Study design

This retrospective, cross-sectional study analyzed deidentified electronic health record data extracted from the Arkansas Clinical Data Repository which included all hospitalizations from the only academic medical center in the state of Arkansas [23]. The University of Arkansas for Medical Sciences is a 535-bed hospital that includes 431 adult beds, 64 newborn bassinets, and 40 psychiatry beds. This study focused on the hospitalizations of adult African Americans with heart failure ranging from January 1, 2014, to December 31, 2023. A hospitalization was defined as an inpatient admission for a continuous stay. The data were accessed for research purposes on February 27, 2024. The data included all study variables related to each patient’s hospitalization. Patient data included pseudonym identifiers, demographics, clinical diagnoses, medical histories, and comorbidities based on each patient’s International Classification of Disease Codes Version 10. The authors did not have access to information that could have identified individual participants during or after data collection.

Logistic regression identified how the configuration of the care delivery teams of African Americans were associated with long LOS. The logistic regression analysis was further stratified to assess the influence of an African American patient’s sex, rurality, overall health (Van Walraven Elixhauser Comorbidity Score), and specific heart health and function (ejection fraction). This additional analysis was performed because prior studies have shown that patents above age 65, patients who are female, patients with poor overall health, and patients with severe heart failure have increased odds for long LOS [1,3–9,23,25]. The Kruskal-Wallis H test further assessed significant differences in the density of each role within the care team configurations. Subsequently, linear regression identified the change in LOS associated with a one-unit change in RNs (i.e., adding or removing one RN) among statistically significant team configurations. Fig 1 provides a workflow and sequence of the exploratory analysis within the study design. The structure and components of the study design was guided by findings and needs identified in prior studies [9,23].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Workflow of exploratory analysis.

https://doi.org/10.1371/journal.pone.0329602.g001

Participant selection and sampling

The study data only included the hospitalizations of African Americans who had a heart failure hospitalization between January 1, 2014, and December 31, 2023. This 10-year period was selected because of data availability. The analysis included all individuals who were African American and 18 years of age and older because they are disproportionately impacted by higher heart failure hospitalization rates and longer lengths of stay than other patients [2–5]. All patients in the dataset had a diagnosis of congestive heart failure, as characterized by left ventricular ejection fraction via transthoracic echocardiogram. Only hospitalizations with at least one or more of the following care roles were included in the analysis because they were previously established as being associated with hospitalization rates and LOS: physician, nurse practitioner, registered nurse, patient care technician, physical therapist, occupational therapist, respiratory therapist, pharmacists, registered dietician, care manager, and social worker [9,12]. Care team configurations with fewer than 15 patients were excluded from the analysis because of the potential for model over parametrization, which is consistent with prior research [9]. All data that met sampling criteria was analyzed. Fig 2 provides a diagram of how the inclusion and exclusion criteria impacted the sample size.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Sample size and participant selection.

https://doi.org/10.1371/journal.pone.0329602.g002

Overall health severity

The Van Walraven Elixhauser Comorbidity Score was also provided for each hospitalization. The Van Walraven Elixhauser Comorbidity Score is a single weighted, numeric score based on each patient’s specific combination of comorbidities and their related mortality risk, an overall measure of health severity during the hospitalization [23–27]. The Van Walraven Elixhauser Comorbidity Score is a scoring system ranges from negative twelve to positive eighty-nine, with the lowest quartile (−12–7) representing patients with the best overall health, and patients with scores ranging from eight to eighty-nine indicating patients with poor overall health [23–27]. Higher Van Walraven Elixhauser Comorbidity Score are interpreted as having a higher likelihood of increased LOS and in-hospital mortality [23–27]. The Van Walraven Elixhauser Comorbidity Score has been previously validated through prior international studies in patients with chronic conditions including congestive heart failure [24–27]. It has also been validated through prior work of the authors which specifically focused on patients with congestive heart failure of all races in Arkansas including those who were African American [23]. These studies have concluded that the Van Walraven Elixhauser Comorbidity Score is a valid and reliable tool for isolating high-risk patients and measuring overall health severity related to the odds of in-hospital mortality [23–27].

Heart failure severity

The dataset also included the ejection fraction rate of each patient as a specific indicator of heart health (the percentage of blood pumped through the left ventricle of the heart) [9,23]. Based on how the patients clinically presented during their hospitalization, patients with less than a 50% ejection fraction reflected cases of heart failure that were the most severe, compared to those with an ejection fraction of 50% or greater [9–11,23].

Care team configuration

An edge list was also provided within the dataset. The edge list linked each patient with each member of the care team who provided their care. The data related to care team members included pseudonyms and care roles by licensure (e.g., physicians, registered nurses, social workers). Unique combinations of the care team roles that provided care were defined as care team configurations.

Care team network density

Network density (a method of social network analysis) was used to map and quantify the relationships between patients and their care team members within each the hospitalization [23,28,29]. Network density represented the number of connections each patient had with each care team member divided by the total number of possible connections the patients could have had with all care team members in the network on the specific day of their hospitalization [29]. Network density values (i.e., ranging from 0–1) represented the percentage of available care team members on each hospitalization date who provided care to each patient (i.e., the percentage of available hospital staff used by each patient) [29]. Network density values were calculated for the overall interprofessional mix of all care roles on a care team (e.g., physicians, nurses, and social workers) who provided care for each patient [23]. Network density values were also calculated by care role (e.g., RN network density) to indicate the percentage of available hospital staff in each care role (e.g., RNs) that was used by each patient.

Statistical analyses

The thresholds presented in Table 1, as well as the dichotomization of age, ejection fraction, Van Walraven Elixhauser Comorbidity Score, and LOS, was consistent with previous studies that applied similar methods, the clinical evaluation of heart failure, and the statistical assumption of the logistic regression analysis (discussed below) [1,9,23,30–34]. Table 1 provides a list of dichotomized variables.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Dichotomous variables.

https://doi.org/10.1371/journal.pone.0329602.t001

Upon receiving the deidentified dataset, Open Refine 2.0 was used to confirm that there was no missingness in the dataset. The data was analyzed using SPSS Version 29. Means, standard deviations, medians, interquartile range, minimum values, and maximum values were calculated to describe LOS by patient characteristic and care team configuration. Logistic regression was used to evaluate how the configuration of the care delivery team of patients was associated with long LOS. Logistic regression was used because the dataset violated the traditional assumptions of parametric tests including normality in the distribution of length of stay, which is consistent with the methodology of similar studies [9,23]. As standard interpretation of logistic regression, the presence of statistically significant odds ratios (p < 0.05, 95% CI) identified team configurations that were associated with long LOS [23]. The regression model tested each team configuration against a reference group which included all patients except for the patients that received the specific team configuration that was being tested (forming two groups: patients who were cared for by the configuration versus patients who were not cared for by the configuration) [9,10,23]. We further tested the extent to which the association between LOS and team configuration held against all patient characteristics (i.e., when stratified by five subgroups of patient characteristics): sex (males and females), age (< 65 and ≥ 65), rurality (resided in rural and non-rural areas), heart failure severity (≤49% and ≥50%), and overall health severity (a Van Walraven Elixhauser Comorbidity Score of ≤ 7 and a Van Walraven Elixhauser Comorbidity Score of > 7).

Two previously established and widely accepted thresholds were assessed to ensure the efficacy of each logistic regression model: the Chi-square omnibus test of model coefficients (assessing overall good model fit) and the area under the curve [6,10,11,23]. A statistically significant Chi-square omnibus test of model coefficients (p-value < 0.05) was used to determine if a regression model with team configuration as a predictor was able to predict the categories of the dependent variable (e.g., long LOS vs. short LOS) better than a regression model without the team configurations as included as a predictor. The discrimination ability of the model was evaluated using the area under the curve. Models with an area under the curve value greater than or equal to 0.70 had acceptable discrimination [11]. Subsequently, network density values assessed the extent to which LOS differed significantly among care teams with a larger proportion of a specific care role. Of the team configurations that were significantly associated with LOS after being stratified by age, sex, rurality, heart failure severity, and overall health severity, a Kruskal-Wallis H test assessed significant differences in the LOS, network density, and the number of care team members by role (e.g., the number of RNs on a team within a team configuration). Finally, a linear regression was used to identify the change in LOS associated with a one-unit change in RNs (i.e., adding or removing one RN) within the team configurations which were significantly associated with LOS after being stratified by patient characteristics.

Ethics statement

The study protocol (#276211) was reviewed and approved by the Ethics Committee of the University of Arkansas for Medical Sciences’ Institutional Review Board. This study was conducted in accordance with all applicable government regulations and institutional research policies and procedures at the University of Arkansas for Medical Sciences. The study team only had access to deidentified data on patients and clinicians. This study presented only minimal risk for loss of confidentiality and a waiver of documentation of the informed consent process was granted by the Ethics Committee, which is the standard for retrospective analyses involving deidentified data. While the Ethics Committee provided a waiver of the informed consent process for this retrospective analysis, the study team recognized the ongoing ethical dialogue regarding patient consent in secondary data use in the era of artificial intelligence. However, artificial intelligence was not used in this study. The Ethics Committee and the study team believed consenting the 2,274 African American patients and their 11,448 care team members would have substantially increased the study’s principal risk for loss of loss of confidentiality due to the need to give the study team access to identifying patient and clinician information solely for the consent process.

Summary of findings

Sixteen care team configurations were identified, of which all were associated with increased odds of having a long LOS in African Americans with heart failure. There were no configurations associated with a short LOS. Of the sixteen configurations, six were associated with a long LOS when stratified for age, sex, rurality, heart failure severity, and overall health severity. Within the six configurations, an increase and decrease in one additional RN on a care delivery team was associated with an increase and decrease in LOS of 8.4 hours (i.e., 504 minutes), respectively.

Implications for research and practice

The results indicate that the configuration of care delivery teams is associated with LOS.

This association implied that team configuration can be used to identify and track African American patients with heart failure who are at increased odds of having a long LOS and increased odds of in-hospital mortality, as evident by the significant associations with team configurations and Van Walraven Elixhauser Comorbidity Score, respectively. All sixteen of the care team configurations that were significant were associated with increased odds of having a long LOS. There were no configurations identified as being associated with a short LOS. This finding indicated that the sixteen configurations assessed were unlikely to be the optimal structure of heart failure care teams for African Americans due to their lack of association with short LOS. Configurations associated with short LOS would have required additional inquiry within experimental designs to narrow configurations for use in optimizing care. Therefore, additional combinations of care roles should be added to roles assessed within the sixteen configurations to explore their association with short LOS and stratified by indicators (e.g., ejection rate, shortness of breath) and pseudo-indicators (e.g., walking without a walker or wheelchair) of relatively good health at discharge [1,15].

Of the sixteen configurations, only six (“P+MD+RN+OT+PT”, “P+MD+RN+PCT”, “P+MD+RN+PCT+RT”, “P+MD+RN+PT”, “P+MD+RN+RD”, and “P+MD+RN+RT”) were associated with a long LOS in African Americans with heart failure when stratified by age, sex, rurality, heart failure severity, and overall health severity. Notably, RNs were the largest care role, and the most engaged care role found within the hospitalizations. Within the context of prior findings, this finding indicates that RNs are positioned to facilitate opportunities for increased inpatient intervention (i.e., heart failure-related education, medication adjustment, laboratory draws, engaging the cardiologists) if used as a new technique for identifying high-need patients in real-time [9,15,35,36]. However, avenues for realizing the real-time identification and management of care delivery team risk scores are needed.

A potential avenue for quantifying the effect of RNs and the teams they work within is social network analysis, as piloted here (i.e., overall network density and RN-specific network density). However, understanding the presence or absence of a specific care team role does not provide sufficient information for developing evidence-based team configurations that drive high-quality care delivery. Social network analysis metrics, including centrality, reciprocity, and transitivity, assess the importance of a team member, linkages, and the likelihood of similar care delivery behaviors among team members [37]. These metrics are critical to understanding how care delivery teams interact and how those interactions influence the outcomes of African American patients with heart failure while the outcomes are being influenced. These metrics can be fueled by metadata that is generally collected in electronic health record systems within the day-to-day workflow of care delivery, data such as who is on a patient’s team. Yet, a range of care delivery factors has historically influenced the LOS and mortality of African Americans, including nursing surveillance, expertise, patient workload, work environment, the quality of interactions with other team members, and clinical decision support. Future research in this domain should focus on providing a deeper understanding of the care delivery team data needed for high-quality nursing care, nurse-related information flow and sharing among care teams, and electronic health record-based communication networks of care delivery teams, of which social network analysis will provide the infrastructure needed to advance [38]. Employing this level of additional investigation will provide specificity on the universal definition of healthcare access, which is limited to a “workforce: [that are] capable, qualified, [and] culturally competent providers” [22]. These findings indicate a need to evolve the definition of healthcare access to include those care delivery team roles that are associated with high-quality care delivery in specific diseases and conditions. Models of care delivery that reduce LOS and mortality in African Americans can be developed to address the disproportionate impact of heart failure on African Americans.

Although the number of physicians, RNs, and the LOS were similar across the six team configurations that were consistently associated with long LOS, the teams differed in the percentage of RNs relative to other roles like physicians or physical therapists. Specifically, these differences were noticeable in how dense RNs were in the overall team’s network of interactions with patients. Some teams had a higher concentration of RNs relative to other roles (i.e., RN-specific network density), which changed how integrated RNs were within interprofessional teams. Notably, more complex patients may have required higher concentrations of RNs relative to other roles. Higher concentrations of RNs relative to other roles could have also occurred because of time. Our approach to additionally calculating network density by care role was innovative and could become a useful and an expanded component of operationalizing social network analysis, as there are no known prior studies that have sought to assess the influence of specific care roles with respect to the overall interprofessional mix of other care team roles to improve the design of care teams [39].

Among teams associated with a long LOS, an increase of one RN on a care delivery team was associated with an increase in LOS of 8.4 hours (i.e., 504 minutes). Inversely, a decrease in one RN on a care delivery team was associated with a decrease in LOS of 8.4 hours (i.e., 504 minutes). Generally, this relationship is explained as the more time a patient spends in the hospital, the more registered nurses would be on their teams. However, our results indicated that the number of RNs on a heart failure patient’s care team explained only 82.4% of the variation in long LOS. This indicates there was an estimated 17.6% of variation that could be due to other explanations.

While this finding advances the understanding of the influence of RNs on the LOS of African Americans with heart failure, neither the current finding nor prior studies provide sufficient information for making evidenced-based decisions about adding or removing a specific number of RNs from a patient’s team. Prior studies have been limited to surveys and retrospective approaches [40]. The interpretation of the association between RN engagement in care and LOS has been largely cross-sectional and limited by an inability to establish that association and/or potentially presume cause (i.e., RNs and their actions related to providing care) precedes the effect (long LOS). The findings indicate a strong potential for prospectively designing care teams to provide high-quality care with the minimum number of RNs needed to optimize LOS, a potential solution for local resource constraints and more global challenges such as the national nursing shortage [41,42]. For example, acute care settings could ensure the specific number of RNs on the care teams of African Americans and the evidenced-based frequency of tasks performed by RNs such as interval clinical assessment, continuing heart failure education, and interval medication adjustment assessment to reduce LOS. However, experimental designs must be employed to control for covariates and confounders when ascertaining the extent to which RNs, team configurations, and their tasks caused short and long LOS. The application of these designs will provide sufficient justification for prospectively designing teams. For example, patients would be provided care for by teams that are not only associated with positive outcomes but also caused them.