Use of record linkage to evaluate treatment outcomes and trial eligibility in a real‐world metastatic prostate cancer population in Scotland

Abstract Purpose New treatments are introduced into standard care based on clinical trial results. However, it is not clear if these benefits are reflected in the broader population. This study analysed the clinical outcomes of patients with metastatic castration‐resistant prostate cancer, treated with abiraterone and enzalutamide, within the Scottish National Health Service. Methods Retrospective cohort study using record linkage of routinely collected healthcare data (study period: February 2012 to February 2017). Overall survival (OS) was analysed using Kaplan‐Meier methods and Cox Proportional Hazard models; a subgroup analysis comprised potentially trial‐eligible patients. Results Overall, 271 patients were included and 73.8% died during the study period. Median OS was poorer than in the pivotal trials, regardless of medication and indication: 10.8 months (95% confidence interval [CI] 8.6‐15.1) and 20.9 months (95% CI 14.9‐29.0) for abiraterone, and 12.6 months (95% CI 10.5‐18.2) and 16.0 months (95% CI 9.8—not reached) for enzalutamide, post and pre chemotherapy, respectively. Only 46% of patients were potentially “trial eligible” and in this subgroup OS improved. Factors influencing survival included baseline performance status, and baseline prostate‐specific antigen, alkaline phosphatase, and albumin levels. Conclusions Poorer prognostic features of non‐trial eligible patients impact real‐world outcomes of cancer medicines. Electronic record linkage of routinely collected healthcare data offers an opportunity to report outcomes on cancer medicines at scale and describe population demographics. The availability of such observational data to supplement clinical trial results enables patients and clinicians to make more informed treatment decisions, and policymakers to contextualise trial findings.

studies frequently describe considerably shorter OS than that reported from clinical trials, including studies analysing the use of abiraterone and enzalutamide, [13][14][15][16] although some studies have found OS rates similar to the pivotal trials. 13,17,18 Consequently, a recent study perspective proposed that OS obtained from clinical trials should only be considered a surrogate endpoint for OS in the real world. 19 As access to cancer medicines is a prominent matter of public interest and healthcare policy, policymakers-particularly in publicly funded healthcare systems such as the National Health Service (NHS) in the UK-would benefit from an improved understanding of whether these treatments deliver the same outcomes in clinical practice as demonstrated in Health Technology Assessments (HTAs), which are almost exclusively based on clinical trials data. There is a growing demand to develop more efficient methods to better identify the benefits and risks of new treatments, and thus to assure the most effective use of available resources, particularly with the increasing pressures of early licensing. Observational data have the potential to supplement the existing appraisal processes with medicines effectiveness data obtained from real-world populations.
In 2016, funding was made available in Scotland as part of the "Beating Cancer: Ambition and Action" Cancer Plan, 20 to deliver the Cancer Medicines Outcomes Programme (CMOP). The CMOP vision is to maximise the use of the existing and evolving local and national electronic datasets to better understand treatment outcomes of cancer medicines in the Scottish population. The methodology is designed to grow scalable and sustainable expertise in cancer medicines intelligence to drive continuous improvement in the safe and effective use of these medicines, and thereby to contribute to the international evidence base. During its initial funding phase, CMOP has sought to deliver an incremental programme of studies to test the availability and record linkage capacity of routinely collected clinical datasets within a region of Scotland.
Prostate cancer was chosen as an initial CMOP case study, as it is the most common male cancer and was the second most common cause of death in men in the United Kingdom in 2017. 21 The West of Scotland Clinical Management Guidelines for mCRPC patients include both abiraterone and enzalutamide, and selection of treatment is made jointly by clinicians and patients taking into consideration individual patient factors and expected side effect profile. 22 Prostate-specific antigen (PSA) is an important molecular marker in the diagnosis and management of prostate cancer and is used alongside the Gleason score and clinical staging to evaluate the prognosis of newly diagnosed patients. 23

KEY POINTS
• Median overall survival in patients with metastatic castration-resistant prostate cancer treated with abiraterone and enzalutamide in clinical practice is less than was observed in the pivotal trials.
• Treatment outcomes were impacted by poorer prognostic factors exhibited by non-trial eligible patients.
• The majority of patients being treated with abiraterone and enzalutamide in clinical practice would not have been eligible for inclusion in the pivotal trials that led to the approval of these medicines.
• Linked, electronic data sources can be used at a population level to describe patient demographics, estimate trial eligibility, and analyse outcomes of cancer medicines in clinical practice.
• This information, alongside clinical trial findings, may enable a more informed discussion of the likely outcomes of treatment, particularly with patients who do not fit clinical trial eligibility criteria; and facilitates a better understanding of the outcomes associated with these medicines in the "real world." The purpose of this study was to demonstrate the feasibility of using electronic record linkage (ERL) to measure real-world outcomes of cancer medicines by determining treatment outcomes of abiraterone and enzalutamide in clinical practice in Scotland. Specific objectives were to: • Calculate OS in patients with mCRPC treated with abiraterone and enzalutamide, and descriptively compare findings to results obtained in the respective clinical trials; • Identify potentially trial-eligible patients, and explore outcomes in this subgroup; and • Analyse factors influencing survival.  24 Dates for cohort inclusion were chosen based on the timing of medicine approval in Scotland and data availability, allowing for sufficient follow-up time. Patients were excluded if they participated in a clinical trial (except where the medicine was used within its product label), or if they received both abiraterone and enzalutamide during the recruitment period. Patients were followed up until death or the end of the study period (February 28, 2017), whichever occurred first.

| Data sources
A range of data, comprising patient demographics, diagnostic details, and information regarding previous, current, and subsequent treatments, were gathered from a number of separate databases used in routine care (Table 1). Records were linked via Community Health Index (CHI) numbers, a unique patient identifier used throughout the NHS to identify individual patients. 25 Trial eligibility criteria were identified through published proto-

| Treatment outcome measures
The primary outcome measure was OS. The duration of treatment was a key secondary endpoint. February 28, 2017, served as the censor date for those patients event free at study end.

| Statistical analysis
Total median follow-up was calculated in two ways. First, for descriptive purposes, using median total observation time (time from treatment initiation until death or censoring), and second, using median Kaplan-Meier estimate of potential follow-up to allow comparability between groups or studies with differing death rates. 29  Charlson comorbidity index (CCI) score and/or number of different medicines used concomitantly at baseline-were created using variables with P < 0.2 from the univariable analyses; the proportionality assumptions were tested using Schoenfeld residuals. In addition to the main analysis, subgroup analyses were undertaken for patients who would potentially have been eligible/non-eligible for inclusion in the respective trials.
All analyses were performed using the R software, version 3.3.3.

| Patient baseline characteristics
A total of 288 patients initiated treatment with abiraterone or enzalutamide between February 2012 and December 2015. Seventeen patients were subsequently excluded due to having received both medicines during this period; hence, 271 patients were included in the analysis ( Figure 1).
Baseline characteristics at treatment initiation for all patients, stratified by medicine and indication and also including those from the respective clinical trials, are summarised in Table 2.

| Treatment outcomes
At the end of the study period, 71 patients (26%) were still alive: 14.0 months (95% CI 11.5-18.2) amongst abiraterone and enzalutamide patients, respectively. Observed median OS were less than those reported from the pivotal clinical trials across all patient groups, as detailed in Tables 3 and 4.

| Factors influencing overall survival
Exploring the underlying prognostic characteristics which might explain differences in OS between the real-world patients and those included in the pivotal trials showed that a variety of individual factors had an impact on survival in univariable analyses in the post-and/or pre-chemotherapy group, including baseline performance status; Gleason score; and baseline PSA, alkaline phosphatase, albumin, and haemoglobin levels (see Table 5 for details).
In the fully adjusted multivariable analyses, factors that remained independently associated with survival were baseline performance status, and baseline PSA and alkaline phosphatase levels in the postchemotherapy group; and baseline performance status and baseline PSA and albumin levels in the pre-chemotherapy group. For details, see Table 5.
In addition, a complete case analysis has been conducted to assess how sensitive the study results were to the missing data on performance status. Briefly, findings were generally in line with those obtained through the adjusted multivariable models containing missing data within the categorical variables; although some effect sizes differed slightly, the direction of effects was the same. Details are presented in Table S4.

| DISCUSSION
To our knowledge, this is the first study to demonstrate the use of ERL of routinely collected data to obtain a comprehensive, population-level assessment of treatment outcomes of cancer medicines. Furthermore, by providing results based on a subgroup analysis of potentially trial eligible patients, this study offers vital information that helps contextualise real-world results in comparison to clinical trials.
In line with other real-world studies, median OS in patients treated with either abiraterone or enzalutamide in both the pre and post-chemotherapy settings were less than in the respective pivotal trials. [13][14][15][16] However, there were important differences between patients observed in clinical practice and the trial populations. In particular, the patients included in this study were older, and had poorer performance status; patients with performance status >2 were included within our study, but were excluded from all pivotal trials. [4][5][6][7] T A B L E 3 Outcomes in the post-and pre-chemotherapy abiraterone populations In addition, patients had higher alkaline phosphatase levels at baseline, and fewer patients had a Gleason score ≤7. Age and alkaline phosphatase were found to be significant prognostic factors in the abiraterone pre chemotherapy RCT, 9 whilst performance status was significant in the enzalutamide post-chemotherapy trial 6 ; conversely, the postchemotherapy trials did not include baseline albumin, 5 [11][12][13]30,31 The underlying impact of differences in patient characteristics as an explanation for the differences observed between RCTs and clinical practice has been confirmed by conducting a subgroup analysis of those patients who would potentially have been trial eligible. In this subgroup, median OS was higher than in the overall study population, Outcomes in the post-and pre-chemotherapy enzalutamide populations

| Strengths and limitations
This study has several strengths: apart from its inclusiveness in terms of study participants, its scope with regards to the medicines studied is unprecedented (by analysing OS for both abiraterone and enzalutamide, pre and post docetaxel chemotherapy). The comprehensiveness and richness of the data available facilitated the identification of potentially trial-eligible patients, which in turn enabled us to conduct the subgroup analysis; this provided a unique opportunity to better compare trial results to those observed in clinical practice.
Nevertheless, there are also limitations to consider. First of all, observational studies such as this are, by nature, non-randomised and sized according to the underlying population. Consequently, there were small numbers of patients in each group, and CIs were wide and overlapped; hence, comparisons between our findings and RCT results should be interpreted with caution. Furthermore, findings are subject to confounding; however, attempts were made to adjust for confounding by using information available from electronic health records. Second, this study did not set out to compare abiraterone and enzalutamide treatments and cannot estimate the relative added benefits of these treatments since there was no control arm. In addition, it was not possible to reproduce key efficacy outcomes as reported from clinical trials; as unstructured, free text, or imaging data were not available for analyses, determining progression-free survival, for example, was not feasible. Third, required data were not always available or were missing; for instance, 36% of patients had no performance status recorded as this only became mandatory on the chemotherapy electronic prescribing system in 2015, potentially impacting the accuracy of results. Finally, the identification of patients eligible for the pivotal clinical trials needs to be interpreted with caution; whilst certain eligibility criteria (eg, demographics and laboratory test results) were easily and unambiguously identifiable within electronic patient records, other criteria were associated with a degree of uncertainty. Identification of prior surgery, for example, was made based on assumptions regarding included procedures.

| CONCLUSION
ERL of routinely collected healthcare data offers an opportunity to report outcomes on cancer medicines at scale and describe patient demographics whilst potentially identifying real-world patients ineligible for the pivotal studies. Such information may be valuable to patients and clinicians to contextualise trial findings, and thus to make more informed shared treatment decisions.

ETHICS STATEMENT
Ethics approval was not required. However, the use of the data use of the data was approved by the Local Privacy Advisory Committee.