Study population
We conducted a retrospective cohort study using data from two large critical care databases: the Medical Information Mart for Intensive Care IV (MIMIC-IV) and eICU Collaborative Research Database (eICU-CRD). We selected these databases because of their comprehensive and diverse patient populations, which augment the generalizability of our findings to heart failure patients in intensive care settings across various regions. The MIMIC-IV database contains comprehensive clinical data from patients admitted to intensive care units at Beth Israel Deaconess Medical Center between 2008 and 201913. The eICU-CRD contains data from over 200 hospitals throughout the United States from 2014 to 201514. Access to these de-identified databases was granted to researchers who successfully completed the Collaborative Institutional Training Initiative (CITI) Program (certification numbers: 60071489 [Zhang] and 52219361 [Tang]). Given the de-identified nature of the data, informed consent and ethical approval requirements were waived. Data extraction was performed using Structured Query Language (SQL) with PostgreSQL (version 13.0) and Navicat software (version 16.0). For patients with multiple measurements of clinical parameters during hospitalization, only the initial values were included in the analysis. To ensure data accuracy and reliability, all variable extractions underwent independent verification by two researchers. This retrospective study was conducted in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines. The studies involving human participants were reviewed and approved by MIMIC-IV and eICU-CRD databases were approved by the institutional review boards of the Massachusetts Institute of Technology and Beth Israel Deaconess Medical Center. Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements. Informed consent was obtained from all subjects and/or their legal guardian(s). Due to the retrospective nature of the study, institutional review board waived the need of obtaining informed consent’ in the manuscript.
Data were extracted using a SQL server for patients fitting the study’s inclusion criteria: adults diagnosed with HF according to ICD-9 or ICD-10 codes (Supplementary Table 1). The initial dataset comprised a total of 31,855 heart failure patients, with 15,208 patients from the eICU-CRD database and 16,647 patients from the MIMIC-IV ICU database. The exclusion criteria applied were patients aged under 18 years, those with ICU stays of less than 24 h, and individuals with fewer than three blood pressure measurements. From the eICU-CRD database, 2 patients were excluded due to age, 3,006 due to ICU stay duration, and 356 due to insufficient BP measurements, resulting in 11,844 eligible patients. Similarly, from the MIMIC-IV ICU database, 2,803 patients were excluded due to ICU stay duration and 97 due to insufficient BP measurements, leaving 13,747 eligible patients. In total, 25,591 patients were included in the final analysis after combining the eligible patients from both databases. This selection process ensured that the analysis focused on a specific group of individuals relevant to the research objectives (Fig. 1). Missing data were handled using multiple imputations, and one randomly selected imputed dataset was used for analysis, with detailed missing data patterns provided in Supplementary Table 2.
Flowchart of patient selection. eICU-CRD, eICU Collaborative Research Database; ICU, intensive care unit; MIMIC-IV, medical information mart for intensive care IV; BP, blood pressure.
Assessment of blood pressure variability
Blood pressure measurements were extracted from each database using structured query language. For each patient, we calculated 24-h BPV from the first ICU admission using the standard deviation (SD) of all recorded values. The coefficient of variation (CV) was calculated as SD divided by mean blood pressure5. These two metrics were chosen due to their clinical relevance and ability to measure both absolute and relative variability across patients with differing baseline pressures. We assessed systolic blood pressure variability (SBPV), diastolic blood pressure variability (DBPV), and blood pressure variability (MBPV) using invasive arterial line measurements recorded during the first 24 h of ICU admission. These accurate and real-time measurements allowed us to evaluate BPV’s impact on patient outcomes.
Additionally, a detailed breakdown of the frequency of blood pressure measurements is presented in Supplementary Table 3, which includes the first values, mean values, SD, variability, and frequency across the eICU-CRD and MIMIC-IV databases.
Covariates
We collected comprehensive baseline characteristics including: (1) demographic data: age, sex, and ethnicity (White or Other); (2) vital signs: body mass index (BMI), heart rate, SBP, DBP, and MAP; (3) comorbidities: hypertension, diabetes, myocardial infarction, atrial fibrillation, stroke, chronic obstructive pulmonary disease (COPD), renal failure, and cancer; (4) laboratory tests within the first 24 h of ICU admission: hemoglobin, white blood cell count (WBC), platelet count, creatinine, blood urea nitrogen (BUN), potassium, sodium, and chloride; (5) treatments: angiotensin converting enzyme inhibitor/angiotensin receptor blockers (ACEI/ARB), beta-blockers, calcium channel blockers (CCB), diuretics, vasoactive agents, hemodialysis, and mechanical ventilation.
Outcomes
The primary outcomes were in-hospital mortality and 30-day all-cause mortality. In-hospital mortality was defined as death occurring during the index hospitalization. Thirty-day mortality was defined as death occurring within 30 days after ICU admission. Both outcomes were assessed separately in the eICU-CRD and MIMIC-IV cohorts. The mortality data were extracted directly from the databases, which maintain comprehensive patient outcome records13,14. For patients with multiple ICU admissions, only the outcomes from the first admission were considered to avoid potential confounding from repeated measurements.
Statistical analysis
All continuous and categorical variables at baseline were presented as means (SD) and percentages. A Chi-square test or independent sample t-test was used to examine the baseline differences between included and excluded participants. Descriptive statistics summarized the baseline characteristics of participants. BPV measures (SBPV, DBPV, and MBPV) were standardized using z-score transformation to facilitate comparison and interpretation.
We constructed four sequential models to analyze the association between blood pressure variability and mortality. Model 1 was unadjusted. Model 2 adjusted for demographic characteristics (age, gender, ethnicity). Model 3 adjusted for Model 2 covariates plus vital signs (BMI, heart rate) and comorbidities (hypertension, diabetes, myocardial infarction, atrial fibrillation, stroke, COPD, renal failure, cancer). Model 4 was fully adjusted by adding laboratory parameters (hemoglobin, WBC, platelet, creatinine, BUN, potassium, sodium, chloride) and treatments (ACEI/ARB, beta-blockers, CCB, diuretics, vasoactive agents, hemodialysis, mechanical ventilation) to Model 3.
Logistic regression was used to examine the association between BPV and in-hospital mortality, while Cox proportional hazards regression was employed to analyze the relationship with 30-day mortality. These regression methods were selected as they account for potential confounders and allow for precise estimation of associations between BPV and mortality. In the Cox regression analysis, patients who were lost to follow-up were censored at their last known status. To visualize survival differences, we constructed Kaplan–Meier survival curves stratified by quartiles of BPV measures, with the log-rank test used to assess statistical differences between these curves.
Stratified analyses were performed to evaluate potential effect modifications by key characteristics, including database source (eICU-CRD vs MIMIC-IV), age (≤ 65 vs > 65 years), gender (female vs male), race (White vs Other), BMI (< 25 vs ≥ 25 kg/m2), comorbidities (hypertension, myocardial infarction, stroke, renal failure), and treatments (ACEI/ARB, CCB, vasoactive agents, hemodialysis). Within each stratum, we calculated odds ratios (ORs) with 95% confidence intervals (CIs) using the fully adjusted Model 4. Interaction tests were performed to assess the presence of effect modification, with a P-value < 0.05 indicating significant interaction. Results were presented using forest plots to visualize consistency of associations across strata.
To assess the robustness of our findings, we conducted sensitivity analyses by excluding patients with missing data. The primary analysis utilized all available data, while sensitivity analysis was restricted to patients with complete data for all variables. We repeated the primary analyses, including fully adjusted models (Model 4) and stratified analyses, in this complete-case cohort.
All analyses were performed using R Statistical Software (Version 4.2.2, The R Foundation) and Free Statistics analysis platform (Version 2.0, Beijing, China, Statistical significance was set at p < 0.05 for all tests.
link