Data Driven Approach to Burden of Treatment Measurement: A Study of Patients with Breast Cancer

Abstract

Chronic disease affects patient quality of life through symptoms of the disease and the work of receiving treatment. While the effects of illness are well investigated, the burden of treatment is not commonly studied or monitored. We developed a method to quantify one dimension of the burden of treatment based on patient encounters with the healthcare system. We applied this method to a population of stage I-III breast cancer patients. As hypothesized and observed, stage IIIpatients had more appointments, spent more time in clinic, and spent more time admitted to the hospital in the first 18 months after diagnosis compared to stage I and II patients. Future work will evaluate the reproducibility and generalizability of this method for quantifying burden of treatment across other clinical settings and chronic diseases. This approach could enable identification of high-risk groups that could benefit from interventions to decrease patient work and improve outcomes.

Introduction

Chronic illness is detrimental to a patient’s quality of life both because of the symptoms of the illness itself as well as the burden of treatment needed to combat the illness¹. The complexity of medical care today makes it difficult for healthcare providers to monitor a patient’s capacity to receive care even though treatment overburden can impact disease outcomes². A high burden of treatment can cause lower compliance in patients with chronic diseases, thus increasing other complications³’⁴. Patients with high burden of treatment tend to have lower satisfaction scores⁵. Excessive treatment can also lead to wasted resources for the medical center⁶, and contribute to a patient’s financial toxicity⁷. Physicians who practice minimally invasive care assess burden and tailor treatment plans that give a patient the maximum likelihood of recovery while taking into consideration the patient’s limitations⁸. To improve the effectiveness of this paradigm, providers and healthcare systems need reliable ways to identify overburdened patients and patient populations.

Cancer patients often undergo intense, multi-modal treatments resulting in diminished quality of life⁹’¹⁰ Figures 1 shows a typical two-week schedule for a breast cancer patient undergoing adjuvant chemotherapy following surgery. This schedule includes 11 appointments over 7 unique days and a total clinic time of 9.5 hours. Additional appointments needed to address complications and comorbid conditions compound the treatment burden for cancer patients. A nationwide study showed that among patients undergoing chemotherapy and radiotherapy, 28% had to schedule appointments to treat side effects, 77% had to arrange for caregivers to accompany them to their appointments, and 43% had some impact to their professional lives¹¹.

Figure 1.

Example of a breast cancer patient’s schedule over a two-week period. This patient had 11 appointments over 7 unique days and a total clinic time of 9.5 hours between July 11 and August 1. MO = Medical Oncologist, NP = Nurse Practitioner.

Burden of treatment measures focus on the impact of the acts of receiving care. Previous research on the topic of treatment burden have focused on qualitative methods to describe factors contributing to patient work¹³. By surveying patients, researchers identified a taxonomy of factors that contribute to the burden of treatment including healthcare tasks and situational factors that exacerbate a patient’s work^14,15. Tran et. al. also developed and validated the Treatment Burden Questionnaire (TBQ), a survey instrument designed to measure paitent burden^5,14. Disease specific questionnaires have also been developed to assess the burden of treatment in specific chronic conditions such as chronic heart failure¹⁶ and end stage renal disease requiring dialysis¹⁷. While these qualitative methods are effective in defining treatment burden, they do not enable population studies of burden or automated monitoring to identify overworked patients who may need intervention.

Research using electronic health record data is most effective when that data is structured and complete¹⁹. Figures 2 displays several dimensions of treatment burden identified by Tran et. al.¹³ and Eton et. al¹⁴ evaluated by approximate completeness and structure of data in the electronic health record. Treatment burden elements in the top-right quadrant of Figures 2 such as financial costs and clinic visits are both highly structured and highly complete by virtue of their relation to billing. These top-right elements are more reliable measures than elements in the bottom-left that are unstructured and less available. For example, a patient’s exercise program is less likely to be captured in an EHR but may be recorded in other systems managed by the patient outside the EHR. While healthcare institutions may differ in how these factors of treatment burden are recorded, many of these data elements are available across different implementations of clinical information systems.

Figure 2.

Relative completeness and structure of data elements related to factors of treatment burden that may be amenable to automated extraction and quantification from the electronic health record.

This study proposes a method to quantify one of the more reliable contributors to treatment burden: clinical encounters. While appointments only contribute to part of a patient’s overall treatment burden, time and effort spent coordinating, traveling for, and waiting for care were among the most commonly mentioned contributors to burden in patient surveys¹³, and the factors most highly correlated with the overall TBQ score⁵. Hospital admissions are also of particular interest to patients and hospital administration. Cancer related readmissions are often unexpected and the most frequent among patients with private insurance¹⁸. Although inpatient time was typically not mentioned in previous burden of treatment literature, admissions are disruptive to patients and their caregivers. To our knowledge, no previous studies have used data from electronic health records to assess treatment burden in patient populations. By using appointment and admission data, our evaluation of treatment burden is reliable, reproducible, and scalable given accurate electronic records.

To evaluate our method, we chose to investigate patients with stage I-III breast cancer. We hypothesized that patients with stage III cancer would have a higher treatment burden than patients with stage I or stage II cancer given the more aggressive therapy they receive for higher risk disease. Additionally, we anticipated that burden of treatment would be greatest in the first few months after the date of cancer diagnosis, and that this burden would decrease over time.

Methods

To study the use of these treatment burden metrics, we applied them to a population of breast cancer patients at Vanderbilt University Medical Center (VUMC). The goal of this analysis was to investigate whether there was differentiation in number of appointments, total time spent in clinic including wait time, and total inpatient length of stay between breast cancer patients with stage I, stage II and stage III disease. We also identified outliers who had abnormally high treatment burden for their stage.

Study population

The patient cohort for this study was chosen from the VUMC Cancer Registry since those were the patients who were diagnosed and receive all or part of their first course of treatment at our institution. We collected 18 months of encounter information from all stage I-III breast cancer patients who were diagnosed over a 17-year period between January 1, 1998 and June 1, 2014. In order to facilitate comparison between sub-populations of patients that received the majority of their care at our institution, we included only patients with at least three appointments from both a Vanderbilt medical oncologist and a Vanderbilt surgical oncologist. We determined which appointments were with a medical or surgical oncologist by mapping their national provider identification number with their specialty in the national patient identifier (NPI) data dissemination file²⁰.

Among the 8161 patients with breast cancer in the VUMC Cancer Registry, 5661 had a date of diagnosis between January 1, 1998 and June 1, 2014. Among these, 4152 patients had stage I-III disease at diagnosis. After collecting 18 months of appointments after the date of diagnosis for these patients, we found that 904 had more than three appointments with a medical oncologist and a surgical oncologist at VUMC (Figures 3).

Figure 3.

Cohort selection

Data Considerations

To get a holistic view of each patient’s treatment burden, we included all patient appointments, including appointments in non-cancer related departments, and all hospital admissions. We extracted outpatient appointment data from the Epic scheduling system that had been in use at VUMC since 1997. Hospital admission and discharge data was extracted from the Medipac admission, discharge, transfer (ADT) system with data going back to 1984. To approximate the total time spent in clinic including waiting time, if a patient had more than one appointment in a day, we calculated the time from the beginning of the first appointment to the end of the last appointment of the day. We used the schedule appointment duration as an estimate of the time the patient spent in clinic understanding that true appointment durations could vary.

Analysis

We compared distributions of the total time spent in clinic, the number of appointments, and the number of admissions over 18 months by stage. We used an ANOVA test to see if there was a significant difference between stages I-III for each of the means of these three metrics. We also compared the estimated time spent in clinics by month over 18 months by stage. The Vanderbilt Institutional Review Board approved this study and granted a waiver of consent since we analyzed a large population of patients in aggregate.

Results

Table 1 summarizes the clinical encounter burden by stage. Among the 904 patients in the final cohort, 419 had stage I, 337 had stage II, and 148 had stage III disease. Across all stages, the average patient in our cohort had 67 appointments on 44 unique appointment days, spent 3.6 hours in clinic per month, and was admitted for 1.9 days over the 18 months after diagnosis. On average, Stage III patients had the greatest number of appointments (92), the greatest amount of time spent in clinic per month (5.4 hours), and the greatest number of hospitalized days (3). Stage II patients had the second greatest totals in each of these parameters and stage I patients had the least. The ANOVA tests for all of the metrics were significant with p-values less than. 001.

Table 1.

Summary of clinical encounter burden for breast cancer patients by stage with ANOVA p-values comparing the difference between stage I, stage II, and stage III.

Breast Cancer Stage		Stage I	Stage II	Stage III	Stage I-III	ANOVA p- value
Number of patients		419	337	148	904	N/A
In first 18 months after diagnosis Mean (Range)	Number of appointments	53(8-164)	74(9-254)	92(15-217)	67(8-254)	<. 001
	Unique appointment days	37(7-126)	47(4-145)	57(6-124)	44(4-145)	<. 001
	Hours of appointment time	26(3-128)	48(3-195)	62(8-152)	40(3-195)	<. 001
	Hours spent waiting between appointments	17(0-97)	28(.5-95)	36(3-92)	24(0-97)	<. 001
	Hours spent in clinic	43(3-163)	76(6-247)	98(12-233)	64(3-247)	<. 001
	Hours spent in clinic per month	2.4(.17-9)	4.2(.33-14)	5.4(.67-13)	3.6(.17-14)	<. 001
	Number of unique admissions	.51(0-10)	.87(0-8)	1.1(0-7)	.74(0-10)	<. 001
	Total inpatient length of stay (days)	1.3(0-41)	2.3(0-33)	3.0(0-29)	1.9(0-41)	<. 001

The boxplot in Figures 4A further shows that the median and interquartile ranges of the number of appointments is greatest for patients with stage III disease followed by stage II and stage I. There were 5 outliers for stage I patients with an unusually high number of appointments while there were 2 outliers for stage II patients. Figures 4B shows the distribution of total length of hospitalized days in the first 18 months after diagnosis. The median patient for stage I and stage II was admitted for one day or less while the median stage III patient was admitted for about two days. Among stage I patients, 67% had no admissions, while only 49% of stage II and 41% of stage III patients had no admissions. With so many patients clustered near zero, there were many outliers for all stages.

Figures 4A and 4B.

Distribution of total number of appointments over 18 months by breast cancer stage (ANOVA p-value <.001), and total length of stay over 18 months by breast cancer stage (ANOVA p-value <.001). The dark line for each boxplot represents the median and the colored box represents the interquartile range (IQR) (25^th to 75^th percentile). The “whiskers” extend to 1.5 times the IQR or to the minimum or maximum value, whichever is closer. Any data points outside the whiskers are outliers and are represented individually as circles.

Figures 5A shows that the median time spent in clinic was greatest for stage III patients followed by stage II and stage I. There are 26 outliers for stage I patients in time spent in clinic and 2 for stage II patients. Over the course of 18 months after diagnosis, the average total time spent in clinic per patient decreased (Figures 5B). In the first month of treatment, stage I patients spent on average 8 hours in clinic while stage II and III patients spent 10 and 12 hours respectively. All three stages saw a decrease in time spent in clinic in the second month but then had an increase in the third and fourth months. In each month after diagnosis, the average time spent in clinic was greater for stage III than stage II or stage I breast cancer patients, although this difference was not statistically significant for any given month.

Figures 5A and 5B.

Distribution of total time spent in clinic over 18 months by breast cancer stage (ANOVA p- value <.001), and average total time spent in clinic per month over 18 months by breast cancer stage.

Discussion

We have developed a simple method for quantifying an important aspect of treatment burden using clinical encounter data derived from the outpatient scheduling and inpatient ADT systems. The results of the ANOVA tests from Table 1 confirm our hypothesis that there is significant separation between stages I-III breast cancer patients with respect to the number of distinct appointments, total time spent in clinic, and total number of hospitalized days. We also observed that treatment burden diminished over time in the 18-month period following diagnosis (Figures 5B). This decrease is expected since the most intense portion of treatment for early stage breast cancer is usually completed in the first 12-14 months after diagnosis. Our metric is concordant with this pattern of treatment.

This approach for assessing burden of treatment is simple and generalizable to other healthcare organizations. Since Stage 1 Meaningful Use was enacted, healthcare systems are incentivized to maintain accurate patient encounter records²¹. Therefore, any healthcare organization could use scheduling data to approximate patient burden. On the other hand, some limitations impede direct comparisons between our study population and those at other institutions. Other institutions may record appointments differently than VUMC. For example, another healthcare system with the patient in Figures 1 could have just one appointment for lab, medical oncologist, and infusion. Meanwhile at VUMC, each of those encounters is recorded as a separate appointment. Comparing patient populations within that institution would still be possible but comparing populations across institutions might be challenging.

Our method of determining the amount of time spent in clinic is a gross estimation. It does not take into consideration situations where patients arrive early, appointments start late, patients leave between appointments, or appointments end earlier than the time allotted. Using data from systems such as VUMC’s outpatient whiteboard would enable detailed analysis of patient whereabouts²². For even more granular data, healthcare systems have used real time locator systems (RTLS) to pinpoint the location of patients as they move through the medical center²³. These advanced techniques would provide precise data about patient burden during patient encounters but would be difficult to generalize to other institutions that do not have infrastructure for patient tracking.

Future work in developing quantitative measures for assessing treatment burden includes incorporating additional factors that influence patient outcomes identified in previous literature. To more accurately capture the patient experience related to appointments, we plan to incorporate commute time into their burden assessment by adding the time to drive from their home address to the clinic address before and after appointments. We will use other structured data such as medication prescriptions to determine the frequency of home medication use, and billing information to approximate other medical encounters not captured directly as appointments. There is also potential for natural language processing of notes to capture other provider recommended activities crucial to outcomes such as exercise or diet changes. Complementary to the need for an accurate assessment of a patient’s treatment burden is the determination of a patient’s capacity for treatment. With burden and capacity, we can compare outcomes for patients for whom burden exceeds capacity against those who receive care within their means.

There are multiple applications where a data driven and quantitative measurement of treatment burden could impact patient, provider, and system level decision-making. Using the electronic health record to characterize burden is much like phenotype classification, where researchers use electronic health data to group similar patients for more personalized care delivery²⁴. Understanding treatment burden can aid in delivering the right amount of care that is prioritized to what each individual can handle. When patients are newly diagnosed with stage I-III breast cancer, Figures 5B could help them anticipate how much time they will need to devote to coming to receive care and how much time they will need to take off from work. Likewise, Figures 4B could help them understand how much time patients like them are admitted in the hospital. Future work could further divide our cohort into patients who chose different treatment paths such as prophylactic contralateral mastectomy followed by reconstruction compared to lumpectomy with radiation therapy. Showing the treatment burden of similar patients could help educate patients about treatment options and their trade-offs.

Our analysis could also help with provider monitoring of patient burden. Our work looking at treatment burden is analogous to using electronic health records to predict risk for readmission. In readmission prediction, clinicians use social and clinical factors to identify patients who are high risk for readmission, which is costly to the medical center and harmful to the patient²⁵. Similarly, treatment burden factors can identify patients at risk for non-compliance and wasted resources which are also costly to the medical center and harmful to the patient. The boxplot in Figures 4A is a good example of how healthcare systems can use appointment data as a proxy for treatment burden to identify outliers. In our population, a provider would notce that there are five stage I patients who had around 150 appointments in an 18-month period. A patient care team for one of those patients could investigate whether the appointments are appropriate. If that care is necessary, the healthcare system may look into ways to help ameliorate burden such as home visits or transportation assistance. Additionally, these types of data demonstrate the potential risks in pursuing alternative payment models such as bundled payment models. Administrators and providers could use these tools to help monitor for patients at high risk for high care utilization.

Focusing on outliers in appointment burden could also identify opportunities for improved care coordination and more convenience for the patient. For stage I patients in Figures 5A, there are many more outliers for total time spent in clinic than for count of appointments in Figures 4A. Since Figures 4A and 5A visualize the same population of patients, an increase in the number of outliers means that some patients who are outliers in time spent in clinic either have longer appointments or more time between their appointments. These patients would be candidates for care coordination interventions such as arranging their appointments closer together on the same day, or assigning them to a medical home clinic.

There were several potentially confounding factors that we attempted to control for in our cohort selection. The first and most significant is that we had to determine which patients received their first course of treatment at VUMC. Although the cancer registry had information about which providers saw patients in the registry and what institutions they were from, the availability of that data was inconsistent. We decided on a data driven approach where we defined patients as having received their first course of treatment at VUMC if they had at least three appointments with both a medical oncologist and surgeon from VUMC in the first 18 months of treatment. This constraint cut our cohort by more than 75%, but enacting the constraint was necessary to ensure that analysis focused on patients receiving care where we had more complete data on their encounters. We limited our analysis to stage I-III patients for a similar reason. Stage 0 and incurable stage IV patients have very different patterns of care making them less comparable in the first 18 months of treatment to stage I-III patients. We chose 18 months as the interval for analysis since a typical course of treatment occurs within that time frame and there is a very low risk of disease recurrence during this time²⁶.

Another limitation of this study is that we were not able to address missing data from patients lost to follow-up during the 18-month time span, or for care patients received outside of our institution. Furthermore, during the 17- year period where we observed our patients, there were changes in the way appointments were recorded in the system. There was a gradual increase in appointments per patient due to an increase in departments using the scheduling system. The effect of this increase in appointment capture should have been minimal on the analysis since they would be equally distributed across patients in the different stages.

Conclusion

With increasing interest in care coordination and value based care, there are already many reasons to limit unnecessary treatment for chronic diseases. Evidence is mounting that appropriate care for patients with a given disease may differ depending on their capacity to handle the burden of treatment. Our study of patient encounters is the first step toward a comprehensive and automated method to assess treatment burden applicable to disease population comparisons. In comparing patients with breast cancer, we found significant differentiation in appointments and admissions for patients with stage I, II, and III disease, as well as a downward trend in amount of time spent in outpatient appointments over time after diagnosis. Future work includes incorporating more factors that influence treatment burden and comparing treatment burden across different types of chronic disease. By better understanding burden of treatment, we can begin to deliver precision medicine not only based on genetic makeup and disease phenotypes, but also on the patient’s capacity to comply with treatment plans in order to maximize the likelihood for improved outcomes.