Healthcare inequalities in palliative and hospice end-of-life care delivery for patients with advanced cancer

<p>In previous work[1-2], we describe care utilization as a time-series signal for each day starting from 6 months prior to death, which we called a signature. This dataset focuses on palliative and hospice care delivery using U.S. Centers for Medicare and Medicaid Services (CMS) administrative claims data. Beneficiaries were 66-99 years of age, diagnosed with poor-prognosis advanced cancers as defined here[2], had at least one hospital visit, and died between April 1st, 2016 and December 31st, 2016. Each beneficiary was assigned to the hospital that provided the preponderance of their care during the last 6 months of life. For hospital-level analyses, we only included hospitals that had at least 11 White people and 11 people of color for both palliative and hospice care delivery.</p>

<p><b>Methods for processing the data: </b>
We combined systems engineering and data science to develop palliative and hospice hospital signatures that represent the inequality of care utilization between two racial groups: White people (WP) and People of Color (POC). Using the beneficiary race code from the Master Beneficiary Summary file, we defined WP as non-Hispanic White patients and POC as Black or African-American, Asian/Pacific Islander, Hispanic, American Indian/Alaska Native, Unknown, or Other patients. Signatures were calculated in two steps. First, for each hospital, the percent of patients having received palliative care was calculated for each day prior to death, for a total of 201 days. This calculation was performed twice, one for WP and one for POC. Second, the difference between these two 201-vector signals was calculated, where the difference signal (WP-POC) is the palliative hospital signature that captures the inequality in palliative care delivery between WP and POC over time. These two steps were then repeated for hospice care. We also calculated palliative and hospice hospital quality measure inequalities in care delivery. This included calculating the percentage of patients who received palliative care by death for the two groups. Next, the difference between the measures for each group was calculated as the measure inequality value with a range potentially of -100 to 100. The same was calculated for hospice care.
For both hospital measures and signatures, we identified high-level narrative descriptions that represent the inequality. For measures, these included more people of color (morePOC) for values less than -5%, more White people (moreWP) for values greater than 5%, and no difference (noDiff) for values in the range of -5 to 5%. For signatures, cluster groups included: more people of color received care (morePOC), more White people received care (moreWP), no difference (noDiff), and two new patterns: more White people received care followed by more people of color receiving care (switchWP), and more people of color receiving care followed by more white people receiving care (switchPOC). </p>

<p> Data-specific information for each file is contained in 01_README.txt. </p>

<p>1 Khayal, I. S., O Malley, A. J., & Barnato, A. E. (2023). Clinically informed machine learning elucidates the shape of hospice racial disparities within hospitals. NPJ Digital Medicine, 6(1), 190.</p>
<p>2 Khayal, I. S., Brooks, G. A., & Barnato, A. E. (2022). Development of dynamic health care delivery heatmaps for end-of-life cancer care: a cohort study. BMJ open, 12(5), e056328.</p>