Breast cancer learning health system: Patient information from a data and analytics platform characterizes care provided

Abstract

Purpose: In a learning health system (LHS), data gathered from clinical practice informs care and scientific investigation. To demonstrate how a novel data and analytics platform can enable an LHS at a regional cancer center by characterizing the care provided to breast cancer patients.

Methods: Socioeconomic information, tumor characteristics, treatments and out-comes were extracted from the platform and combined to characterize the patientpopulation and their clinical course. Oncologists were asked to identify examples where clinical practice guidelines (CPGs) or policy changes had varying impacts on practice. These constructs were evaluated by extracting the corresponding data.

Results: Breast cancer patients (5768) seen at the Juravinski Cancer Centre between January 2014 and June 2022 were included. The average age was 62.5 years. The commonest histology was invasive ductal carcinoma (74.6%); 77% were estrogen receptor-positive and 15.5% were HER2 Neu positive. Breast-conserving surgery(BCS) occurred in 56%. For the 4294 patients who received systemic therapy, the initial indications were adjuvant (3096), neoadjuvant (828) and palliative (370). Metastases occurred in 531 patients and 495 patients died. Lowest-income patients had a higher mortality rate. For the adoption of CPGs, the uptake for adjuvant bisphosphonate was very low, 8% as predicted, compared to 64% for pertuzumab, a HER2 targeted agent and 40.2% for CD4/6 inhibitors in metastases. During COVID-19, the provincial cancer agency issued a policy to shorten the duration of radiation afterBCS. There was a significant reduction in the average number of fractions to the breast by five fractions.

Conclusion: Our platform characterized care and the clinical course of breast cancer patients. Practice changes in response to regulatory developments and policy changes were measured. Establishing a data platform is important for an LHS. The next step is for the data to feedback and change practice, that is, close the loop.

Introduction

Data generated through day-to-day clinical practice holds valuable clinical insight into the impact of treatment and care delivery on patient outcomes. This was recognized by the Institute of Medicine (IOM) through the learning health system (LHS); a continuous cycle or feedback loop in which scientific evidence informs clinical practice, while data gathered from clinical practice and administrative sources informs care and scientific investigation.¹ In practice, the latter part of the loop is often missing because patient data are frequently siloed and/or unstructured, making it unsuitable for analysis with conventional statistical techniques. While manual retrospective chart reviews can allow such data to be used for research and quality improvement, they are time-consuming and are limited by the scale of data that can be captured. The IOM recognized that having powerful, nimble and secure digital infrastructure is key to supporting an LHS.²

By leveraging routinely collected electronic data on patient care, an LHS can inform patient/oncologist decision-making and institutional policy. Examples of questions that are important to a breast cancer LHS are whether clinical practice is consistent with guidelines, how practice changes in response to emerging evidence, how social determinants of health influence patient outcomes, and how patients are managed in scenarios where randomized trial data are not available.

The electronic health record (EHR) contains an abundance of patient information.³ However, harnessing clinical data from the EHR and making it available for research and quality assurance is a daunting challenge.⁴

The aim of our research is to create a breast cancer LHS at the Juravinski Cancer Center (JCC), a regional cancer center affiliated with McMaster University in Hamilton, Ontario. The JCC is the hub of cancer treatment services in south central Ontario, serving a catchment area of approximately 2.5 million people. Along with three affiliated community oncology clinics, it delivers all the systemic cancer therapy in the region and most of the radiation therapy.

In a previous pilot project, we demonstrated that the clinical course of 50 patients with locally advanced breast cancer could be characterized using a form of artificial intelligence called Natural Language Processing (NLP) applied to the clinical notes stored in the hospital’s EHR.⁵ Building upon the work in the pilot study, we scaled up to a breast cancer data and analytics platform that automatically extracts and integrates data from our hospital’s information systems and clinical documentation on a nightly basis.⁶ DARWEN, a commercially available medical NLP engine, was used to extract structured data from unstructured clinical documentation.⁶ In the current report, we demonstrate how this platform can characterize clinical care by building a comprehensive view of the patient journey and we describe opportunities provided by the platform to conduct research and quality improvement that define an LHS.

Introduction

This study was approved by the Hamilton Integrated Research Ethics Board (Project Number 15338-C).

2.1 Data source

Our data platform currently includes 141 data elements prioritized by our clinicians.⁶ The data elements span many topic areas including demographics, patient encounters, pathology, medical imaging, radiation treatment, systemic therapy information, quality of life, recurrence, death and social determinants of health.

2.2 Patients

The cohort for the current report was assembled from 7019 new breast cancer patients seen at the JCC between 1 January 2014 and 3 June 2022. To be included in the cohort, histology had to be available. A total of 5986 (85.3%) patients had histology reported; invasive ductal or invasive lobular (4992), invasive mixed (169), ductal carcinoma in situ (DCIS) or lobular carcinoma in situ (LCIS) (677) and other (148). Of the 1033 patients with missing histology, almost half were missing because they had been referred to the JCC only for genetics assessment, pain control or social work consultation where such information is not required for referral. Patients with invasive ductal, invasive lobular or invasive mixed histology who did not have any biomarker information (hormone receptors, HER 2 Neu) were then excluded (118). Thus, the final study cohort consisted of 5868 patients.

2.3 Interventions

Systemic therapy included chemotherapy, endocrine therapy and molecular targeted agents. The systemic therapy regimens were coded by the JCC pharmacy as adjuvant, neoadjuvant or palliative (Figure 1).

2.4 Outcomes

Mortality, including death from any cause, is reported along with the occurrence of a patient’s first metastasis and site of metastases. The Edmonton Symptom Assessment Scale (ESAS), which is patient-reported, is routinely collected at each clinic visit.⁷ The effect of adjuvant chemotherapy on patients is reported for several questions in ESAS before treatment, at mid-course and after completion of adjuvant chemotherapy.

2.5 Social determinants of health

The importance of some social factors, for example, poverty and marginalization, on health has been recognized.⁸ Accordingly, we are building measurements of social determinants into the LHS. To explore the potential impact of the social determinants of health on our cohort we employed the Ontario Marginalization Index.^9, 10 We considered four dimensions of marginalization (material deprivation, dependency, ethnic concentration and residential instability) and analysed the distribution of patients treated at the JCC based on quintiles of the Ontario population (quintile one being the least marginalized in Ontario and quintile five being the most marginalized). A more detailed description of this index and the dimensions of marginalization are provided in the supplement (Table S1).

2.6 Uptake of clinical practice guidelines and impact of policy changes

To evaluate the ability of our data platform to identify opportunities for research and quality improvement related to practice change, JCC medical oncologists specializing in breast cancer were asked to identify, based on their experience, two examples from the previous 10 years where a new clinical practice guideline (CPG) was published, and clinical practice changed immediately to be consistent with the CPG. These examples would be characterized as “game changers”. Example 1: In 2013, pertuzumab received regulatory approval for patients with metastatic HER 2 positive breast cancer.¹¹ Example 2: In 2016, CD4/6 inhibitors received regulatory approval for the treatment of estrogen receptor-positive metastatic breast cancer and their use was recommended through CPGs.¹² The medical oncologists were also asked to provide an example from the previous 10 years where a CPG was published, but oncologists were reluctant to follow it, that is, it was less clear-cut, and the response was mixed. Example 3: In 2017, guidelines recommended that bisphosphonates which block bone resorption be used in the adjuvant setting to prevent breast cancer recurrence and thus improve survival in post-menopausal women.¹³

Since radiation therapy practice in Ontario is tightly coupled to policy guidance from the provincial cancer authority, Cancer Care Ontario (CCO) radiation oncologists were asked to identify a recent provincial policy change that may have had a significant impact on practice. Example 4: On 25 April 2020, in response to the COVID-19 pandemic, CCO recommended decreasing the number of fractions for postoperative breast irradiation.¹⁴

2.7 Analysis

Descriptive statistics were used to analyze the patient population and the care provided.

Practice change in radiation therapy was evaluated using an interrupted time series design, a robust quasi-experimental method for evaluating the impact of quality improvement initiatives and large-scale health interventions when randomization is not possible.¹⁵ We employed an Autoregressive Integrated Moving Average (ARIMA) model for this analysis to account for seasonality and autocorrelation.¹⁶ Monthly values for mean fractions per patient were calculated by attributing all fractions provided to patients to the month in which their radiation therapy was initiated.

Analyses were carried out in Python (version 3.9.6) using the following packages: matplotlib (3.5.3), pandas (1.4.2), seaborn (0.11.2) and stats (0.13.2) models.

Results

3.4 Social determinants of health

The baseline socioeconomic characteristics of the cohort are described in Figure 2. There is a relatively even distribution of patients with respect to material deprivation, which is closely related to poverty, with slightly fewer patients from the most deprived quintile. Conversely, there was a higher proportion of patients from the most deprived quintiles who have died.

3.5 Uptake of clinical practice guidelines and impact of policy changes

The adoption of pertuzumab was identified as an example of rapid uptake of a guideline by the JCC medical oncologists (example 1). A total of 72 (64.3%) of 112 patients with metastatic HER2-positive disease received pertuzumab as a part of their antineoplastic systemic regimen. The second example of rapid uptake identified a priori was CD4/6 inhibitors. Of 303 patients over 50 years of age with ER-positive metastatic breast cancer, 122 (40.2%) took a CD4/6 inhibitor. Their adoption into practice is shown in Figure 3. To further examine the uptake, a locally estimated scatterplot smoothing (LOESS) plot was performed on the scatter plot data (Figure 4). The medical oncologists postulated that there would be poor uptake of the use of adjuvant bisphosphonates (example 3). Of 2264 patients over 50 years of age with breast cancer who received systemic adjuvant therapy after 1 January 2017, only 180 (8.0%) received adjuvant bisphosphonate.

In response to COVID-19, CCO recommended decreasing the number of fractions for postoperative breast irradiation.¹³ As a result of the CCO recommendation (example 4), there was a marked reduction in the average number of fractions, from 17 prior to the pandemic to 12 from May 2020 onward (Figure 5).

Discussion

The overall goal of our research was to describe the care provided to breast cancer patients at our regional cancer center using information from a novel data and analytics platform.⁶ We have shown that individual patient data on patient factors, for example, comorbid medical conditions, social determinants, tumor characteristics (e.g., hormone receptors), HER2 Neu expression, diagnosis, treatments and follow-up outcomes can be extracted from this platform and assembled to characterize the care patients receive throughout their journey. Having such information is essential to the creation of an LHS. It can be used both to identify areas for quality improvement and research, as well as to monitor the impact of change efforts related to new clinical guidelines and policies.

Increased use of granular neighborhood-level socioeconomic data has been recently highlighted as an important opportunity for cancer research.^17–19 Our data reveals that while our cancer center serves a relatively even distribution of the population with respect to material deprivation, survivorship for breast cancer appears to be much better for the rich than for the poor. The same is true for housing instability; patients from stable neighborhoods appear to be less likely to die than patients from neighborhoods with high instability. These are interesting observations coming from a system with universal healthcare access. We did not analyze whether the poorest patients or those with the most unstable housing received different treatments than the other patient groups. This is for future work.

An important function of a data platform such as ours is to measure uptake of CPGs in routine clinical practice. The validity of our data platform was established by assessing the degree with which the data on treatments received agreed with theoretical constructs. A priori, medical oncologists from the center’s breast cancer disease site team identified two significant practice changes that were reflected in guidelines and would be manifest in our analyses. They also identified a scenario where they predicted uptake would be low. The uptake of pertuzumab, a recombinant humanized monoclonal antibody that binds to a separate antigenic region of the HER2 extracellular domain than trastuzumab, was relatively high as predicted a priori by the medical oncologists.¹¹ Approximately one-third of patients with HER2-positive metastatic disease did not receive pertuzumab. Possible explanations are the presence of cardiac disease, decreased cardiac ejection fraction because of prior exposure to adjuvant trastuzumab, or patients who received pertuzumab at one of the community oncology sites and this data was not captured in the data platform. Our results show the uptake of CD4/6 inhibitors was high once approved by the government regulatory agency.¹² It is not surprising that the use of palbociclib was much higher than ribociclib, likely because it was introduced first.²⁰ However, we anticipate that soon our data will reflect that ribociclib use will overtake that of palbociclib because of the recently reported improvement in survival with ribociclib.²¹ The uptake of adjuvant bisphosphonate was low as predicted. The published CPG was controversial.¹³ Reasons for this were questions about the validity of the data, lack of clarity on their mechanism of actions, small magnitude of benefit, side effects and uncertainty about the logistics of therapy (time to start and dosing intervals).^13, 22

Post hoc consideration of the patient data provides an opportunity for us to speculate on the congruence of the data with preconceived notions. For example, the most common adjuvant or neoadjuvant chemotherapy regimen used was DDACT in keeping with CPGs.²³ For metastatic disease, less toxic chemotherapy was chosen, for example, oral capecitabine, weekly doxorubicin or weekly paclitaxel. We speculate that this real-world evidence supports the construct that a goal of treatment for metastases is to control the cancer and maintain quality of life for as long as possible.²⁴

Another important function of a data platform is to measure changes in practice resulting from policy changes. The COVID-19 pandemic resulted in an abrupt change in practice at the JCC and worldwide.^25, 26 To reduce patient visits to the JCC there was a substantial change in radiation therapy to shorter radiation treatments. Although this was a pragmatic decision, it was based on guidelines from CCO¹⁴ and emerging data on ultra-hypofractionation.²⁷ Our platform clearly captured the change over time, a valuable ability for a rapid LHS that needs to monitor the impact of its change initiatives.

The routine use of patient-reported outcome measures for clinical decision-making in oncology clinical practice is being advocated.²⁸ In 2007, CCO encouraged all their regional cancer centers to routinely collect information on symptoms at each clinic visit with the ESAS (patient self-reported).^29, 30 The pattern of ESAS scores in patients receiving adjuvant chemotherapy was consistent with expectations. Patients had high anxiety levels on their first visit to the clinic. We speculate that they are waiting for information on their treatment and prognosis. Once there is a plan, anxiety drops considerably. Pain increased over time and after completion of treatment. This is likely related to the neuropathic pain from paclitaxel that can last for months after completion of chemotherapy. These two examples are encouraging and support that in our LHS, patient-reported outcomes can be connected to other clinical data.

Administrative data sources such as SEER have been used to generate evidence from real-world practice. However, these sources do not tend to enable cancer centers to leverage their own data to monitor and improve care in their regions. The potential of the EHR to be a valuable resource for regional research and quality improvement in cancer has been touted for a long time.^3, 31 Progress has been slow, due in large part to the siloed and unstructured nature of clinical documentation. The Oncoshare project has linked multiple local and national breast cancer data sources to enable population-based³² and NLP research topics.³³ The potential for integrated data to enable LHS activities was recently described by Hernandez-Boussard, who combined data from Stanford University EHR, the California Cancer Registry and the Veteran’s Health Administration to characterize care provided to their patients.³⁴ Recently, Morin and colleagues described the creation of an information technology infrastructure using data from cancer patients at the University of California San Francisco³⁵ to generate prognostic models in breast cancer using NLP. In our study, we extracted data from a platform to describe the clinical course of and care provided to breast cancer patients at a regional cancer center from diagnosis to last follow-up or death. To the best of our knowledge, this is the first time this has been reported.

Conclusion