Developing a platform to investigate the heterogeneity of outcomes for patients with ovarian cancer

Identification of centres

Six UK sites were carefully selected to include large and small centres, and centres from Cancer Alliances with high and low surgical resection rates identified from the OCAFP.1 The participating centres were selected to reflect the full range of UK practice and included one centre which lies above the 99th centile for resection rates, and one centre lying below the 1st centile, with four centres representing mainstream practice.

A data dictionary was generated to define all clinical variables to be collected. Details of the dictionary have been previously described.4 Briefly. the data dictionary comprises (1) patient factor data which include demographic, comorbidity using the validated ACE27 system,5 and germline BRCA status, (2) tumour-related data, including histological type, grade and stage, physiological response data and radiological distribution of disease and (3) treatment-related data including type and outcome of each treatment modality. Deprivation indices were generated from patient postcode using the Indices of Deprivation Score.6 Three-year follow-up data were also recorded.

Each participating centre was funded to collect data, according to the data dictionary, for all patients registered with a diagnosis of ovarian cancer (ICD56), between 1 January 2018 and 31 December 2019.

Data were transferred to the University of Manchester and stored in a data repository in a pseudonymised fashion. Missing data were managed with median replacement. Data were analysed with Kaplan-Meier, Cox proportional hazard and multivariable Cox models using Rstudios (V.4.2.3).

As part of the data collection process, sites were also asked to complete a resource requirements document, to better understand the time taken to obtain the data at each site. This information was not requested from the one centre, as data collection was complete prior to the commencement of this project.

Data completeness

A total of 1117 patient records were submitted, representing all cases registered with each centre for the 2-year period from 1 January 2018 to 31 December 2019. The numbers registered per centre were 94, 115, 123, 173, 304 and 308, respectively. The numbers per centre varied, as expected, dependent on the varying sizes of the local population.

Performance status, CA125, Deprivation Index, Federation Internation Gynaecology & Obstetrics (FIGO) stage and grade were generally well-recorded and available for analysis. In contrast, while ACE27 (Adult co morbidity evaluation) scores and ethnicity were well recorded in some centres, they were very poorly recorded in others (table 1). Although germline BRCA status was also heterogeneously recorded it should be noted that the study period coincided with the national rollout of germline BRCA testing with centres taking this on at various points during the study period (table 1).

Data collection processes

The resource required to carry out retrospective data collection is rarely calculated. We wanted to generate an estimate of the time and resource required to generate a comprehensive data record, which often requires access to multiple IT systems. Data collectors therefore recorded the time taken to complete each record. Time taken to collect data for each patient varied across centres with a mean score of 27 min per patient, table 2. There was no evidence of time per record decreasing with increasing experience. There was also no correlation between number of patients and time taken to collect data. The number of data systems required to obtain the data (electronic and paper-based) ranged from 2 to 7 in the five centres that submitted resource requirements information (table 2).

Explaining heterogeneity

Median follow-up for the whole cohort was 36 months. However, to explore heterogeneity, we focused on patients with FIGO stage 2–4 disease. Median survival for stage 2–4 patients was 27 months. Kaplan-Meier analysis, stratified by treatment centre, showed marked survival differences between patients in each centre (p<0.001, log-rank test) (figure 1).

Demographics and clinical factors varied between centres. There were marked differences in indices of deprivation between the six centres, with two centres having large proportions of patients with low (1–3) deprivation scores (table 3 and figure 2A).

There were also differences between centres in age, WHO performance status, ACE27 Comorbidity Index,5 smoking status, ethnicity, Body Mass Index, FIGO stage and histological diagnosis (table 3).

However, most notable was the heterogeneity seen within treatment patterns between the six centres (figure 3). For stage 2–4 disease, the combination of surgery and chemotherapy, which can be delivered with surgery either prior to or during chemotherapy,7 is considered gold standard treatment for patients with advanced ovarian cancer. Median rates of gold standard treatment in this cohort were 64.2% but ranged from 29.4% in one centre to 75.9% in another. There were similar variations in rates of ‘no treatment’ (median 12.0%, range 2.1%–31.2%).

Cox proportional hazard modelling showed that the patient and tumour factors that impacted on survival included age (HR for death 1.04), performance statue (PS) 3 or 4 (HR 2.06 and 16.5, respectively), FIGO stage 3 or 4 (3.13 and 4.30, respectively) and albumin level (HR 0.97).

When treatment was incorporated into the model, there was no effect of receiving neoadjuvant chemotherapy and interval surgery compared with primary cytoreductive surgery. Unsurprisingly, receiving chemotherapy alone without surgery (HR 3.03) was associated with shorter survival times (HR 4.0).

However, when multivariable Cox models were generated, treatment centre remained a predictor of 1 year survival independent of patient, tumour factors and treatment choice.