,12 The complexities of ABC care require the information to be comprehensive, granular, and contextual, to guide decision making in The Netherlands and in other countries.
This study aims to investigate the real-world resource use and costs of ABC in The Netherlands from the hospital perspective. First, we assessed patients' resource utilization and the associated costs; second, we estimated the average lifetime costs of patients with ABC; and, finally, we investigated which factors contribute to the heterogeneity of costs between patients.

Patient and Data Collection
This study used patients included from the Southeast Netherlands Advanced Breast Cancer (SONABRE) Registry. 13 The ongoing registry was initiated in 2010 and aims to include all patients with de novo or recurrent ABC, diagnosed at age 18 years or older, from 12 participating hospitals in the southeast of The Netherlands. For this study, we used data from the 5 hospitals, in which information on resource use was collected from 2010 through 2017. These hospitals were selected with the intention to obtain a representative mixture of different hospital types and sizes, and they account for approximately 7% of the hospitals in The Netherlands. Patients were included if they were diagnosed with ABC between 2010 and 2014. No exclusion criteria were employed. Data lock was October 23, 2017.
Patients were identified prospectively, and clinical data were, retrospectively, retrieved manually from electronic medical records by trained registration clerks and entered into a registry database. Collected patient and disease characteristics included age, survival time, comorbidities, and tumor characteristics (initial hormone receptor [HR] and human epidermal growth factor receptor 2 [HER2] status). Moreover, we retrieved data on hospital resource use for the following categories: medicines/systemic therapies (chemotherapy, endocrine therapy, targeted therapy, bone-modifying agents/bisphosphonates, and transfusions), consultations/hospitalizations, radiotherapy, and diagnostic and surgical procedures.

Handling of Incomplete and Missing Data
Because our study was based on routinely collected data, missingness was unavoidable. This problem was mostly limited to medications, for which the administered dosage was missing, while the number of administrations was reported. To avoid the creation of implausible observations, we used hot deck imputation to replace missing values with observed values from another patient, matched by the resource they used. 14 For hormonal therapies and bone-modifying agents/bisphosphonates, consumed resource units were not recorded in the registry but were computed based on treatment durations and respective standard doses. 15 When a patient's treatment duration was not reported, resource consumption was imputed using the agent-specific average duration. If the HER2 (n = 58) and/or HR (n = 3) receptor status is not tested, patients are being treated as if they had negative receptor status. In our study, such patients were classified accordingly as triple negative (TN).

Resource Use and Associated Costs
For all patients, we assessed the hospital resource use and associated costs from the date of diagnosis (ie, date of pathological conformation, or else date of imaging) until date of death or censoring. Costs associated with resource consumption were derived by multiplying the units of resource consumption with the respective cost prices. Relevant unit costs were taken from Dutch costing guidelines, medicijnkosten.nl, and, if otherwise unavailable, from individual studies or the financial department of Maastricht University Medical Centre (see Appendix Table 1 in Supplemental Materials found at https://doi.org/10.1016/j.jval.2020. 12.007). 16,17 All costs are expressed in 2017 Euro. If necessary, costs were inflated to the price level of this year, using the consumer price index. 18 For each type of resource, we assessed the average consumption and the average costs, conditional and unconditional on resource use, as well as the share on total costs. Since cost estimates are provided for use in health economic decision modeling, we only report averages and their bootstrapped 95% confidence intervals, which are commonly used in models, instead of, for example, the median, as a more stable measure of central tendency. For drugs, consumption is expressed in terms of the most commonly administered dosage. The hospital costs associated with administering drugs intravenously are reported separately, as they have been previously reported to account for a considerable share of total costs. 19

Lifetime Costs
We calculated the average costs per patient, as well as the costs per patient month. Since survival time and lifetime costs tend to be correlated, the observed costs per patient are not an unbiased estimate of lifetime costs, in the presence of censoring. Including patients based on the date of diagnosis enabled us to apply adjustment methods to extrapolate the observed costs for any patient who was censored over a lifetime horizon. To adjust the costs for censoring, we used the approach described by Lin et al 20 : the observation period was split into intervals, based on dates of death and censoring. For each interval, we computed the mean and weighted it by the Kaplan-Meier estimator (ie, the probability that patients survive until the beginning of the interval). We then summed up the weighted means to derive an estimate of mean lifetime costs of ABC. To compute standard errors around means, we used bootstrapping with 1000 iterations. When appropriate, we refer to the observed costs as being unadjusted.

Heterogeneity of Costs Between Patients
We further investigated the heterogeneity in costs, that is, the share of the variability that can be explained by patient and tumor characteristics. The following factors, identified from the literature and/or clinical expertise, were taken into consideration 12 : survival time, age, year of diagnosis, interval between primary diagnosis of breast cancer and diagnosis of metastatic disease, initial HER2 and HR status, death (in versus outside of a hospital), systemic therapy (was any systemic therapy initiated?), and locoregional aggressive treatment (defined as breast surgery or radiotherapy with 15 or more fractions within the first year after diagnosis of metastatic disease).
We investigated the association between these variables and 2 outcomes: observed total costs per patient and costs per patient month. Although the former was the main focus of this study, survival time was assumed to explain a large proportion of the variability in total costs, so that inference about other variables may be limited. For both outcomes variables, we fitted generalized linear models, with a gamma distribution and a log link. This method, which is frequently used to model cost data, was chosen to evade the shortcomings of ordinary least squares, specifically with regard to right-skewed data, heteroscedasticity, and the strictly non-negative values in cost data. 21 An ordinary least squares model may, for example, predict negative expenditure (ie, gains) for short periods of survival time. We used backward elimination to successively remove predictors and find the model with the lowest Bayesian information criterion. 22 As a goodnessof-fit measure, we report the McFadden's pseudo R 2 value. A sensitivity analysis was conducted to investigate whether effect estimates for the subsample of deceased patients differed compared to the full cohort (including deceased and censored patients). Alternative generalized linear model specifications (Gaussian distribution with a log link and Gaussian distribution with an identity link) were also tested to validate our model choice.

Ethical Approval
The Medical Research Ethics Committee of Maastricht University Medical Centre1 reviewed and approved the SONABRE Registry. The need for informed consent was waived because of the observational nature of this study.

Patient Population
After the application of the inclusion criteria, 597 patients from the SONABRE registry were included in the study. The number of patients per year of ABC diagnosis varied only a little between 2010 and 2014, with the minimum being in 2014 (n = 103; 17%) and the maximum being in 2011 (n = 128; 21%). Overall, 436 patients died and 161 were censored. The median survival time was 24.5 months (95% CI 22.7-27.5). The median follow-up time (ie, time until censoring) was 55.2 months (51.7-58.7). A large majority of the patients in our cohort had a HR1/HER2-receptor status at the time of the initial diagnosis (n = 417; 70%), followed by TN (n = 69; 12%); HR1/HER21 (n = 65; 11%), and HR-/HER21 (n = 46; 8%). A total of 199 patients (33.3%) had any of the following comorbidities: metabolic disease (n = 90; 15%), cerebral disease (n = 32; 5%), cardiovascular disease (n = 63; 11%), other malignancy (n = 62; 10%), or pulmonary disease (n = 43; 7%). Further patient characteristics are provided in Table 1. Table 2 provides an overview of the hospital resource use and the associated costs in our cohort as observed during the study period. Shown are the number of patients who used a particular resource and the associated share in total costs, as well as the average units of resource consumption and the average costs per patient, conditional on resource use (ie, the average for those patients, who used the resource). In addition to the figures for aggregate resource categories, data on individual resource items are reported if the respective share in total costs was at least 1%. Total costs refer to the sum of costs of all included patients over the entire observational period. For around 1.2% of the consumed resources, the number of consumed units was not reported and had to be imputed. HR/HER2 subtype-specific resource use and cost figures are provided in Appendix Tables 2-5 in Supplemental Materials found at https://doi.org/10.1016/j.jval.2020.12.007.

Resource Use and Associated Costs
A few specific points about the results shown in Table 2 are worth highlighting: First, the majority of costs (56%) of ABC treatment were related to systemic therapies. The main drivers of these costs were targeted therapies (37%) and bone-modifying agents/bisphosphonates (6%), whereby the individual agents trastuzumab (15%), everolimus (10%), bevacizumab (9%), and pamidronic acid (2%) accounted for the highest shares in costs. Second, of the 597 patients included in the study, a vast majority (89%) received 1 or multiple systemic therapies. Third, most patients had frequent contact with healthcare providers, and utilization of these services accounted for 30% of total costs. On average, outpatient clinics were visited about 26 times per patient, and 458 (77%) patients were admitted to a hospital at least once; admission was, on average, 18 inpatient days. Finally, many patients underwent comprehensive diagnostic testing and had multiple radiographic examinations (eg, CT scans, MRIs, x-rays), which contributed to the share of 9% on total costs. Table 3 provides the mean costs per patient with ABC, both unadjusted (ie, as observed) and adjusted for censoring of patient time and costs. The unadjusted mean (95% CI) total costs per patient were V44 277 (41 949-44 277). The distribution of costs per Heterogeneity of Costs Table 4 shows the predictors retained in the final generalized linear models for total per patient costs and costs per patient month. Consistently for both, death in-and outside hospital and an HR/HER2 receptor status other than HR1/HER2were associated with higher costs, and patient age at diagnosis with lower costs. In addition, systemic and locoregional aggressive therapy were retained as predictors of higher costs and cerebral comorbidity was retained as a predictor of lower costs in the model to explain total per patient costs, but not in the costs per patient month. Moreover, survival time showed a negative association with costs per patient month (mean coefficient estimate = -0.23), but a positive association with total costs per patient (= 0.31). A sensitivity analysis only including

Discussion
Our study provides an overview of the real-world costs of patients with ABC in The Netherlands. After adjusting for censoring, the average (SE) lifetime hospital costs of ABC were estimated to be V52 709 (405). However, there was large variation in these costs between the HR/HER2 subtypes, ranging from V29 803 (1130) for patients with TN ABC to V92 272 (910) for patients with HR1/HER21 ABC. Our analyses further revealed that, underlying this variation, there were considerable differences in the structures of costs. In all groups, medicines, and more specifically only a few expensive agents, accounted for a large share of total costs. In addition to the HR/HER2 receptor status, several other patient characteristics were independently associated with the costs per patients and/or per patient month, including age, survival time, locoregional aggressive treatment, death in hospital, and cerebral comorbidity.
To our knowledge, this is the first study to investigate the realworld lifetime costs of ABC in a cohort, including patients with any HER2/HR receptor status as well as de novo and recurrent ABC, in The Netherlands. A previous real-world study by Frederix et al 23 included only 88 patients with ABC HER21, treated in 3 hospitals in The Netherlands between 2004 and 2010. They estimated the lifetime costs to be V48 996. Even though the data collection and analysis methods were similar to ours, the reported resource use and cost estimates were much lower than what we found for both patients with the HR1/HER21 (V92 272) and the HR-/ HER21 subtype (V69 079). Apart from the shorter follow-up time of 2 years for each patient in the study by Frederix et al, the increasing availability of targeted therapies in recent years should be considered a main reason for the apparent difference in ABC lifetime cost estimates. 23 Van Kampen et al 24 also used resource Table 3. Costs of patients with advanced breast cancermean (95% confidence interval) in V.     27 Differences in the organization of health service delivery and in drug reimbursement prices between countries are likely to contribute to the apparent differences in costs. 4 Nevertheless, across settings, inpatient days and costs for medicines, in particular for trastuzumab, which is specifically highlighted in all 3 aforementioned studies, are found to be main drivers of ABC lifetime costs. 12,23,27 The patient cohort included in our study is somewhat different from other epidemiological cohorts, such as the French ESME 28 : half of the metastases in the ESME population were found asymptomatically (by screening), whereas screening for distant metastases is not standard practice in The Netherlands. Furthermore, the ESME cohort excludes patients not systemically treated, whereas all newly diagnosed patients, including the 11% of patients who did not receive any systematic treatment, were included in SONABRE. In addition, ESME includes comprehensive cancer centers, whereas a mixture of hospital types are participating in SONABRE. These differences could all explain the lower median overall survival observed in our SONABRE population (ie, 24 months) when compared to, for example, the ESME cohort (ie, 37 months).
In contrast to previous cost of illness studies, which either focused on a selective ABC population (eg, including only patients with HER2 positive 23 or negative 24 ABC) or had only a small sample size, 11,27 we provide precise cost estimates for a general cohort of patients with ABC in The Netherlands, as well as for the most relevant subgroups. Continued consistent data collection by the SONABRE Registry will ensure that our findings can be updated and refined in the future.
There are also several limitations to our study that deserve mention. We investigated costs of ABC from the hospital perspective and collected data in 5 hospitals. Resources that were used in other healthcare sectors (eg, primary care, nursing care, hospice) or in other areas of society (eg, education, judiciary, productivity losses) were not taken into consideration. Hospital resource use was also only collected in 5 hospitals in the southeast of The Netherlands, and even though we do not have any reason to expect major systematic differences, those may not be representative of the country as a whole. Moreover, we did not account for vial wastage, and due to the retrospective data collection, some resource uses may have been missed, which may have led to an underestimation of costs. However, because of the financial implications for the hospitals, the amount of not recorded resource consumption is probably low. For some types of resources, especially for surgical interventions, reference prices 17 were not available, and it was not always possible to retrieve unit costs from other publicly available sources. In several instances, we had to use internal cost prices from the Maastricht University Medical Centre1, which are confidential and cannot be reported. In other instances, we had to impute missing resource use information, using the hot deck imputation method. Even though the amount of missing data was relatively small, and it is unlikely that it introduced relevant bias into the analysis, it should be noted that the method has important limitations. First, we did not account for important patient characteristics and may have thus generated implausible resource use patterns (eg, use of mutually exclusive treatments). Second, hot deck imputation may lead to an underestimation of the uncertainty around mean estimates. In addition, there are several methods available to estimate lifetime costs in the presence of censoring. Lin's method has shown to provide accurate results, 20 but up until now, it is not clear whether a different method would have been more precise. Since some of our analyses were conducted using the observed (ie, unadjusted) patient costs instead of adjusted lifetime costs, results might have also been affected by the censoring of patient time and costs. To account for this in the multivariable regression model, we included death, inside or outside of the hospital, as an explanatory factor, but in other analyses, resource use and costs were probably underestimated. The reported costs thus represent a conservative estimate and should be interpreted as a lower bound. Finally, it should be noted that reported resource use and cost estimates may be context specific. Treatment patterns can be expected to differ between countries and change over time. During the study period, modern, expensive targeted agents, such as CDK4/6 inhibitors in HR1/HER2-ABC, pertuzumab and T-DM1 in HER21 ABC, and atezolizumab in TN ABC, were available only to a small proportion of patients or were not available at all. The increasing use of these treatments may have already affected the management and the lifetime costs of patients with ABC in The Netherlands, and it can be expected that this trend will continue with the advent of new therapies in the next few years.
Our study provides essential information about the real-world hospital costs of ABC that provides insight into the complex structure of costs in the heterogenous population of patients with ABC. We report several complementary measures of costs, facilitating a comparison of adjusted and unadjusted lifetime costs and monthly costs per patient, which can be used by others in health economic evaluations and to inform health policy. For comparisons with current or past costs, cost estimates can be adjusted using the respective consumer price index. 18 We also identified several factors that were independently associated with the total hospital costs per patient: in addition to HR/HER2 receptor status, this includes age at diagnosis, survival time, death inside or outside the hospital, any systemic and locoregional aggressive treatment, and the presence of a cerebral comorbidity. When combined with effectiveness data and put into context, the reported estimates can help to improve the quality of decision analytical models and enable more precise subgroup analysis in patients with ABC, which, ultimately, can help inform sound decision making. More research is required to better understand if, and if so, which factors are predictive of healthcare spending in patients with ABC. Future studies should aim to also take into account the longitudinal structure of healthcare costs over time and investigate its temporal dynamics. Furthermore, real-world cost data from outside the health sector are required to complement the information reported in this study.

Conclusion
We investigated the real-world hospital cost of patients with ABC in The Netherlands. The comprehensive description of resource use and associated costs provides researchers and policy makers with key figures for economic evaluations and budget impact analyses. Our analyses offer new insights into the structure and clearly shows the large heterogeneity of hospital costs of patients with ABC in The Netherlands. A better understanding of the real-world costs of ABC will be increasingly important to inform priority setting and resource allocation in healthcare, as novel and expensive therapies become available.