Within-hospital hospice racial disparity signatures

<p>In previous work, we describe healthcare utilization as a time-series signal from 6 months prior to death until the day of death [1]. This dataset specifically focuses on hospice utilization. We combined clinical information, systems engineering, and data science to develop hospital signatures that represent hospice utilization disparity between people of color and white people daily for six months before death for each hospital. We included hospitals with at least 11 people of color and 11 white people with poor-prognosis cancers. Specifically, for each patient with advanced cancer, hospice claims were used to identify which days prior to death they were enrolled in hospice. This data was converted into a 200-day vector of values (i.e., signal), with 0 for days not in hospice and 1 for days in hospice. For each hospital, patients were classified into one of two groups: white people (Non-Hispanic white) or people of color (Black or African-American, Asian/Pacific Islander, Hispanic, American Indian/Alaska Native, Unknown, Other) using the beneficiary race code from the Master Beneficiary Summary file. The signals for all patients belonging to a group were summed to generate the people of color and white people signals for each hospital. The signals were then normalized by dividing the signal values by the number of patients in each group. Next, a straightforward vector difference was calculated by subtracting the white signal from the minority signal to produce the difference signal, which results in positive or negative values from -100 to 100. In the case when the daily value is an event indicator variable, the difference described here is a difference of percentages. Positive values represent the percentage of higher hospice utilization by white people relative to people of color, and negative values represent the percentage of higher hospice utilization by people of color relative to white people. Each hospital difference signal represents the disparity signature for that hospital, which in the simple case of binary indicators for hospice status on a given day corresponds to the average difference in the number of days in hospice spent by white people compared to people of color at that hospital. Each row represents a time point labeled from 0 to 200, where 200 represents the day of death. Each column represents a hospital labeled by the hospital's CMS Certification number. There are also 7 columns for each of the “average” signals calculated for each of the hospital signature clusters (1-7) identified using machine learning from the related publication citation.<p>

<p>[1] Khayal IS, Brooks GA, Barnato AE. Development of dynamic health care delivery heatmaps for end-of-life cancer care: a cohort study. BMJ Open 2022;12:e056328. doi: <a href="https://doi.org/10.1136/bmjopen-2021-056328" target="_blank">10.1136/bmjopen-2021-056328</a></p>