Cluster analysis of COVID-19 recovery center patients at a clinic in Boston, MA 2021–2022: impact on strategies for access and personalized care

COVID Recovery center

The Brigham and Women’s Hospital (BWH) COVID-19 Recovery Center (CRC) was designed to incorporate strategies to address inequities in care. Our structured approach to comprehensive care in the CRC is particularly important for patients who are minorities, vulnerable, or disadvantaged. The CRC is a part of BWH (an academic medical center) and Brigham and Women’s Faulkner Hospital (an affiliated community hospital). Patients are referred through the hospital system (largely internally through primary care physicians or other clinicians in the system) or from the community. We began seeing patients in April 2021 and the effort is a multi-divisional collaboration, with subspecialty care including primary care, neurology, psychiatry, otolaryngology, cardiology, gastroenterology, rheumatology, allergy, dermatology, sleep medicine, physical medicine and partnerships with pharmacists and social workers. At the time of this study, the CRC had seen 1285 patients. Our core equity group designed strategies that promote equity through education and community partnerships, with a robust system of monitoring to ensure patients whom we serve in our recovery center reflect those who have borne a disproportionate burden of COVID-19. The multidisciplinary team also includes a community resource specialist who is an integral member in championing the equity mission within neighboring communities. To modify care barriers, we also reserved funds for transportation reimbursement, and created social support groups. This work was an iterative process, adapted throughout the year to focus on different aspects of community partnership building and use of tools that prioritized care for patients from communities most impacted. The CRC also provides opportunities to enroll eligible patients in the NIH Recover Study, supporting ongoing efforts to better understand post COVID impairments.

Data collection: Variables, metrics and indicators for care delivery

Performance indicators designed prior to deployment of the CRC included indicators for completeness of data collection, proportional visits (patients seen in the CRC who were previously hospitalized, goal benchmark 30%), community engagement (patients referred to the CRC through community partnerships, goal benchmark at least 40%), interpreter access (percent of non-English speaking patients with timely access to interpreter services) and resource referral (percent of patients referred to social work or additional community resources). We prospectively collected data for research purposes as patients referred to the CRC. These data included demographic variables (including age, sex, race/ethnicity, sex); COVID-19 vaccination history, smoking, insurance status, COVID-19 related symptoms, support utilization, and additional variables described in detail below.

Statistical analysesOverview of study design

We included study participants with complete data for sex, race, language, and insurance status. We used these four socio-demographic variables as inputs into a clustering algorithm to classify individuals into subgroups, or clusters. We applied multivariable linear and logistic regression frameworks to test the association of clusters with anthropometric, clinical outcome, symptom, and resource utilization measures.

Cluster analyses

We used socio-demographic variables (sex, race, language, and insurance status (“government”, “commercial”, “other/none”)) to calculate Gower distances and create a distance matrix for non-continuous data. Gower distances provide a value between 0 and 1 that describes the dissimilarity of two data points; this approach, in contrast to Euclidean distance-based methods, allows one to quantify distances for mixed data types, including categorical, binary, and continuous variables [15, 16]. We then performed hierarchical clustering with complete linkage using the R cluster package. We chose these socio-demographic factors variables based on prior knowledge of SES risk factors (as described in the background [2, 5]) ease of variable acquisition, and completeness in the data. For specific variables, prior knowledge on the role as risk factors for acute infection or prolonged symptoms during COVID-recovery are described, such as for race [2], language/interpreter service use [17], sex [18], and insurance status [19]. To determine the optimal number of clusters, we computed the within-cluster sum-of-squares over a range of cluster numbers and chose the number of clusters at the inflection point at which additional clusters does not substantially improve error (i.e., inflection point on an “Elbow plot”). We assigned each patient a corresponding cluster number based on this hierarchical clustering algorithm.

Predictors

We coded cluster assignment as a categorical variable in which lowest ICU admissions cluster (cluster 2) was used as the reference group – that is, the group with the lowest initial disease severity.

Outcomes

We selected outcomes related to post-COVID-19 care based on clinician input. We examined the association of each predictor with the need for an ICU admission, symptoms, and utilization of SES-targeted resources. Symptoms included cough, dyspnea on exertion, anxiety, fatigue, and brain fog. SES-targeted resources and support utilization included social support services that provided one of the following services: psychoeducation for chronic illness, help navigating government benefits, help addressing financial/housing concerns and obtaining community resources, arranging for support groups, and providing care coordination.

Model specifications

We utilized multivariable linear and logistic regression models, as appropriate. We performed unadjusted and adjusted analyses. We adjusted all models for age in years, but not other socioeconomic factors as these variables were incorporated into the clustering analyses.

We performed all analyses in R v4.0.3 [20], a programming language and environment that supports data analyses and advanced graphical solutions. We performed linear regressions with the “lm” function and logistic regressions with the “glm” function. We assessed all variables for normality by visual inspection of histograms and Shapiro–Wilk tests. We compared continuous variables with Student t-tests or Wilcoxon tests, as appropriate. We compared categorical variables with analysis of variance (ANOVA) or Kruskal–Wallis tests, as appropriate. We reported 95% confidence intervals and considered two-sided p-values below 0.05 to be significant.

留言 (0)

沒有登入
gif