Association between High Cumulative Numbers of Elevated Heart Rate and Mortality in Neurological ICU patients: Retrospective Analysis of eICU Collaborative Research Database


 Background

Heart rate is routinely measured in Neurological intensive care unit(NICU), but its prognostic value remains debated. We sought to evaluate the association of high cumulative numbers of elevated Heart Rate (HcneHR) with mortality in NICU patients.

Methods

We used a large observational eICU Collaborative Research Database (eICU-CRD), where continous heart rate monitoring every 5 minute was available. We collected periodic heart rate, disease severity (APACHE IV score), NICU and hospital mortality and other information in 8347 patient admissions from the eICU-CRD. The cumulative numbers of Heart Rate (cneHR) were defined as >100 beats/min in first admittion 24 hours, and if cneHR ≥10,then was defined as higt cneHR(HcneHR). The primary outcome was NICU mortality. The other outcomes were hospital mortality, length of NICU stay and APACHE IV score. Multivariable logistic regression was used to assess for association for HcneHR and other covariance with NICU and hospiltal discharge status.

Results

The mean age of patients were 63 years, and the most frequent disease categories of NICU in eICU-CRD were postoperation (25%), stroke(19%), traumatic brain injury(14%). The mean APACHE IV score was 50. Overall NICU mortality of the cohort at discharge was 4%, and hospital mortality was 8%. The NICU mortality of HcneHR patients was 7%. Adjusted logistic regression for HcneHR showed a significantly increased risk of NICU death with odds ratio 1.61(confidence interval, 1.26-2.06; P <0 .001).

Conclusions

In adult neurocritically ill patients, we found a significant association for HcneHR with elevated mortality and several others important patient-centered outcomes.


Background
The burden of critical illness is higher than generally appreciated and will increase as the population ages, preventive and therapeutic interventions that are generalisable across ICUs are needed [1]. Heart rate is routinely measured in ICU but its prognostic value remains debated. It is well known that critically ill patients are prone to develop tachycardia due to various conditions such as anxiety, pain, fever, infection, hypovolemia, exaggerated sympathetic activation, heart failure, hypoxia, anemia, drug effects, primary arrhythmia and so on [2][3][4]. It has been shown that there is an association for high heart rate(HR) and poor outcomes in critically ill patients [5], and HR is an important part of most ICU prognosis scores [6].
The critical care unit is a data intensive environment [7]. However, as yet, heart rate from bedside monitor in critical care is evaluated by clinicians almost in much the same way as before, little information is available on the critical duration of tachycardia, especially in neurological intensive care unit(NICU), which focuses on the optimal management of acutely ill patients with life threatening neurologic and neurosurgical disease or with life-threatening neurologic manifestations of systemic disease [8]. We may get more information on the continuous heart rate monitoring.
Therefore, in this study, we hypothesized that an elevated HR of >100 beats/min for a prolonged period, where we showed it as high cumulative numbers of elevated HR(HcneHR) during the rst day of NICU stay, may be associated with decreased survival in neurocritically ill patients. This may help clinicians to better identify high risk patients admitted in NICU.

Page 3/24
Methods Setting This study used data stored in the high-resolution database, the eICU-CRD(eicu-crd.mit.edu), which comprises 200,859 patient unit encounters for 139,367 unique patients admitted between 2014 and 2015. The database contains parameters that were available in the ICU clinical information system, including vital signs, laboratory measurements, medications, APACHE components, care plan information, admission diagnosis, patient history, time-stamped diagnoses and similarly chosen treatments. The elaborate description of eICU-CRD is available elsewhere [9].
The eICU-CRD was exempt from institutional review board(IRB) approval due to the retrospective design, lack of direct patient intervention, and the security schema, for which the re-identi cation risk was certi ed as meeting safe harbor standards by Privacert (Cambridge, MA) (Health Insurance Portability and Accountability Act Certi cation no. 1031219-2). Due to the HIPAA compliant de-identi cation in this database, our institutional IRB requirement was waived. After completing a National Institutes of Health(NIH) web-based training course(Protecting Human Research Participants), the author(dw zhou, certi cation number: 28795067) was approved to access to the database for research aims.

Study Population
All patients in the eICU-CRD were eligible for inclusion in the present investigation. As for those who admitted in NICU for more than once, we considered them as different patients. We selected all adults patients admitted in NICU with length of stay >24 hours, and excluded those who used heart rate controlling drug during the rst day in NICU. Heart rate controlling drug contained β-blocker and calcium antagonist. We collected all periodically recorded heart rates for the rst 24hours of NICU admission, and patients without periodical heart rate data were excluded. Patients without APACHE IV score, pulse oxygen saturation(SpO2), admission temperature, admission blood pressure, gender information were excluded.

Clinical Variables and Outcomes
Data on the following information during the rst 24 hours of admission were extracted: age, gender, ethnicity (caucasian, african american, hispanic, asian, ative american, other/unknown), admission disease categories (postoperation, ischemic stroke, traumatic brain injury, intracranial hemorrhage, seizures, sepsis, subarachnoid hemorrhage, coma, others), admission vital sign(temperature, respiratory rate, heart rate, mean blood pressure, SpO2 ), heart rate related indexes for continuously periodical monitoring(cneHR, HcneHR, maximum heart rate, minimum heart rate, average heart rate), acute physiology score, APACHE IV score, the rst day hemoglobin, comorbidities, including atrial brillation, cardiovascular diseases, tracheal intubation, mechanical ventilation, vasopressors, length of NICU stay, the NICU and hospital discharge status(alive or expired). All patients admitted from the operating room or recovery room had a surgical diagnosis, even though some surgical patients had a medical reason for admission to the NICU.
Periodical data refers to data which is consistently interfaced from bedside vital signs monitors into eCareManager. Continuously measured heart rate was originally collected at 1-minute intervals, with 5minute medians archived in eICU-CRD. That is to say, every patient in the rst day of NICU had 288 heart rate values. The cneHR was de ned as cumulative numbers of heart rate >100 beats/min in rst admission 24 hours archived in eICU-CRD "vitalPeriodic" table, and if cneHR (10,then was de ned high cneHR(HcneHR), and cneHR<10 was de ned as Low cneHR(LcneHR).
The primary endpoint of our study was NICU mortality, de ned as death observed during NICU stay. Other study endpoints included hospital mortality, NICU length of stay and APACHE IV score. NICU length of stay was de ned as the difference between date of NICU admission and discharge.

Statistical Analysis
Continuous variables are shown as mean (standard deviation) or median (interquartile range), and were analyzed with Wilcoxon rank-sum tests. For dataset with a large sample size, the Anderson-Darling test can be very sensitive to a small deviation from the normal distribution. So we switched off the normality testing and still used mean and standard deviation to describe the data [10]. Categorical variables were reported as numbers and percentages, and were analyzed with chi-square test. As for missing data onto the rst day hemoglobin, we use the mean hemoglobin of the NICU alive patients to imputed the missing data onto the NICU alive patients, so was for the NICU expired patients. Some patients did not get the heart rate monitoring immediately after admission to NICU, the missing heart rate data were replaced using next observation carried backward(NOCB) method.
Receiver operating characteristic (ROC) analysis was used to evaluate the predictive ability of cneHR. The best cut-off value was determined using the Youden index, and was used to de ned HcneHR. Uadjusted crude outcomes including NICU mortality, hospital mortality, NICU length of stay and APACHE IV score were compared between HcneHR and LcneHR patients.
Logistic regression model was tted to adjust to odds ratio of NICU and hospital mortality. Three models were created: model 1 adjusted to APACHE IV score; model 2 adjusted for demographics and APACHE IV score; model 3 adjusted for model 2 covariance plus disease categories, admission vital sign( temperature, respiratory rate, heart rate, mean blood pressure,SpO2 ), the rst day hemoglobin, use of mechanical ventilation, use of vasopressors, length of NICU stay, comorbidities, including atrial brillation, cardiovascular diseases. A stepwise backward elimination method with a signi cance level of 0.05 was used to build the nal model.
We determined the predictive e cacy of HcneHR in conjunction with currently accepted known risk factors for mortality. Here wo chosen APACHE IV score, use of mechanical ventilation, HcneHR and use of vasopressors to construct prediction model for NICU expired, and showed it in nomogram with R package "rms". Each variable was represented by a bar. A given value of a variable can be mapped to the point bar at the top of the graph and there is a point value for that given value. After each variable was assigned to a point number, they were summed and mapped to the total point bar. Then there will be a value in the "Risk of NICU Expired" bar corresponding to those total points. ROC analysis was used to evaluate the predictive ability of the predictive model. For disease subgroup analysis, we used model1 to calculate odds ratio of NICU and hospital mortality for different diseases, and showed them with forest graph.
Data extraction was performed using PostgreSQL (version 10.5, www.postgresql.org) and pgADmin PostgreSQL tools(version 4). R (version 3.5.1, www.r-project.org) was used for statistical analysis. A twosided P value of <0.05 was considered statistically signi cant.

Patient Characteristics
The eICU-CRD contained 200,839 patients, including 14,451 NICU patients. Of the remaining 14,451 patients, we excluded those with length of stay (24h, age<18, and those who used heart rate controlling drug. A total of 8,347 patients met our inclusion criteria after excluding those with missing data (Fig 1), of which 339 were expired when discharged from NICU, giving an NICU mortality rate of 4%. Demographics and baseline characteristics between alive and expired patients are presented in Table 1. The mean age were 63 years, and most patients ethnicity were "Caucasian" (75%). The main disease categories were postoperative patients (25%) , stroke(19%), traumatic brain injury(TBI)(14%) , intracranial hemorrhage(10%), and so on. The mean APACHE IV score was 50. On the rst 24hours, cneHR was signi cantly higher in NICU expired patients than in alive patients(72.5±91.58 vs. 34.26±65.39; p<0.001), also in hospital expired and alive patients(71.41±90.8 vs. 50.52±73.18; p<0.001)( supplementary g.1). The number of missing data onto hemoglobin was 1127(13.5%).
HcneHR and the Outcome Analysis Table 2 showed crude outcomes by cneHR category. Patients with HcneHR had signi cantly higher NICU and hospital mortality. ROC analysis showed area under curve(AUC) of the predictive ability of cneHR was 0.655, and the best cut-off point was 9.5(supplementary g.2), that is why we chosen 10 as the cutoff point of HcneHR and LcneHR. The percentage of HcneHR in NICU expired patients was 61%, while the alive one was 36%. Fig.2 showed the different percentage of HcneHR between NICU alive and expired patients over time. The length of NICU stay and APACHE IV score were signi cantly different between the HcneHR and LcneHR patients (table 2). Table 3 showed the ORs of NICU and hospital mortality in the different logistic regression models. The unadjusted OR of HcneHR for NICU mortality was 2.84(2.27-3.55, p<0.001 ). After adjustment for demographics , APACHE IV score and other clinical parameters, the OR decreased but remained statistically signi cant for HcneHR patients. As for subgroup analysis of different diseases, in a univariable logistic regression, HcneHR signi cantly associated with NICU mortality in postoperation, ischemic stroke, traumatic brain injury, intracranial hemorrhage and others groups. Furthermore, after APACHE IV score was adjusted to a multivariable logistic regression analysis, HcneHR remained signi cantly associated with NICU mortality in ischemic stroke, intracranial hemorrhage and others groups( g. 3), and signi cantly associated with hospital mortality in traumatic brain injury, intracranial hemorrhage and others groups (supplementary g.3). In a post hoc analysis, we explored the ORs of different heart rate and corresponding cneHR. We calculated the adjusted odds ratio of death using a minimally adjusted logistic regression model (including age, gender, and ethnicity) for each combination of cumulative numbers and heart rate thresholds. The results were shown in Figure 5, where each square showed the corresponding odds ratio of NICU death numerically. There was an increasing trend of OR as the heart rate and cneHR increased.

Discussion
In this analysis of a large and heterogeneous NICU clinical database, we found a signi cant association for HcneHR with elevated NICU mortality, hospital mortality, ICU length of stay and APACHE IV score. As cneHR can be readily got from the ECG monitoring system, which is common in NICU, it may be a good tool to better identify high risk patients admitted in NICU. In contrast to other heart rate parameters, cneHR integrates many parameters, like elevated heart rate, volatility and duration.
Elevated resting heart rate (RHR) is associated with all cause mortality in the general population [11,12], and lower RHR is related to lower all-cause mortality [13]. Elevated preoperative HR is associated with impaired cardiopulmonary performance consistent with clinically unsuspected, subclinical cardiac failure, and is associated with myocardial injury, myocardial infarction, and mortality after non-cardiac surgery [14,15]. In critically ill patients, prolonged elevated HR increased incidence of major cardiac events and was associated with a signi cantly longer ICU stay and reduced survival [5,16]. Our ndings are in line with these studies.That is to say, in critically ill patients admitted in NICU, HcneHR was associated with NICU and hospital mortality. In the time series characteristics of HcneHR percentage, the NICU alive patients had a gradually downward trend, and tended to be stable after 12 hours. However, the HcneHR percentage of NICU expired patients was higher, and jumped over time (Fig.2).
The disease pro le of our study contained surgical diagnosis, ischemic stroke, traumatic brain injury,intracranial hemorrhage, seizures, sepsis, subarachnoid hemorrhage, coma, and others groups. There are a number of studies which have evaluated the association for increased HR and mortality in neurocritical patients. Preoperative elevated heart rate is associated with mortality after neurosurgery [15].
In moderate and severe TBI patients, a correlation between elevated baseline HR and 60th-day mortality was shown [17]. After SAH, prolonged elevated HR was associated with 3-month poor outcome and an increased hazard for delayed cerebral ischemia, myocardial injury and pulmonary edema [18]. Among intracerebral hemorrhage and ischemic stroke patients, higher admission HR was independently associated with death and poor functional outcome [19,20]. Elevated HR prior to partial seizure onset of those attacks which become secondarily generalized compared to seizures which remain localized, and this may be relevant to the understanding of sudden death in epilepsy [21]. In our study, subgroup analyses showed HcneHR were associated with NICU mortality in ischemic stroke, intracranial hemorrhage and others groups, and signi cantly associated with hospital mortality in traumatic brain injury, intracranial hemorrhage and others groups.
There are myriad causes of elevated HR in neurological critically-ill patients, however, including but not limited to pain, agitation, hypovolemia, infection, fever and systemic in ammation, or vasoactive therapy with vasopressors,and multiple potential causes may coexist, making it di cult or impossible to get a precise cause [2][3][4]. Elevated HR can increase metabolic oxygen demand, increase blood pressure from increased sympathetic activity, activate in ammatory pathway, or cause endothelial dysfunction [22,23]. The precise mechanism between elevated HR and poor outcome still remain di cult to unravel.
This was an observational analysis and therefore we cannot draw causal conclusions, and other unknown confounders may in uence the result. It is of paramount importance to analyze covariance in a retrospective cohort study. We attempted to include a wide range of clinically important confounders to reach valid results and performed PSM to account for group differences.We used the multivariable logistic regression to adjust several clinically important confounders, like age, gender, ethnicity, temperature, mean blood pressure, SpO2, mechanical ventilation use, vasopressors use, comorbid heart disease, admission hemoglobin and so on. This association was maintained after adjusting to these clinically important confounders. APACHE IV scoring system had good discriminatory capability for critical patients [24,25]. In the multivariable logistic regression analysis, we found that HcneHR was associated with NICU and hospital mortality after APACHE IV score was adjusted, indicating that HcneHR may provide prognostic information independent of APACHE IV score. The increasing trend of OR as the heart rate and cneHR increased may suggested dose-response relationship.The nding can help with prognosis prediction and raise the question about a causal relationship.
Our ndings posed the question as to whether therapeutic reduction of HR is bene cial to neurocritical patients. For patients in septic shock, β-blocker was associated with reductions in heart rates to achieve target levels, without increased adverse events, and with a lower twenty-eight day mortality [26]. A metaanalysis of nine observational studies showed in adults with acute TBI, exposure to β-blockers after TBI was associated with a reduction of in-hospital mortality (pooled OR 0.39, 95%CI: 0.27-0.56; I2=65%, p<0.00001) [27]. There is a growing body of evidence showing that perioperative β-blocker exposure was associated with lower rates of 30-day all-cause mortality in cardiac high-risk patients undergoing noncardiac surgery [28]. Nevertheless, the right timeframe for intervention and the appropriate heart rate goals are currently unde ned. Further studies are required to investigate strategies for HR control in NICU patients. Appropriately randomized controlled multi center trials are required to con rm these questions.
The neurocritical care unit is a data intensive environment. Utilization of these dense data onto real time decision making for patient care represents the art of neurocritical care practice [7]. Despite advances in medicine, data onto vital signs of bedside monitors in neurocritical care are evaluated by clinicians almost in much the same way as before. In one study of NICU patients, the investigator rather than just displayed the maximum daily temperature as the measure of fever, calculated the area under the curve (AUC) above a speci c cut off (such as >38.5 °C) and provided a more robust measure of 'dose' of fever [29]. In Olaf Sander's study prolonged elevated HR (PEHR) group was de ned as >95 beats/min for >12 hours in at least one 24-hour period of ICU stay starting at admission [5]. And in Veit Sandfort's study a prolonged eHR (peHR) episode was prespeci ed as 11 hourly heart rate measurements >100 beats/min in any given 12-hour interval [16]. These are examples of 'dose' parameters. In other studies, measures of HR volatility (standard deviation, % HR >120 bpm, and %HR<60 bpm) were used to predict prognosis [30,31]. In our study, cneHR was used as an index to evaluate HR, it was also a 'dose' and 'volatility' parameter. It can re ex duration of elevated HR, on the ip side may present HR volatility. Most importantly, cneHR can be displayed in real time at the bedside because it can be acquired and calculated automatically.
In general, the large sample size, inclusion of multiple NICUs, and the excellent quality of the clinical data collection by the treatment centers, for example, the periodic heart rate is available for 96% of patients, are advantages of this analysis. The chosen variable to de ne 'dose' and 'volatility'of elevated HR was found to be suitable to identify patients with an increased risk of mortality in the NICU. However, there are several limitations of our study. Firstly, the study is retrospective, its post hoc nature should be taken into account when considering the ndings. Residual confounding may also in uence our ndings, although we attempted to account for this through several adjustments and different models. Secondly, although the datasets we used are large and comprise routinely collected clinical data, some patients had to be excluded owing to poor-quality data recording or missing data. The amount of missing data onto the hemoglobin assessed in the study is a potential limitation. In the stepwise regression analysis, hemoglobin was contained in the nal model. Because of the missing data, we dropped it in the predictive model. Thirdly, in eICU-CRD, all patients admitted from the operating room or recovery room had a surgical diagnosis, even though some surgical patients have a medical reason for admission. That may to some extent change disease spectrum. Although our ndings do support an association with HcneHR and NICU mortality, stronger evidence is necessary to establish a causal relationship.

Conclusions
In adult neurocritically ill patients , HcneHR in the rst day of NICU stay is independently associated with higher NICU mortality and several others important patient-centered outcomes. We thank the anonymous reviewers for insightful and helpful comments.
Funding:This study was supported by grants from key support projects of "Yangfan Plan" of Beijing Medical Administration(NO.ZYLX201836).

Availability of data and materials
The datasets generated and analysed during the current study are available in the eICU Collaborative Research Database(eicu-crd.mit.edu).
Authors' contributions DWZ, JXZ and GZS were responsible for the concept and participated the design of the study. DWZ, ZML and SLZ participated the data collection and the data analysis and interpretation. DWZ, JXZ and GZS participated the statistical analysis. DWZ and GZS wrote the rst draft of the manuscript. All authors commented on the manuscript revisions and approved the nal version.

Ethics approval and consent to participate
The eICU-CRD was exempt from institutional review board approval due to the retrospective design, lack of direct patient intervention, and the security schema, for which the re-identi cation risk was certi ed as meeting safe harbor standards by Privacert (Cambridge, MA) (Health Insurance Portability and Accountability Act Certi cation no. 1031219-2). Due to the HIPAA compliant de-identi cation in this database, our institutional IRB requirement was waived. Patient's consents were waived due to the retrospective design of the study.        odds ratio for various heart rate and cumulative numbers over corresponding heart rate

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download.