Opportunities and challenges when using record linkage of routinely collected electronic health care data to evaluate outcomes of systemic anti-cancer treatment in clinical practice

Key requirements for observational studies assessing outcomes of cancer medicines

Crucial aspects to be considered when conducting studies evaluating SACT treatment outcomes in clinical practice fall into three major categories:

• Data – demographic and clinical information to allow patients, treatments, and outcomes to be described;

• Information governance – permissions to process data; and

• Interdisciplinary collaboration – the strategic and operational groups enabling and advising the work.

Data requirements mainly relate to the four areas of patient identification, treatment exposure, patient baseline characteristics, and treatment outcomes. While death records and hospitalisation data are vital to facilitate the assessment of treatment outcomes, complete regimen and indication data for the population are primary pre-requisites for any study assessing SACT treatment – allowing the identification of study eligible patients and the reliable definition of the exposure of interest, thereby decreasing the risk of bias. Summarising patient baseline characteristics such as age or comorbidities is necessary to support the interpretation of study findings, and to enable comparisons to trial populations or other published studies; for example, several disease-related factors are recognised to be important confounders and/or predictors of treatment effectiveness and/or survival, including cancer staging and the Eastern Cooperative Oncology Group Performance Status (ECOG PS).[14]

Proper information governance demands that data are processed according to relevant laws and guidance, which in Europe includes, for instance, the General Data Protection Regulation. Formal ethics committee approval may not be required if evaluation of outcomes is considered to be part of clinical audits – a component of good clinical practice – but the specifics of implemented procedures may differ depending on context. There is usually a requirement to highlight the potential benefit of a proposed study to patients and/or a health care system, as well as provisions relating to data anonymisation, researcher training, and the safety and security of work environments and associated IT infrastructure.

Considering the complexity of the topic and the amount and nature of data required to conduct studies focusing on outcomes of cancer medicines in clinical practice, interdisciplinarity of the study team and broad academic and health services support is essential; clinical expertise, especially in assessing patients for SACT and SACT prescribing, is crucial. Furthermore, governmental/health system engagement is recommended to facilitate rapid translation of findings into clinical practice.

The Cancer Medicines Outcomes Programme

Following publication of the “Cancer Action Plan” and in line with the national Digital Strategy, [15,16] in 2016, the Scottish Government made funding available for the Cancer Medicines Outcomes Programme (CMOP), a collaborative project aimed at evaluating the feasibility of using routinely collected electronic health data to determine outcomes of SACT in clinical practice in Scotland. The programme team comprises academics, quantitative and qualitative researchers, and clinicians – spanning a variety of disciplines, specialties, and areas of expertise. The main objective of the initial three-year programme was to set up a series of incremental studies to test existing linkage capabilities and assess data quality with regards to cancer treatment outcomes in local populations.

The purpose of this paper is to discuss the opportunities and challenges when applying an electronic record linkage methodology with respect to SACT by showcasing a series of CMOP studies conducted between 2017 and 2020. Although firmly situated within a Scottish setting, the majority of challenges encountered are likely not exclusive to Scotland.

Study context

Scotland has a population of approximately 5.4 million, and all residents are covered by the National Health Services (NHS) Scotland – a tax-based system offering healthcare free of charge at the point of delivery, including hospital care and prescription medicines. Cancer services are provided through three regional cancer networks: the West of Scotland Cancer Network (WoSCAN), covering four Health Boards including NHS Greater Glasgow & Clyde (NHS GGC), the largest Health Board in Scotland (Health Boards are regional organisations responsible for all health care provisioning to their respective populations); the South-East Cancer Network (SCAN), which covers four Health Boards and includes the capital city, Edinburgh; and the North Cancer Alliance (NCA), comprising six Health Boards spread out across the Scottish Highlands and Islands, mostly covering large geographical areas but relatively small populations.[17–19] Due to close collaboration with NHS GGC, initial CMOP projects, serving as pilot studies, focused on NHS GGC, which covers approximately 1.2 million people (or 20% of the overall Scottish population); subsequent projects were intended to test the scalability of the methodology and aimed to include all patients residing within WoSCAN, covering almost 50% of the total population of Scotland.[3]

Data sources

In Scotland, a range of routinely collected, national-level administrative datasets are available for research purposes, ranging from medicines dispensed in primary care to death records. In addition, clinical datasets providing radiotherapy treatment information and laboratory test results, curated by individual Health Boards, can be made available. Details of the main datasets used for CMOP studies are described in table 1.

Study design

All CMOP studies conducted to-date (as listed in table 2) have been designed as retrospective cohort studies. For practical reasons and to ensure that study cohorts were composed appropriately, patients were identified based on cancer diagnosis and/or treatment, within a defined time frame, through the Chemotherapy Electronic Prescribing and Administration System (CEPAS), a system used to record the prescription of SACT.[20] Record linkage with other relevant datasets has been enabled through the availability of a unique patient identifier in Scotland, the Community Health Index (CHI) number; CHI numbers are assigned to every resident (at birth or entry into the Health Care System), and their use is mandatory across the Health and Social Care system in Scotland.[21] While all studies had a very similar design, analytical methods were adapted to fit specific questions depending on cancer type.

Data management

Permission to use all requested datasets for CMOP studies has been approved by the Public Benefit and Privacy Panel for Health and Social Care,[22] and access has been granted through the Glasgow Safe Haven – a secure, closed environment hosted by the University of Glasgow and supported by NHS GGC.[23] Data has been pseudonymised, and no identifying information has been made available to researchers; results to be released from the Safe Haven are subject to statistical disclosure procedures. Therefore, additional ethical approval was not required.

Patient and Public Involvement

While patients or the public were not directly involved in the design, conduct, or dissemination of individual studies, the CMOP Board – responsible for the general oversight of the project – includes patient advocacy representatives.

Identifying patients

For the majority of CMOP studies, the study populations were initially identified through CEPAS, where patients are assigned both an indication for treatment (the diagnosis) and a regimen (the treatment), accompanied by a treatment intent (e.g. adjuvant or palliative); the Cancer Registry (SMR06), while being used to confirm diagnoses, did not provide sufficient information to be used for cohort identification. In the first CMOP study for example, focusing on patients with metastatic castration-resistant prostate cancer (mCRPC),[24] patients were identified based on the treatment prescribed on CEPAS (abiraterone or enzalutamide), and the indication was confirmed by the attached mCRPC diagnoses in CEPAS as well as the patients’ diagnoses of prostate cancer as recorded in SMR06. Similarly, patients for a study intended to analyse SACT outcomes in patients with metastatic melanoma were identified within CEPAS through a combination of diagnosis and treatment regimen, and eligibility for study inclusion was confirmed via ICD10 codes as recorded in SMR06 (see also table 2 for further details).

In contrast, identification of patients with some other cancers was more challenging. For instance, the identification of patients newly diagnosed with a range of gynaecological cancers (cervical, endometrial, and vaginal/vulval cancer) subject to neo-adjuvant treatment required a considerable amount of clinical input, for two main reasons: first, neo-adjuvant therapy was not easily identifiable in CEPAS for some cancer types, as this sub type was not necessarily included within the prescribing setup (where diagnoses can be selected from a drop-down menu); and second, the complex nature of these cancers impacts the reliability and consistency of the data available, since diagnoses may evolve as more information becomes available over time (e.g. from biopsies or scans). Consequently, identifying patients comprised several steps and required information from a number of datasets in order to exclude study ineligible patients (e.g. those with recurrent disease), including: details of received SACT; timing between diagnosis and initiation of SACT; and previous/concurrent treatment for the cancer of interest. The flowchart in figure 1 illustrates the complexity of the patient identification process.

• Download figure
• Open in new tab

FIGURE 1: Generic flowchart detailing identification of a neo-adjuvant chemotherapy patient cohort in gynaecological cancers

ARIA – radiotherapy treatment records dataset; CEPAS – Chemotherapy Prescribing and Administration System; NACT – neo-adjuvant chemotherapy; SACT – systemic anti-cancer treatment; SMR01 – Scottish Morbidity Records, inpatient dataset; SMR06 – Scottish Cancer Registry; XRT – radiotherapy

Defining exposure

Exposure to cancer medicines is comprehensively collected within CEPAS, including the name of the regimen, the drugs given along with their prescribed dose, and dates of administration. While defining exposure can be straightforward – as, for instance, in the case of mCRPC patients were the treatments of interest (abiraterone and enzalutamide) were both single agent oral drugs given at a fixed dose – other therapies may require further interpretation of records to enable reliable categorisation (e.g. due to complex treatment schedules involving multiple drugs with some variability of sequencing and/or dosing of medication).

As an example: patients with metastatic colorectal cancer (mCRC) may be treated with a range of different regimens depending on circumstances, comprising both chemotherapy (most prominently fluorouracil (5-FU) and/or oxaliplatin) and/or targeted treatments (e.g. aflibercept or cetuximab). If applicable and appropriate, doublet or triplet therapy (combining more than one drug) is preferred over monotherapy. The number and combination of drugs given might, however, change over time for various reasons; changes may include “stepping up” of treatment, i.e. adding a drug to the previously prescribed drug – this might occur if there are initial concerns with regards to potential side effects and a patient’s fitness; or “stepping down” of treatment, i.e. removing a drug from an originally given regimen, usually due to a patient experiencing side effects relating to one of the drugs included in the combination.

Unfortunately, there is some variability in how these cases are handled in CEPAS, and it was not always possible to determine the reason for changes in treatment. Following in-depth discussions with clinicians involved in treating mCRC patients within NHS GGC, a set of rules was developed to facilitate the simplification of exposure while acknowledging the prescribing intention; these rules were based on a) the sequence of drugs prescribed; b) the number of cycles given; and c) the timing of different regimens and/or drugs. Figure 2 illustrates two examples.

• Download figure
• Open in new tab

FIGURE 2: Examples of re-coding exposure due to stepping up or down of treatment

5-FU – 5-fluoro-uracil; FOLFOX – folinic acid, fluorouracil, oxaliplatin

Describing patient baseline characteristics

In Scotland, data on age and sex are incorporated into a person’s CHI number, and are included in all datasets. In addition, a number of geographical and socio-demographic information is available from the majority of health-related datasets; these usually include the Health Board of residence and the postcode area; an urban/rural classification;[25] and the Scottish Index of Multiple Deprivation (SIMD).[26] Information regarding co-morbidities can be obtained from hospital and outpatient clinic records, and data on medicines prescribed and dispensed in primary care may be used as a proxy for long-term conditions that do not require hospital admission; however, comorbidities might be missed if an individual has not been hospitalised and medication cannot be easily used as a proxy for the condition in question. Hospital records also comprise details with regards to procedures undertaken (see also table 1).

Regrettably, additional information relevant to cancer treatment is not always available in Scotland for record linkage studies for a number of reasons, with potential implications on the accuracy and comparability of results. Data stemming from (molecular) pathology – such as BRAF or RAS status – for instance are recorded locally by the departments conducting these tests, and stored in non-standardised formats (e.g. spreadsheets); while the presence of CHI numbers theoretically enables these datasets to be linked to other data sources, locating, accessing, and processing these files requires considerable resources, both in terms of time and technical expertise. This particularly affects studies aimed at combining patients across Health Boards, as resources and/or processes may differ considerably between regions. The Cancer Registry, in contrast, routinely captures patients from across Scotland; however, completion of all fields is not mandatory. While the registry offers the opportunity to collect additional useful information related to the disease including pathology/cell type and cancer staging based on the TNM (Tumour-Node-Metastases) system [27] or cancer-type appropriate methods (e.g. FIGO stage in patients with gynaecological cancers [28]), some of this data is frequently incomplete. In addition, records are entered at time of diagnosis and are usually not updated at recurrence – i.e. the location of metastases, for example, is only recorded if these were present at the time of diagnosis, but not if patients subsequently develop metastases. Furthermore, fields such as ECOG PS were non-mandatory within CEPAS in WoSCAN prior to July 2015, leading to incomplete data for earlier study cohorts.

Defining treatment outcomes

The most common outcome measures as reported from both clinical trials and observational studies of cancer medicines are median overall survival (OS) and landmark survival figures (e.g. % alive after five years). Death records, comprising both date and cause of death, can easily be linked to other patient records in Scotland; therefore, analyses of these outcomes among cancer patients are well supported. CMOP studies conducted thus far have provided insights into OS among mCRPC patients treated with abiraterone or enzalutamide;[24] patients with metastatic melanoma or mCRC undergoing treatment with chemotherapy, targeted treatment, and/or immunotherapies; and patients with gynaecological cancers (cervical, endometrial, and vulval cancer) subject to neo-adjuvant treatment.

Nevertheless, progression-free survival (PFS) is increasingly used to establish early efficacy in advanced cancers in RCTs with small sample sizes and/or short duration.[29,30] PFS for solid tumours is usually based on imaging data; however, in haematological malignancies such as multiple myeloma, it is determined from changes in blood plasma levels of, e.g., paraprotein.[31] Laboratory test values are, to a certain extent, available for record linkage in Scotland; however, these are collected and stored locally within individual NHS Health Boards, and the existing infrastructure does not easily support the combination of data from different areas across Scotland for various reasons (including, but not limited to, the use of different lab equipment with diverging measurement standards and variability in the coding of tests). This, in effect, currently limits our capability of conducting studies across regions where laboratory test data is required for analyses. Imaging data – and, in extension, any other data not provided in a standardised format (such as patient-reported outcomes (PROMs) or free text information) – is not routinely available for use in record linkage studies. National Language Processing and other Machine Learning techniques enabling analyses of these data are, however, in development, and have in certain areas already proven their usefulness;[32,33] with further progress being made, PFS might be added to the list of treatment outcomes readily available for analyses in the future.

Key challenges

While the description of patient populations and the calculation of median OS are supported by record linkage, potential pitfalls span the entire breadth and depth of the research process – from data acquisition to selecting variables for statistical models. The following issues are not necessarily restricted to cancer studies but deserve particular consideration in this context.

Data

Some data that would be useful for analyses is currently not routinely available for record linkage; or cannot easily be obtained. Most importantly, there are difficulties collating laboratory and molecular pathology data from across Health Boards, potentially affecting the scope of studies. In addition, across Health Boards, different versions of some data sources (particularly CEPAS) are used, making comparability of data and mapping and merging of datasets from different regions challenging.

The non-availability of specific variables might influence the ability to answer certain research questions; information regarding some potential toxicities of medicines is, for instance, not routinely captured across Scotland. Particular attention needs to be paid when attempting to use information that is contained in non-mandatory data fields of a dataset, such as socio-economic data (including, e.g., ethnicity or marital status); this data has most likely a high degree of missing-ness. Unfortunately, the same is true for data pertaining to treatment modifications and/or side effects in CEPAS. Furthermore, there is a considerable lag time (currently approximately 12-18 months) in the collection of Cancer Registry data, meaning that very recent data will likely be unavailable.

Most of these issues are, however, not unique to Scotland. Challenges with incomplete or out-of-date information in Cancer Registry data has, e.g., also been reported from the United States,[45,46] whereas details relating to SACT and laboratory test results appear to be incomplete in the Netherlands;[47] previous research conducted in the United States also advised caution when using data based on laboratory tests – not least due to diverging data recording conventions between health service providers.[48] Calls to standardise data collection, coding, and storage, and to implement nation-wide infrastructure with possible re-use of data in mind, have been made for many years.[49–51]

For the majority of datasets regularly used for research purposes, metadata is readily available; it is advisable to consult these prior to making any decisions with respect to study data. Nevertheless, it is highly recommended to collaborate with clinicians familiar with, ideally, both the treatment area of interest and the local systems used to generate and collect data. Many of the issues encountered while conducting the aforementioned studies – spanning from difficulties identifying the correct patients to uncertainties with regards to the appropriate categorisation of treatment exposure – were resolvable by discussing clinical procedures and treatment details with oncologists and other health care professionals working within these clinical settings.

Information Governance

In Scotland, the current approval process to access routinely collected electronic data is both complicated and time consuming. Permissions to use data have to be sought from the Public Benefit and Privacy Panel for Health and Social Care,[22] potentially followed by an application to use one of four existing local Safe Havens [52] depending on how data is intended to be accessed. Subsequently, extraction, anonymisation, sense checking, and uploading of data onto the selected Safe Haven to grant researchers access may require input from several parties, depending on the datasets in question. Overall, this process might take several months – this needs to be taken into account at the planning stage of a study.[53]

Interdisciplinary collaboration

CMOP has highlighted the benefit of researchers and clinicians collaborating on ambitious projects aimed at generating results useful for clinical practice by harnessing the wealth of routinely collected data. Study findings have been shared with health care professionals both within and outside WoSCAN, and clinicians are beginning to use the data to inform discussions regarding treatment decisions with their patients. In addition, initial conversations have taken place with the Scottish Government, intended to potentially impact future decisions related to data collection, recording, and access. Activities are already ongoing with regards to improving the Scottish Cancer Registry (level of completeness); upgrading CEPAS to version 6 (with the possibility of unifying data collection across Scotland and adding further information); and – potentially – implementing a Scotland-wide laboratory resource.

Next steps

The Scottish Government has provided additional funding, enabling CMOP to continue its work.[36] Building on the experiences of the first three years, the next phase is aimed at solving some of the existing challenges; upscaling select projects to cover Scotland-wide patient populations; and further developing the applied methodology. While emphasising a drive towards quality improvement in Scotland, CMOP will also investigate the utility of data stemming from clinical practice to inform medicines assessment and reimbursement decisions; and explore the integration of patient reported outcomes measures related to cancer medicines into the Health Care System.