Overall and linked blood pressure variabilities in the first 24 hours and mortality after spontaneous intracerebral hemorrhage: a retrospective study of 1,036 patients
Article information
Abstract
Background
This study aims to establish the individual contributions of blood pressure variability (BPV) indexes, categorized into overall and linked variability, to mortality following intracerebral hemorrhage (ICH) by examining the risk factors.
Methods
Patients with spontaneous ICH (n = 1,036) were identified with valid blood pressures (BP) from the first 24-h systolic BP records in the Medical Information Mart for Intensive Care IV version 2.2 database (MIMIC IV). Information on the baseline characteristics, including age, sex, initial Glasgow Coma Scale (GCS) and National Institutes of Health Stroke Scale (NIHSS) scores, ICH location, Charlson comorbidity index score, and presence of diabetes with or without complications, were collected. Three indexes of BPV—range, standard deviation (SD), and generalized BPV (GBPV)—were calculated using the first 24-h systolic BPs. An automated stepwise variable-selection procedure was used to develop the final logistic model for predicting in-hospital mortality.
Results
Out of 1,036 patients, 802 (77.4%) survived and were discharged after spontaneous ICH. Factors associated with mortality included age; male sex; ICH in the brainstem, ventricle, or multiple locations; low GCS score (< 9); high NIHSS score (> 20); and diabetes with complications. Mean systolic BP, SD, and GBPV were also linked to mortality. Higher GBPV notably increased the risk of in-hospital death, with an odds ratio of 3.21 (95% confidence interval, 2.10 to 4.97) for every + 10 mmHg/h change in GBPV.
Conclusions
This study underscores the additional impact of GBPV, herein linked to BPV, on mortality following ICH, providing further insights into the management of blood pressure in the early stages of ICH treatment.
INTRODUCTION
Intracerebral hemorrhage (ICH) is a severe subtype of stroke that significantly contributes to global morbidity and mortality. Despite significant advancements in medical and surgical interventions, ICH remains a major cause of worldwide morbidity and mortality. Prompt management of blood pressure is crucial to prevent rebleeding and secondary brain damage during ICH treatment [1,2]. Blood pressure variability (BPV), which reflects fluctuations in blood pressure readings over time, poses an additional risk factor that is associated with poor outcomes in ICH patients. Recent studies [3,4] have further demonstrated a link between BPV and increased mortality in individuals suffering from ICH.
However, the exploration of BPV has been limited due to the lack of consensus on a representative value that captures the variability among clinically accessible measurements. The BPV indexes are divided into two primary categories based on their inherent characteristics. The first category includes range and standard deviation (SD), which serve as indicators of the overall variations observed over a specific timeframe. These variables are classified as representative of overall BPV. Conversely, the second category captures variations between consecutive measurements, such as successive variation (SV) [3], generalized BPV (GBPV) [5], and functional successive variation (FSV) [3]. These variables are classified as “linked” BPV, focusing on the interconnectedness of adjacent measurements rather than the overall variation. Among them, GBPV is the most intuitive to interpret.
The Medical Information Mart for Intensive Care IV (MIMIC-IV) database provides the high-density vital records required for our study, capturing extensive patient data from intensive care units. We aimed to determine whether and how much linked BPV, derived from the high-density intensive care unit (ICU) blood pressure recordings within the first 24 h, possesses additional explanatory power compared to overall BPV indexes in predicting a fatal outcome in patients diagnosed with ICH.
MATERIALS AND METHODS
Cohort characterization
The MIMIC-IV database (version 2.2) was used for this retrospective study [6,7], containing healthcare information of over 40,000 patients admitted to the ICU of Beth Israel Deaconess Medical Center from 2008 to 2019. To obtain rights and permission to use the database, all authors completed the Collaborative Institutional Training Initiative Program following the protocol required by the Massachusetts Institutes of Technology.
Using nine International Classification of Diseases 10th Revision codes to identify patients with spontaneous ICH, we extracted data on 1,359 adult patients (> 18 years) who had at least one of these codes indicating spontaneous ICH in their diagnoses (Table 1).
ICD-10 Code for Extracting Cohort of Spontaneous Intracerebral Hemorrhage
Severity and comorbidity grading
For the risk stratification of patients with ICH, general information was collected, including age groups, sex, presence of diabetes with or without complications, Charlson Comorbidity Index (CCI) score, National Institute of Health Stroke Scale (NIHSS) score, Glasgow Coma Scale (GCS) score at arrival, and the location of the ICH of interest (cerebellar, in the brainstem, with intraventricular hemorrhage, and multifocality).
Blood pressure and calculation and exploration of BPV indexes
Among 164,332 BP measurements, 163,198 were valid BP measurements from 1,036 ICH patients. Valid BP measurements satisfy two criteria: (1) systolic blood pressure (SBP) - DBP ≥ 5 mmHg, and (2) more than 3 records during the first 24 h.
Analyzing 24-h SBP recordings, we derived several variables from the SBP data, including the occurrence (1 or 0) of maximum SBP >160 mmHg and minimum SBP < 140 mmHg, cumulative duration (h) with SBP > 160 mmHg or < 140 mmHg, and the 24-h mean SBP. Additionally, we calculated the SD (mmHg), SBP range (maximum–minimum, mmHg) as overall BPV indexes, and GBPV (mmHg/h) as linked BPV indexes. GBPV calculations followed the method described by Mascha et al. [5].
In the formula, ‘i’ represents the order of blood pressure measurement, and ‘n’ the total number of measurements during a single 24-h period. We exploratively analyzed the correlation structures behind the 24-h mean SBP and its derivatives—SD, range, and GBPV—using principal component analysis.
Main analysis and sensitivity analyses
The primary research question was whether in-hospital mortality is influenced by SD and GBPV after adjusting for patient characteristics and general SBP derivatives following an ICH event. For clarity, numeric characteristics were categorized into age groups, CCI > 3, NIHSS groups (< 6, 6–13, 14–20, 20 <), and GCS groups (< 9, 9–13, 13 <). Associations between these characteristics and mortality were estimated as ORs before and after applying a stepwise variable-selection technique.
Three approaches were designed for the sensitivity analyses. Initially, to prevent overfitting [8], 1,000 bootstrapping iterations (resampling size, n = 315) were performed, generating 1,000 odds ratios (ORs) from regression analyses. This method allowed for robust summarization of ORs using nonparametric statistics, including median OR and the 2.5th and 97.5th percentiles. The resampling size was set as 315, calculated by multiplying 15 by 21, the number of predictors including categorical variables from the main analysis. To control for potential bias from an unmeasured strong confounder, we generated an inverse propensity weight for GBPV changes and stabilized it using marginal densities as numerators and conditional densities as denominators. The association between GBPV and mortality was then reassessed using the logistic regression model weighted by the inverse propensity weight. Finally, after identifying outliers in GBPV measurements—defined as values beyond Q1 minus 1.5 times the interquartile range (IQR) and Q3 plus 1.5 times IQR—we recalculated the OR using the dataset excluding these outliers (n = 1,003).
Presentation and employed tools
To clarify the primary outcome, we first obtained unadjusted ORs for in-hospital mortality using separate logistic regressions for each candidate predictor. Subsequently, the adjusted ORs for the final selection of predictors, including mean SBP, SD, and GBPV, were calculated through a penalty-based stepwise variable selection process using Akaike’s information criterion as a penalty measure. This process was automated using the MASS package (Venables WN and Ripley BD, Modern Applied Statistics with S. 2002). The ORs were presented along with their 95% confidence intervals [CIs]). All records related to the condition were compiled from the database.
After constructing the relevant dataset from Google BigQuery (Google. (n.d.). BigQuery. Retrieved Dec 23, 2023, from https://cloud.google.com/bigquery/), we utilized R software version 4.2.3 (R: A language and environment for statistical computing. R Foundation for Statistical Computing) for manipulation and analysis. Apart from those derived from the bootstrapping procedure, all displayed OR values were manually estimated via log-odds conversion from the logistic regression output. For ease of interpretation, we presented the OR for SD and GBPV per every + 10 mmHg/h change. Our statistical analysis primarily focused on the 95% CI to ensure robust inference, omitting P values in this context.
RESULTS
Among 1,036 patients, 946 had 24 or more SBP measurements within the first 24 h, 39 had fewer than 12, and 51 had between 12 and 23 measurements. In the cohort, 77.4% (802 out of 1,036) of patients survived and were discharged following spontaneous ICH. The mean duration of hospitalization was 12.1 days, with a median of 8.0 days (Q1–Q3 range: 4.0–15.0 days). The average GBPV for the first 24 h was 11.5 mmHg/h, median 10.5 mmHg/h, ranging from Q1 at 8.0 mmHg/h to Q3 at 13.2 mmHg/h, showing that the central 50% of the population fell within a 5.3 mmHg/h interquartile range.
Except for the initial 24-h mean SBP, derivatives such as range, SD, and GBPV were lower in patients who survived (Fig. 1). From principal component analysis, the most influential factor on the first principal component was SD and range, on the second principal component it was influenced by mean and inversely by GBPV, and on the third principal component it was influenced by GBPV and inversely by range (and SD), highlighting a complex correlation structure (Table 2).
Summary plots for four SBP derivatives: (A) mean systolic BP, (B) range of SBP, (C) standard deviation of SBP, and (D) generalized BP variability. Each boxplot includes a box (Q1 and Q3), a median line (Q2), whiskers (lines extending to the farthest points within +/- 1.5 times the interquartile range from the box's edges), and outliers. Note that the Y-axis of subplot (D) is transformed using the common logarithm. SBP: systolic blood pressure, GBPV: generalized blood pressure variability.
Relationship of BP and Its 3 Derivatives Explained Through Principal Component Analysis
Variables associated with in-hospital mortality were presented as unadjusted OR and its CI (Table 3). Significant factors included older age, certain brain locations (brainstem and IVH), low GCS score, and high NIHSS score. After a stepwise selection process, regardless of separate statistical significances, 13 variables remained in the equation. These included older age groups (65–83 and > 84), specific ICH locations (brainstem, IVH, and multifocality), low GCS (< 9), NIHSS scores (6–13, > 20), DM with complications, and 4 BP derivatives: mean SBP, cumulative duration of SBP > 160 mmHg, SD, and GBPV. GBPV still showed a strong association with high mortality, with an OR of 3.21 (95% CI, 2.10 to 4.97) for every change by + 10 mmHg/h, and an OR of 3.45 (95% CI, 2.14 to 5.65) for SD (Table 4). The ORs of GBPV in the sensitivity analyses were summarized in Table 5, further validating the robustness of our primary findings.
Predictors and Unadjusted ORs for In-hospital Mortality in 1,036 Patients with Spontaneous ICH
Final Selection of 13 Predictors for In-hospital Mortality with Associated ORs in 1,036 Patients with Spontaneous ICH
Comparisons of ORs of GBPV from the Main and Sensitivity Analyses
DISCUSSION
The odds of in-hospital death increased with rising GBPV, showing a more than threefold increase in odds for each incremental +10 mmHg/h increase in GBPV, alongside the effect of SD. Considering that the central 50% of our study participants had a GBPV variation around 5.3 mmHg/h, such an increase in odds indicates a notable impact on outcomes. These findings underscore that elevated GBPV acts as an independent predictor of mortality in ICH patients. This data suggest that GBPV could serve as a viable clinical tool for risk stratification and guiding treatment decisions in ICH patients.
The association between BPV and adverse outcomes in ICH patients has been previously reported in several studies [3,4,9-11]. However, previous studies focused on indices such as SD [12] or the coefficient of variation [13], which are unaffected by the sequence of blood pressure readings. SV and FSV have been explored by Divani et al. [3], but these were not analyzed in relation to the type of BPV. Compared to FSV, GBPV is a relatively new metric in ICU medicine and is simpler to calculate, as it does not require mathematical integration, thus operating in a more intuitive unit, mmHg/time. Our results show that GBPV had a low variance inflation factor (Table 3), suggesting its potential to act independently from other predictors. Our study extends the existing literature by highlighting GBPV's usefulness as a linked BPV index in predicting poor outcomes in patients with spontaneous ICH.
Our findings align with previous studies indicating that factors such as age [14], male sex [15], ICH location [14,16], lower GCS score [14], and higher NIHSS score [16] are associated with poor outcomes in ICH patients. It is important to note that elevated mean SBP was inversely associated with mortality. This seemingly paradoxical observation can be explained by the consistent association between high blood pressure and high BPV. Therefore, in logistic regression analysis estimating the effect of BPV, the impact of high BP was inherently integrated into the effect of high BPV. After variable selection, only the effect of low BP on mortality remained.
Several limitations need to be addressed. Firstly, the size of the intracerebral hemorrhage (ICH), a significant risk factor considered by neuro-intensivists [17,18], was not included because it was not documented in our database. Secondly, our retrospective study design may not fully capture the complete clinical picture of patients with ICH. Specifically, due to the lack of temporal data in the database about the timing of ICH events, our blood pressure recordings during the first 24 h may not accurately represent the initial post-ICH period. This limitation could significantly affect the causal relationships we aim to infer. The absence of temporal data on surgical procedures in the MIMIC database also poses a limitation. To prevent confusion in interpreting the analysis results, surgical-related information was excluded. Thirdly, conducting the study at a single center using a database might limit its generalizability to broader populations; however, the robustness of our findings, supported by sensitivity analyses, may help alleviate some of these concerns.
In summary, our study demonstrates that beyond the influence of overall BPV, GBPV, the representative measure we used for linked BPV, is associated with increased mortality in ICH patients. GBPV shows promise as a valuable clinical tool for risk stratification and decision-making in ICH management. Prospective studies are needed to confirm these findings and establish a therapeutic threshold for GBPV in ICH management.
Notes
FUNDING
This research was supported by Dongguk University Research Fund 2024.
CONFLICTS OF INTEREST
No potential conflict of interest relevant to this article was reported.
DATA AVAILABILITY STATEMENT
To obtain rights and permission to use the dataset extracted and analyzed during the current study, readers should complete the Collaborative Institutional Training Initiative Program following the protocol required by the Massachusetts Institutes of Technology.
AUTHOR CONTRIBUTIONS
Writing - original draft: Hangyul Cho, Taehoon Kim, Younsuk Lee. Writing - review & editing: Younsuk Lee, Dawoon Kim, Hansu Bae. Conceptualization: Younsuk Lee. Data curation: Younsuk Lee. Formal analysis: Younsuk Lee. Methodology: Younsuk Lee. Project administration: Younsuk Lee. Funding acquisition: Younsuk Lee. Visualization: Younsuk Lee. Investigation: Younsuk Lee. Resources: Younsuk Lee. Supervision: Younsuk Lee. Validation: Younsuk Lee.