Population and data

The study population consisted of all women resident in England, enrolled in UKCTOCS, who were diagnosed with EC between April 2001 and December 2009, and for whom HES records were available. Patients with a concurrent diagnosis (synchronous) of primary peritoneal or ovarian cancer in addition to primary endometrial cancer, missing stage or incomplete HES data (possibly due to patients transferring their care outside the English NHS) were excluded.

The notification of EC diagnosis was received from multiple sources, the cancer registries (through the Health and Social Care Information Centre (HSCIC) and HES (Hospital Episode Statistics) which were searched using the ICD-10 codes of C54 (malignant neoplasm of corpus uterus) and C55 (malignant neoplasm of uterus, part unspecified), from self-reporting on the postal follow-up questionnaire or direct contact from the volunteer or their physician. Hospital records which included surgical and pathological reports were retrieved and independently reviewed by a clinician using an Outcomes Review form who confirmed EC diagnosis, stage using the FIGO classification,[26] grade and histological subtype.

UKCTOCS trial data provided patient characteristics including date of birth, Body Mass Index (BMI) collected at recruitment, hospital records of women diagnosed with EC and date of death. For confidentiality reasons only the month and year was available for the date of birth and date of death; each entry was arbitrarily assigned to the 15^th of the relevant month.

HES provided date of admission, date of transfer or discharge, diagnosis (as ICD-10 codes),[27] procedures undertaken (coded using the Office of Population Censuses and Surveys (OPCS) Classification of Surgical Operations version 4 codes),[28] treatment specialty, Healthcare Resource Group (HRG)[29] codes for reimbursement, and patient deprivation (measured as the Index of Multiple Deprivation (IMD) which is a geographically determined index based on areas of approximately 1500 people).[30]

We included all HES inpatient and outpatient episodes related to EC. Assignment for inpatient procedures was based primarily on OPCS-4 procedure codes taking into account the HRG code and treatment speciality. Outpatient records were less complete and procedure codes were rarely available. Outpatient records lacking procedure codes were only included where the fields coding the main specialty or the treatment speciality contained a code relating to either oncology or gynaecology. Investigations occurring within six months prior to diagnosis were included. Procedures were further classified into surgery (hysterectomy, BSO and pre-surgical investigations), adjuvant therapy, and further treatment (including management of complications).

Estimating costs of inpatient and outpatient episodes

Costs were assigned to episodes of care on the basis of the associated HRG code.[29] This is a type of Diagnosis Related Groups (DRG) used to assign a reimbursement tariff in the English NHS. Version (3.5) was used, which allowed assignment of the 2005 Payment by Results (PbR) tariff to episodes from the entire period 2001 to 2010.[31] We then inflated the cost to 2012/13 prices using the Hospital & Community Health Services Index.[32] Finally we multiplied each cost by 1.08 which is the average Market Forces Factor for Hospitals in the English NHS.[33] Market Forces Factors are applied to payments generated from HRG codes to adjust payments for unavoidable differences in costs relating to the geographic location of each hospital.[34]

In addition to estimating cost on the basis of HRG code we adjusted costs for extended length of stay. A maximum length of stay is specified for each HRG code and length of stay beyond this point is reimbursed at a daily tariff specific to that HRG. Where patient length of stay during a spell in hospital exceeded the maximum specified for the HRG code we applied the appropriate excess bed day adjustment to the estimated cost.

PbR tariffs distinguish inpatient and outpatient procedures, and (for inpatient procedures) elective versus emergency admissions. We assumed that all inpatient admissions were elective procedures as we lacked data on the type of inpatient admission. Hospital admissions are remunerated per spell (from admission to discharge). Generally a spell consists of one episode of care, but occasionally it may consist of more than one episode. In the latter situation hospitals are remunerated according to the dominant procedure for the component episodes (typically the most expensive). We applied the costs for the most expensive episode for spells of care consisting of multiple episodes.

Inpatient episodes with missing HRG code (20%) were assigned a code on the basis of OPCS procedure codes. The majority of outpatient episodes had no HRG or OPCS code assigned. We identified records of chemotherapy and radiotherapy on the basis of the interval between admissions and the assigned treatment specialty (medical or clinical oncology). These records were assigned the appropriate outpatient PbR tariff for radiology or chemotherapy. The remaining outpatient episodes were assigned the average outpatient cost for 2012/13 of £135 per episode.[32]

Censoring of cost data

We had missing cost data in the form of censored observations where five year follow-up date and date of death exceeded 31^st March 2010 (data available from HES). We had further missing data on histological subtype (0.4%), grade (3.2%), IMD (0.4%), and BMI (0.8%). IMD scores were used to partition patients into quintiles. We applied a binary classification of histological subtype which grouped Atypical Endometrial Hyperplasia (AEH) and Endometrioid EC versus all other EC histological subtypes. We used multiple Imputation (MI) to impute the missing data.[35] MI is a principled approach which fully captures the additional uncertainty generated in the imputation process.[36–37] Further details on the application of MI are provided in the supporting information (S1 File).

Analysis of cost data

Costs arising in years following the first year after diagnosis were discounted at 3.5% according to recommendations on technology assessment by the National Institute of Health and Care Excellence (NICE).[38] FIGO stage was specified as follows: stages IA & IB; stage IC; stage II; stage III; stage IV. We subdivided costs into three categories: surgery; adjuvant therapy and further treatment as detailed above. Costs accumulated at two years after diagnosis are reported for the subset of patients with a minimum of two years follow-up in HES (diagnosis prior to 31^st March 2008). Total costs (after imputation of missing data) were determined for all patients at five years from diagnosis, with non-parametric 95% confidence intervals estimated from 1,000 bootstrap replicates of the data (details in appendix). Costs accumulated at five years after diagnosis for patients with complete cost data (diagnosis prior to 31^st March 2005) are provided in the supplementary material (Table A in S1 File).

Regression modelling was undertaken to explore the impact of patient characteristics on cost. We fitted Ordinary Least Squares (OLS) regression models to each of the three cost categories and to total costs. We pre-specified the following covariates: age, year of diagnosis, histological subtype (as Endometrioid Carcinoma/AEH or other), grade, stage, IMD quintile, Charlson score[39] and BMI. We categorised BMI as under 18.5; 18.5 to 30; over 30. Year of diagnosis was specified as the number of years following the earliest diagnosis date in the sample (January 2002). Charlson scores were determined from ICD-10 codes recorded for the hysterectomy (or the first inpatient procedure following diagnosis in the absence of hysterectomy) after exclusion of codes for cancer or metastases.

Analysis of survival data

Dates of death were available until June 2014. Deaths were assigned to EC where the original underlying cause of death was EC (ICD-10 codes C54.0, C54.1). Regression modelling was used to investigate the impact of patient characteristics on survival over the total observation time considering both all cause and cancer specific mortality. Cox proportional hazards models were fitted to the data that adjusted for the same covariates selected for the cost analysis (detailed above). A Therneau and Grambsch test was applied to assess the appropriateness of an assumption of proportional hazards.[40]

All statistical analysis was undertaken in Stata, version 13.1.[41]

Main findings

This is the first study that we are aware of that demonstrates increasing cost of treatment for EC in women diagnosed with advanced stage disease. We show that treatment costs are nearly three times as high for cancers detected at stage III compared to stage IA or IB with majority of costs for stage I / II EC incurred in the first six months after diagnosis whereas for stage III / IV considerable costs accrued after the first six months. Consequently, in addition to survival advantages there are significant cost savings if patients with EC were detected earlier. The differences in cost arise from the increased need for adjuvant therapy and increased costs of complications and follow-up procedures in patients diagnosed with advanced stage disease. These findings are important in the consideration of the cost-effectiveness of screening for EC.

Strengths and limitations

Our study reported on EC management costs of a population based cohort of women participating in an ovarian cancer screening trial. The women are therefore more representative of those in the general population than a hospital based series. Data on treatments is drawn from a large administrative database representing clinical practice across the NHS in England, rather than any particular trial protocol. Data collected in UKCTOCS allowed access to independently reviewed stage, grade and histological subtype at diagnosis, and BMI for each woman alongside data on deprivation and comorbidities captured from HES. As a result we were able to explore the patient and tumour characteristics that have been shown to influence survival, and might be expected to influence cost.

Whilst resource use data from administrative databases is deemed to be representative of clinical practice, this source is subject to some limitations. HES only collects data on women treated by the NHS in England; women opting for privately funded treatments and those undergoing treatment in hospitals in Wales and Scotland would not have those treatments captured in HES. As a result, we observed incomplete data for 16% of women who we knew from hospital note review had surgery and had to exclude them from the analysis. Research has shown that the quality of data capture in HES is good, but it is unlikely to match that of a clinical trial [42–43]. Moreover, it is likely that not all treatments related to EC were captured and this will have resulted in underestimation of costs. It is possible that the rise in cost over the period 2002–2010 observed in the regression analysis reflects improvements in data capture in HES rather than changes in treatment protocols.

The cost estimates presented here are based on a macro-costing approach in which costs are assigned to the dominant procedure for a spell in hospital. We were able to adjust costs for excessive length of stay in hospital but we did not adjust costs for any other variation in treatments. This approach is the recommended approach for costing HES in a recent comparison of methods [44] and represents the way in which hospitals in the NHS are remunerated.

The available HES data was to the period ending 31^st March 2010 and consequently we did not have complete data on all women in the cohort. This is a common problem in assessing longer term costs of illness. Various methods have been developed to adjust for censoring in cost data and recover unbiased estimates of total cost [45–47]. We used a principled approach which has previously been applied to missing cost data [35,48]. Use of MI allowed regression analysis of total cost with appropriate consideration of the additional uncertainty introduced by imputation, and also allowed us to incorporate additional data on mortality (available until June 2014), an important predictor of cost.

Our analysis did not include the costs of primary care visits and referrals/all investigations as our study was limited to HES data. We did however cost all investigations documented in HES as occurring in the six months prior to diagnosis. We limited cost analysis to the five year period following diagnosis. It is possible that this underestimates costs for patients diagnosed at stages II and III who die of the disease after five years. The cost data presented here describe treatment related to EC, the impact of EC on unrelated medical costs has not been quantified [49]. We would expect this impact to be negative (cost saving) in as far as EC reduces the life expectancy of women and hence their potential to consume unrelated medical care. Finally, our analysis did not consider quality of life. Earlier diagnosis and treatment may lead to improved quality of life as well as survival.

Interpretation

Previous studies on the cost of management of EC are predominantly from the US and have reported hospital costs for hysterectomy [15–22]. Costs range from $3790 (possibly $1996)¹⁶ to $36,487 (possibly $2011) [19]. Data on stage in these studies is frequently unavailable and none report costs by stage. Further limitations include short follow-up, and the use of hysterectomy to identify patients, which is likely to lead to under representation of patients with late stage disease. Our study benefitted from use of a population based cohort of EC patients, thus avoiding selection bias, detailed data on patient and tumour characteristics and long term follow-up. The cost gradient across stages is larger. The findings are important for future assessment of the cost-effectiveness of screening interventions. They indicate the potential for significant reduction in costs through earlier detection and down-staging particularly from stage III/IV to stage I/II. This is particularly relevant in view of the recent data from UKCTOCS indicating a possible reduction of ovarian cancer mortality with screening [50]. Should this reduction be confirmed on further follow up, the cost-effectiveness of ovarian cancer screening will be a driving factor in implementing a national screening programme. Although endometrial cancer screening in the general population is currently not warranted [51], earlier detection may be achieved within an ovarian cancer screening programme and therefore lead to lower costs of EC treatment.

We found a modest increase in mortality associated with lower socio-economic status, a finding in line with the broader literature on cancer survival [52,53]. In contrast to some reports from the US [54,55] we found a modest increase in resource use with increasing deprivation, which suggests that any mortality difference do not arise from access to treatment. Our study had the advantage of access to patient data on BMI and comorbidities, allowing us to control for these potential confounders. It is possible that some deaths in women with EC were wrongly attributed to cancer and our observations are attributable to elevated non-cancer mortality in deprived populations which has been previously reported in the UKCTOCS cohort [56].