Hidden figures underlying quality measures: revealing hidden racial inequalities in end-of-life cancer care delivery: a cohort study

Summary

Background

Calls for healthcare systems to reduce disparities in cancer care access and outcomes draw on comparisons of existing measures across race and ethnicity subgroups. This approach may hide inequities driven by systematic bias in the timing of care delivery. The goals of this study were to: (1) identify differences in the timing of care delivery between racial groups, and (2) determine whether these differences could be identified from quality measures.

Methods

Retrospective decedent follow-back study of hospitals treating Medicare fee-for-service beneficiaries with advanced cancer aged 65–99 who died April–December 2016. Among hospitals serving at least 11 decedents of color (including Black or African-American, Asian/Pacific Islander, Hispanic, American Indian/Alaska Native, and Other) and 11 White decedents, we calculated hospital-level differences between White decedents and decedents of color for 1) any use of palliative care and hospice (Measures) and 2) daily counts of palliative care and hospice use for each day in the 6 months before death (Signatures).

Findings

The cohort included 30,319 decedents across 217 hospitals, of whom 7,852 (25.9%) were people of color (POC). The median of the hospital-specific aggregate measure difference was −5.35% (IQR = 12.83) for palliative care, indicating more POC received any palliative care, and 3.66% (IQR = 12.45) for hospice care, indicating more White people (WP) received any hospice care. We identified 5 high-level cluster-group descriptions of inequality from signatures. Inequality information from signatures matched those from measures in only 46.5% and 39.2% of hospitals for palliative and hospice care, respectively.

Interpretation

Signatures incorporating timing of care delivery using longitudinal data revealed patterns of racial-ethnic inequalities in end-of-life cancer care otherwise missed by traditional aggregate quality measures.

Funding

This work was supported by the American Cancer Society Award (RSG-22-128-01-HOPS).

Research in context.

Evidence before this study

In the United States, racial and ethnic inequalities have been identified across patients with many diseases (e.g., cancer, diabetes, heart disease). Methodologically, racial inequalities have been calculated as differences in quality measures between racial groups. These quality measures are typically calculated at a single time point over a time range (e.g., the last 6 months of life). However, racial inequalities have also been identified in time-to-treatment studies, signaling timing of service as a critical element of quality. And yet, quality measures only capture “if” a service was provided rather than “if and when” a service was provided. For example, for previously endorsed National Quality Forum measure NQF#0215, “Proportion Not Admitted to Hospice,” a patient may either be included or not in the numerator if they received hospice care, but with no consideration as to when they received this care, despite the quality of hospice being directly reflected by the duration of time in hospice. Currently, the only approach to address time was to create a new measure is (e.g., NQF#0216, “Percentage of Patients Admitted to Hospice for Less than 3 Days”) focused on a specific time range. In this case, 3 was selected based on clinical judgment as the minimum number of days required for achieving “some benefit” from hospice, but clinicians agree that earlier hospice enrollment is better. Creating new measures that can focus on one specific aspect of time is not an efficient or tractable approach to incorporating time.

To identify studies that sought to understand quality measures calculated dynamically or longitudinally over time, we searched PubMed from inception to August 19, 2024, using the query “(“Quality measures” [All Fields] OR “Quality metrics” [All Fields] OR “Performance Metrics” [All Fields]) AND (“Dynamic” [All Fields] OR “Longitudinal” [All Fields] OR “Time-Varying” [All Fields] OR “Temporal” [All Fields] OR “Time [All Fields]”).” We identified no studies that incorporated daily time points in the quality measure calculation.

Added value of this study

In this study, we identified racial inequality in care delivery using (1) traditional aggregate measures of end-of-life cancer care quality and (2) longitudinal signatures constructed using the same numerator and denominator but calculated daily receipt for 6 months before death (i.e., a classic engineering approach to modeling temporal information). The traditionally calculated measure is the final value in the set of values that make up the hospital signature. In more than half of the hospitals evaluated, determining racial inequalities based on the final value alone (i.e., the last chance to determine if care was delivered) missed racial inequality in when care was delivered. Therefore, signatures can reveal longitudinal racial inequalities in palliative and hospice care use that are hidden in the aggregated measures. This suggests that quality measures identify racial inequalities between population groups based on “if” a service was delivered, but fundamentally lack the ability to identify racial inequalities in “when” service was delivered.

Implications of all the available evidence

Methodologically, the classic approach of using quality measures to identify racial inequalities within hospitals captures only one type of racial inequality based on “if” care was delivered and misses another key racial inequality based on “when” care was delivered. Consequently, operationalizing the construct of racial inequity using aggregate-level quality measures may hide racial inequities that exist when the timing of care delivery is a salient aspect of care quality.

Introduction

Healthcare equity refers to fair and equal access to healthcare services, regardless of socioeconomic status, race, ethnicity, age, or other factors.¹ The Institute of Medicine defines equity as one of the 6 domains of healthcare quality.² Systematic approaches to healthcare quality improvement should aim to increase healthcare equity.³ Quality improvement relies on quality measurement. Quality measures are typically calculated cumulatively at a specific time point (e.g., 30-day hospital mortality measured 30 days after admission). Disaggregating this measure by racial groups is the most common approach to identifying inequalities in cancer care delivery.

Racial healthcare inequities have garnered national attention because addressing them is crucial for achieving high-quality healthcare and improving population health. We hypothesize that the current approach of disaggregating standard metrics is insufficient and potentially misleading. Specifically, existing methods of calculating quality metrics that aggregate events over a particular time frame may hide racial/ethnic inequalities in timeliness of care delivery. In fact, timeliness is a critical element of care quality² that has been shown to be a major source of inequality in care delivery. For instance, racial inequities in cancer have been identified in time-to-treatment for cancers of the breast,4, 5, 6 lung,⁷ prostate,⁸ colon,⁹ melanoma skin,¹⁰ endometrium,¹¹ bladder,¹² and brain.^13,14 Racial inequities have also been identified in the timing of specific services, such as time in hospice care.15, 16, 17, 18 While we use cancer as an exemplar related to our cohort, this time-to-treatment problem exists in all clinical conditions.

To address the missing time axis in quality measurement, we applied a dynamic modeling approach that effectively calculates delivery system quality metrics for each day. For this paper addressing end-of-life care quality, we use each day prior to death for 6-months before death to create a longitudinal set of metric values, which we coined a signature (i.e., a set of 200 values, approximately one for each day in the 6 months before death).¹⁹ The last value in the signature is the traditionally calculated metric. We applied this approach to a cohort of patients with advanced cancer, focusing on receipt of palliative and hospice care. Palliative care, often delivered concurrently with cancer-directed treatment, has been demonstrated to improve the quality of life for patients with advanced cancer and may lead to increased uptake of hospice care.²⁰ Hospice enrollment is widely recognized as an important component of high-quality end-of-life cancer care. Recognizing the value of time in hospice, the National Quality Forum sought to incorporate time into the metric by defining enrollment in hospice care for 3 or more days before death as a cancer care quality measure rather than simply any receipt of hospice before death.

In this paper, we explore whether signatures of palliative and hospice care elucidate racial inequities that are otherwise hidden in traditional approaches to calculating quality measures.

Methods

Patient cohort and claims data

The patient cohort included a 100% sample of Medicare fee-for-service beneficiaries diagnosed with poor-prognosis advanced cancers, defined as metastatic cancers and primary cancers associated with a high mortality risk.^21,22 Inclusion criteria were: (1) death between April 1-December 31, 2016, (2) age 66–99, and (3) availability of a 6-month look-back period. For more information about sample selection, see Wasp²² and Khayal¹⁹ et al. Centers for Medicare and Medicaid Services (CMS) files included the Master Beneficiary Summary, Medicare Provider Analysis and Review (MedPAR), Physician/Supplier Carrier, Outpatient, and Hospice records. We assigned beneficiaries to hospitals based on the majority of their hospitalizations during the last six months of life. The Dartmouth Committee for the Protection of Human Subjects IRB (CPHS) concluded that this study does not qualify as human subjects research, as it solely involves data from deceased individuals (CPHS# STUDY00033050).

Calculating traditional hospital measures of quality

We focused our analysis on two indicators of high-quality care for advanced cancer: (1) receipt of palliative care²⁰ (palliative measure), calculated as the proportion of decedents who received palliative care in the 6 months before death, and (2) receipt of hospice care²³ (hospice measure), calculated as the proportion of decedents enrolled in hospice before death. Receipt of palliative care was based on the diagnosis code Z51.5 (encounter for palliative care) found in the MedPAR and Carrier files. Admission to hospice was determined from the Hospice file (National Quality Forum #0215).²⁴ While practice guidelines recommend integrating palliative care into oncology care, no formal quality measure has been defined for palliative care. Therefore, we use the phrase ‘measures of quality’ to refer to our study metrics and ’endorsed quality measures' when writing about formally endorsed measures (i.e., CMS, NQF).

For each hospital, decedents were categorized into either the non-Hispanic white (white) group or the people of color (POC) group. The latter includes Black or African-American, Asian/Pacific Islander, Hispanic, American Indian/Alaska Native, and Other decedents. While the POC group conflates heterogeneous population sub-groups with unique healthcare experiences, they share a history of marginalization in the U.S. This classification was based on the beneficiary race code (modified by the RTI race algorithm) from the Master Beneficiary Summary file.²⁵ At each US hospital, the measures of palliative care and hospice receipt were independently calculated for two demographic groups: White people (WP) and POC. We included only those hospitals with a minimum of 11 decedents of color and 11 White decedents, adhering to the CMS suppression rules for data inclusion. We additionally calculated inequality in measures of quality as the difference between WP and POC at each hospital. We characterized hospital-level differences into three categories: more people of color receiving the service (morePOC) for WP-POC differences of −5% or less, no difference (noDiff) for differences between −5% and 5%, and more White people receiving service (moreWP) for differences greater than 5%. We chose the 5% cutoff as a perceived meaningful difference.

Calculating hospital signatures

Signatures extend measures of care delivery and quality by retaining, rather than aggregating, the dimension of time.^19,26

This approach is based on the novel dynamic modeling of healthcare engineering systems.^27,28 Specifically, we calculated signatures for the (1) Proportion of patients who received palliative care for each day in the last 6 months of life (palliative signatures) and (2) Proportion admitted to hospice for each day in the last 6 months of life (hospice signatures). At each hospital, we calculated signatures for White people and people of color and subtracted these group signatures (WP-POC) on a day-by-day basis to calculate the inequality in care delivery for the last 6 months before death (for both hospice and palliative care measures). Each signature was then classified using an agglomerative hierarchical machine-learning algorithm. Methodological details have been previously published.¹⁹ Briefly, the algorithm (see Supplementary Material A) grouped similar signatures into clusters, which we then grouped into high-level cluster groups. Next, we checked each classification and adjusted the cluster group if needed. Finally, we re-ran the unsupervised machine-learning algorithm for each cluster group to identify the number of clusters for each cluster group.

Comparing measures to signatures for individual care services

For both palliative and hospice care services, we conducted Kruskal–Wallis tests to evaluate whether significant differences existed between measure-values and signature-clusters. At the hypothesis testing p-value threshold of p < 0.05, we conducted Conover–Iman tests to identify which specific group differences were statistically significant. To highlight the contrasting insights between measures and signatures, we developed a colored, bar-stacked histogram to visualize the differences in informational yield from each. In addition, we included hospital exemplars showcasing similar measures and signatures, similar measures with different signatures, and different measures with similar signatures. We also performed a sensitivity analysis to assess the sensitivity of the results to specific race sub-groups in our people of color group. For each hospital, we identified if the people of color group included any specific race sub-groups with at least 11 patients. For each racial sub-group, we calculated the hospital signature as described above by replacing the POC group with the specific race sub-group. We then identified the high-level signature-cluster for each racial sub-group. Finally, we compared the new signature clusters for each hospital with the signature-cluster identified by the POC group analysis. If any signature-clusters were different, we marked the hospital as having changed the signature-cluster. Otherwise, the hospital would be marked as having no change.

Comparing measures and signatures for combined palliative to hospice services

We represented the transition from palliative care to hospice care as flows in a Sankey diagram showing the flow from palliative care to hospice care for measure groups and signature-cluster groups. In addition, we used color shading to denote differing measure groups and signature-cluster groups when comparing palliative and hospice care.

All analyses and visualizations were completed using Python 3.9.

Role of the funding source

The study sponsor had no role in the study design; in the collection, analysis, and interpretation of data; in the writing of the report; or in the decision to submit the paper for publication.

Results

Cohort and hospitals

We identified 217 hospitals with sufficient patients to calculate both palliative and hospice values. Our cohort included 30,319 beneficiaries, of whom 7,939 (25.9%) were people of color (see Table 1 in Supplementary Material B). The percentage of people of color attributed to each hospital ranged from 6.3% to 79.0% with a median of 27.2%.

Hospital measures of quality

Fig. 1 visualizes palliative and hospice histograms of measure values for a cohort of White people (WP), people of color (POC), and the difference between the two groups (WP-POC). For palliative care, utilization was greater among POC than among WP, with a negative median difference of −5.35% (Interquartile Range [IQR] = 12.83). For hospice care, utilization was greater among WP than among POC, with a positive median difference of 3.66% (IQR = 12.45). Within a hospital, we identified 111 out of 217 hospitals (51%) that showed more POC receiving palliative care but more WP receiving hospice care. Only 18 hospitals (8%) showed the reverse: more WP receiving palliative care but more POC utilizing hospice care. 62 hospitals (29%) showed more WP receiving both types of care, whereas 26 hospitals (12%) showed more POC receiving both services.

Fig. 1.

Central measures of tendency and dispersion of hospital-level measures of quality for palliative (A) and hospice (B) care utilization by White people (WP), people of color (POC), and the difference between the two groups (WP-POC). Difference values in blue represent more utilization by people of color (morePOC), black represents no difference in utilization (noDiff), and red represents more utilization by White people (moreWP). Despite finding more people of color receiving palliative care (larger blue area under the curve for palliative), we found more White people receiving hospice care (larger red area under the curve for hospice).

Hospital signatures

We clustered palliative and hospice signatures into 11 and 12 clusters, respectively, and 5 high-level cluster-groups, as shown in Fig. 2. The cluster-groups described signatures where (1) more people of color received a specific service consistently over time (morePOC), (2) more White people received a specific service consistently over time (moreWP), (3) no difference between people of color and White people over time (noDiff), (4) more White people received a service initially, followed by a switch to more people of color (switchWP), and (5) more people of color received a service initially, followed by a switch to more White people (switchPOC). The clusters within the five cluster groups identified sub-groups that relay specific disparity differences in time and magnitude. For example, palliative cluster #1.1, reveals racial differences emerging around 90 days before death and increasing to an average of a 10% disparity of more people of color receiving palliative care. In palliative cluster #1.2, racial differences emerge almost 180 days before death, with more people of color receiving early palliative care to a magnitude beyond a 20% difference.

Fig. 2.

Signatures for palliative (A) and hospice (B) care utilization for each hospital over the last 6 months of life. Each signature is color-coded based on its high-level cluster group and placed into its respective cluster sub-plot. For each cluster sub-plot, the thick black line represents the average signature. Signatures beyond a single pattern (e.g., morePOC, noDiff, moreWP) were identified across time (e.g., switchWP, switchPOC).

Measures to signatures for individual care services

We visualized the parity of information between measures and signatures for each service individually in two ways. First, through boxplots that display quality measure-values segmented by cluster number for palliative and hospice care, in Fig. 3A and B, respectively. We found no statistically significant differences in the medians between either the morePOC (blue) and switchWP (pink) clusters or the moreWP (red) and switchPOC (purple) clusters, likely because the measures of quality at the end are similar. Therefore, clusters or cluster-groups cannot be identified from the traditionally-used measures of quality. However, we could impute the quality measure from the quality signature with 100% accuracy as it is the last value in the signature. Second, we extended Fig. 1 histogram into a bar-stacked histogram showing the breakdown of high-level cluster-groups within each bin for palliative and hospice care in Fig. 3C and D, respectively. Within the same bin, the distribution of multiple high-level cluster groups suggests that different information is captured in measures and signatures. Measures and signatures matched, both showing morePOC, moreWP, and noDiff, in 82 (38%), 11 (5%), and 8 (4%) hospitals for palliative care and in 26 (12%), 52 (24%), and 7 (3%) hospitals for hospice care, respectively. Therefore, signatures could be predicted from measures in less than 47% of hospitals for palliative care and in less than 40% of hospitals for hospice care.

Fig. 3.

Box plot values of palliative (A) and hospice (B) measure differences between white people (WP) and people of color (POC) for hospitals in each of the respective signature clusters and cluster-groups described in Fig. 2. Hospitals with different signature inequalities could not be distinguished (i.e., no significant difference) by measure values (e.g., morePOC & switchPOC and moreWP & switchPOC). Bar-stacked histograms of palliative (C) and hospice (D) measure differences, with signature clusters stacked within each bin. Most histogram bins include some level of similar measures and signatures (i.e., same color) indicating the same inequality insights from measures and signature. However, the histograms are very heterogeneous showing that each measure color may be categorized into most other signature colors, indicating very different inequality insights from measures and signatures.

In Fig. 4, we showcase the three possible patterns of measure-and-signature combinations. First, similar measures and signatures (4A) highlights that in some instances, hospitals have similar measure-values and similar signature-cluster-groups. In other words, two hospitals may believe they are similar because they have similar end values and they have similar longitudinal data, which suggests that they have similar patterns of inequities. Second, similar measures with different signatures (4B) highlight that despite having similar measure-values, the signature patterns of inequities that occur longitudinally prior to the last endpoint are different and not represented by the measure value. This pattern showcases how measures calculated at the last point can “hide” inequities occurring earlier in time. In other words, two hospitals may believe they are similar because they have similar end values, but longitudinal data suggests that they have different patterns of inequities. Third, different measures with similar signatures (4C) highlight that similar longitudinal patterns can exist over time, which vary only at the very end (of the timespan) with different measure-values. This pattern showcases how measures calculated at the end may not be representative of the overall pattern of inequities occurring over time. Here, hospitals may believe they are different based on their measure-values, but longitudinal data suggests that they have similar patterns of inequities.

Fig. 4.

Hospital exemplars (boxed numbers correspond to a hospital's Medicare CMS Certification Number—CCN) with similar measures and signatures (A), similar measures with different signatures (B), and different measures with similar signatures (C). Measures are depicted with circles, representing moreWP (in red), noDiff (in black), and morePOC (in blue). (A) highlights how measures and signatures can relay the same information. (B) highlights how hospitals with similar measure-values may believe they have similar inequities, but their signatures reveal different patterns. (C) highlights how hospitals with similar longitudinal patterns over time may vary only near death with different measure-values.

We performed a sensitivity analysis to assess signature sensitivity by specific race sub-groups. We identified three race sub-groups with sufficient patient numbers: Black, Hispanic, and Asian. Of the 217 hospitals, 118 and 121 hospitals had sufficiently large sub-groups for sensitivity analyses for hospice and palliative signatures, respectively. For hospice signatures, 106 hospitals included at least one race sub-group, with 98 hospitals showing similar cluster-groups and 8 hospitals showing differences; therefore, analyses by sub-group versus POC led to similar results in 94% of hospitals. For palliative signatures, 106 hospitals included at least one race sub-group, with 102 hospitals showing similar cluster-groups and 4 hospitals showing differences; therefore, analyses by sub-group versus POC led to similar results in 96% of hospitals. For hospice and palliative care signatures, the same three large hospitals included all three race sub-groups, with 1 hospital showing similar cluster-groups and 2 hospitals showing differences, therefore, analyses by sub-group versus POC led to similar results in 33% of hospitals; and the same seven hospitals included two race sub-groups, with 4 hospitals showing similar cluster-groups and 3 showing differences, therefore, analyses by sub-group versus POC led to similar results in 57% of hospitals. This analysis shows that at predominately large hospitals with multiple race sub-groups, the increased statistical information obtained from combining race sub-groups may lead to potential differences in signature cluster-groups.

Measures and signatures for combined palliative to hospice services

This section showcases the parity of information within the same hospital between palliative and hospice services. Since palliative care tends to be delivered prior to hospice care, we visualized the flow from palliative care to hospice care using Sankey flow diagrams that start with a palliative source node on the left that flows into a hospice target node on the right with a thickness density representing the count of hospitals. By converting measure-values and signature-clusters into meaningful inequity groups of morePOC, noDiff, moreWP, switchWP, and switchPOC, we can describe each hospital by its palliative and hospice inequity description, based on measures in Fig. 5A and signatures in Fig. 5B. The colored links indicated hospitals with different information between measures and signatures, while the light gray links represented hospitals with similar information from palliative to hospice care. For instance, the inequity in hospice care could be predicted from palliative care in only 28.6% (62/217) and 38.7% (84/217) of hospitals for measures and signatures, respectively, demonstrating the insufficiency of quality measures that do not consider dynamic shifts in care quality and continuity for different groups over time.

Fig. 5.

Sankey diagrams visualizing the flow of inequity information from palliative to hospice measures (A) and palliative to hospice signatures (B). Flows from palliative to hospice inequity information that are the same are shown in light gray, and those that are different are shown in color. The inequity in hospice care could be predicted from palliative care in only 28.6% (62/217) and 38.7% (84/217) of hospitals for measures and signatures, respectively. Consequently, the traditional strategy of improving hospice inequities by changing upstream palliative care may not be successful if it doesn't consider actual patterns of hospital-specific racial inequalities—explained in detail in the results through an example based on data in the green box.

The actionable value of this analysis can be clarified through an example in Fig. 5A (green box). This boxed region highlights hospitals where more white people received hospice care (hospice moreWP). A classic quality improvement intervention strategy, based solely on hospice data, would be to increase upstream palliative care delivery for people of color. This strategy could be evaluated using a data-driven approach, examining a hospital's current palliative and hospice care practice behavior and equity at a hospital. The Sankey diagram shows that 14 hospitals had more palliative care use by white people, suggesting that increasing palliative care use by people of color may be effective. However, the figure also shows that 35 hospitals—more than double the previous group—had no difference in palliative care use between people of color and white people, and yet, more white people received hospice care. This raises the question of whether increasing palliative care use among people of color would translate into more hospice care, since it did not occur initially. Interestingly, 50 hospitals showed a pattern where more people of color received palliative care, yet more white people received hospice care. This suggests that simply increasing palliative care for people of color may not address hospice care disparities. The data indicates that a deeper, more nuanced understanding of why palliative care consultations are not leading to hospice care is needed. Such an understanding is essential for designing targeted interventions that address the specific inequities at each hospital. The goal of this data-driven approach to incorporating time and multiple services is to provide information to hospitals that can serve to guide where to look within their hospital and electronic medical record (EMR) to develop tailored solutions that address the underlying behavior uncovered.

Discussion

In this study, we identified hidden patterns of racial/ethnic inequity in end-of-life cancer care that cannot be identified from quality measures despite using the same administrative claims. Our findings suggest that signatures reveal racial inequalities hidden in the longitudinal time domain that are missed using traditional quality measures; therefore, current quality measures are insufficient to reveal the full extent of racial inequities. Signatures do not directly identify causes of inequality; instead, they provide hospital-specific information about the timing and extent of inequalities to guide internal hospital quality and equity improvement processes to identify causes of inequality. Elucidating longitudinal dynamic behavior of cancer care delivery across multiple services can support decision-making for quality and equity improvement.

Inequities identified in signatures suggest that current methods of calculating measures based on static or aggregated time unfairly hide a form of racial inequality that manifests in the longitudinal time dimension. These current methods of treating time statically or, in the aggregate, fundamentally ignore the literature showing racial disparities in ‘time-to-decision,’ ‘time-to-treatment,’ and ‘length of stays’ across many cancers, as well as other conditions,4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14 and suggests that identifying the receipt of care may be as critical as the time of receipt of care. In addition, quality measures, especially process measures, provide receipt-of-care information that lacks a connection to value-to-patient information. For example, achieving racial equality of hospice use in the last few days of life does not indicate equitable use of hospice. Given that the value of care to the patient is associated with time spent in hospice,²⁹ it becomes clearer how an analysis incorporating the timing and length of care delivery allows for a closer description of perceived quality by patients and their families. Similarly, receipt of palliative care shortly before death may not be as valuable relative to an earlier time in the patient's end-of-life trajectory, with the potential to change trajectory to one that matches their preferences. Consequently, quality measures are insufficient to reveal the full extent and complexity of racial inequities.

Reformulating the concept of quality to include time supports the shift from “does” to “when does” an inequity exist? Signatures provide detailed information about the timing and extent of inequalities for each specific hospital. Signatures are not intended to incorporate or directly identify factors causing these inequalities (i.e., causation). Instead, they are intended to elucidate dynamic macro-level system behavior that is otherwise hidden in quality measures. Feeding back this information to hospitals may serve as bespoke operational feedback that can be actionably used by hospital quality and equity teams to support their efforts to triangulate inequity within a rich electronic medical record to support internal improvement processes such as, for example, root cause analyses. Causes of inequities vary and can be related to patient, provider, or hospital-specific factors. This needs to be investigated independently and varies for each hospital. Therefore, inequity in time information may provide insights into when and potentially where to intervene to target inequity improvement. This is especially critical for measures, such as hospice, that require upstream intervention, such as palliative care integration and/or discussion.

The dynamic information captured in signatures can also elucidate how different populations utilize the combination of services, consequently providing a higher-level understanding of the behavior of end-of-life care utilization that cannot be understood by studying the individual services alone. Prior work suggests that single-value measures provide minimal insights into how multiple services, such as palliative and hospice care, are being used relative to each other.³⁰ The variation in the utilization of multiple services indicates an underlying behavior that differs in how racial groups utilize the combination of services. For example, in a hospital with more utilization of hospice by white people, quality decision-makers may logically envision an equity improvement intervention strategy that increases palliative care services for people of color. However, we found several examples where more white people utilize hospice care and more people of color utilize palliative care. Therefore, the underlying behavior change may not be in the number of people who are made aware of hospice care (potentially from palliative care providers) but in the next step of deciding to enter hospice care. A deeper EMR review could potentially identify why people of color receiving palliative care do not enter hospice (e.g., system-level factors, such as hospice service availability in locations where people of color live, or provider-level factors, such as provider palliative care specialty level or referral practices, or patient-level factors, such as personal choice to not enter hospice). The success or failure of any equity improvement strategy must consider the actual patterns of hospital-specific racial inequalities. Developing analytical strategies that incorporate “time” in understanding multi-service dynamics can help teams design bespoke strategies tailored to the behavior of their individual hospitals.

Our findings should be interpreted with a number of potential limitations in mind. We applied this method to hospitals with a large enough patient pool, 11 or more racially and ethnically diverse patients, and 11 or more White patients. This approach excludes many hospitals. Future studies will incorporate multiple years of data to address this limitation. In this cohort, we included only Medicare decedents seen at a hospital in the last six months of life; therefore, signatures only reflect racial inequities for the population that included a hospital visit. Future work will not impose hospital visit inclusion criteria. Instead, the focus will be on distinguishing inequities between decedents who had a hospital visit, decedents who had an outpatient visit from primary or palliative care, and decedents who had neither a hospital nor clinic visit in the last six months of life. This would allow for identifying other possible inequities in non-hospital-based services. Use of the Z51.5 palliative care code likely varies across hospitals and health systems; this may lead to varying sensitivity of the Z51.5 code for identifying palliative care across hospitals and health systems. However, unless there is heterogeneity in coding Z51.5 within a hospital system by race, we expect our results to stay the same since we calculate differences between two groups at the same hospital. Future work will explicitly examine the varying sensitivity across hospitals and any variations of coding within hospitals. We grouped sub-groups of heterogeneous marginalized populations together to increase the sample size. As part of a sensitivity analysis, we also restricted the population to specific racial sub-groups (Black persons, Hispanic persons, Asian persons) and assessed if a hospital cluster classification changes. In future studies, an alternative may be to apply a Bayesian approach to identify any significantly different patterns. We applied this methodology to Medicare fee-for-service claims data, meaning that our analysis does not encompass patients covered by Medicare Advantage or commercial insurance. To address the prior limitations, future work will also seek to develop hospital-facing tools to allow hospitals to produce racial signatures applied to their own locally available data. This would allow hospitals to explore their data and identify hospital-specific patterns to support the development of hospital-specific tailored solutions.

Conclusion

In conclusion, by constructing a signature of traditional measures of care quality calculated each day in the last months of life, we revealed longitudinal racial inequalities in palliative and hospice care that were not seen in traditionally aggregated measures calculated at the final end-point of death. This suggests that operationalizing the construct of racial inequity using aggregate healthcare use metrics may hide racial inequity in care that exists when quality is viewed longitudinally.

Contributors

ISK secured funding for this work, developed the idea and methodology, analyzed the data, and wrote the first draft. JS analyzed the data and revised drafts of the manuscript. AJO reviewed the methodology and statistics. AMF contributed to the foundational methodology and analysis. GAB, AEB, and TL contributed to iterative data analysis and interpretation. ELA contributed to interpretation. All authors read and approved the final manuscript.

Data sharing statement

Our Data Use Agreement with the Centers for Medicare and Medicaid Services does not allow for the sharing of individual patient data. Hospital-level datasets generated and/or analyzed during the current study are available in the Dartmouth Dataverse repository, https://doi.org/10.21989/D9/EOQOCP.

Declaration of interests

ISK, AEB, ELA, TL, and AJO declare financial support through a National Institute of Health grant (NIH P01AG019783). All authors declare no financial or non-financial competing interests.