Background

Autonomic changes are evident from the onset of acute pathology requiring critical care. Cardiac autonomic function can be derived by analysing variability between heart beats to yield time domain and frequency domain (power spectral density) measures that reflect autonomic modulation of cardiac frequency [1, 2]. Heart rate variability (HRV) appears to contribute diagnostic and prognostic value in various cardiometabolic conditions associated with subclinical autonomic dysfunction that predispose to critical illness including hypertension, coronary artery disease, heart failure and diabetes [3,4,5,6,7]. Similarly, HRV has been proposed to serve as a potential diagnostic and prognostic tool in critically ill patients [8].

However, HRV measures in critically ill patients are fraught with potential problems. [9] Although population norms for HRV parameters have been reported in healthy populations [10], the impact of multiple physiological, procedural and technical factors in critically ill patients has not undergone systematic scrutiny in critical care medicine [11]. Moreover, the validity of HRV as a tool to interrogate autonomic function is increasingly under physiological scrutiny [12, 13], since a strong correlation between HRV and morbidity/mortality appears to be largely attributable to incident heart rate. In addition, recording technique, clinical context and adjustment for incident heart rate are key factors to consider when interpreting the translational relevance of HRV in critically ill patients.

Here, we sought to systematically evaluate the methodology and design of HRV studies in critical care medicine. We focused on whether recommended standards for measurement and reporting have been employed [14, 15], with the aim of identifying areas to refine in future HRV experimental design in critical care medicine.

Methods

Identification of studies

A literature review was performed based on the Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) guidelines for systematic reviews [16]. The summary of the search strategy employed is shown in Additional file 1.

We searched the electronic databases PubMed, EMBASE, MEDLINE and the Cochrane Central Register of Controlled Clinical Trials for articles investigating HRV measurement in intensive care patients. Inclusion criteria were full-text studies written in English involving adult patients, published after 1996 (following published guidelines) and reporting traditional time and frequency domain parameters [15]. Studies which reported newer analysis techniques of HRV (e.g. entropy analysis) were excluded, as we focussed on those reporting measures in line with recent European guidance [17]. The following Medical Subject Headings (MESH) were used to identify pertinent articles: “Heart rate variability OR HRV AND Sepsis”, “Heart rate variability OR HRV AND multiple organ dysfunction OR MODS”, “Heart rate variability OR HRV AND critical illness”, “Heart rate variability OR HRV AND intensive care OR ICU”. The last search took place on 9 November 2016. We screened articles by title search and abstract review. Relevant articles were analysed for eligibility, and further articles were identified from reference lists. Articles were excluded based on the following criteria: experimental studies, incorrect target population (adult; >18 years old), medical field other than intensive care, not original research, topic not within scope or traditional HRV parameters not reported.

Data extraction

Data was extracted by two independent reviewers (S.K and A.S) and recorded into a standardised excel sheet recording: author, year of publication, study design, number of subjects, mean patient age, proportion of male subjects, risk stratification score, comparator groups, study aim and outcome, study design, protocol for measurement, processing, analysis and reporting of HRV parameters, adjustment and reporting of confounding factors and quality assessment. We identified the following clinical confounding factors: age, gender, average heart rate, average respiratory rate, co-morbidities, drugs, sedative drugs, vasoactive drugs, enteral nutrition and mechanical ventilation. Full details of the impact on HRV of these parameters are provided in Additional file 1. For reporting and analysis purposes, we selected the most commonly used time and frequency domain HRV parameters [15].

Risk of bias and study quality assessment

The quality of studies was assessed by two assessors independently (SK, SM) using two established tools (Newcastle–Ottowa scale, Downs and Black Instrument). The Downs and Black instrument is recommended by the Cochrane Collaboration for use in non-randomised and observational studies (Additional file 1) [18, 19]. Inter-observer reliability evaluating quality within five domains: reporting, external validity, bias, confounding and power. Five questions were omitted because they are designed for interventional trials. The version which we employed in this study therefore has a maximum score of 22. Differences between reviewers were resolved by panel consensus opinion following further review of the article(s) in question by the senior author.

Results

Study selection

We identified 238 studies which underwent screening by title search and abstract review. From these, 31 articles involving 2090 patients (including controls) met the inclusion criteria for assessing the role of HRV in critically ill patients [20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53]. Two articles analysed the same cohort of patients [34, 37].

Study characteristics

Demographic and clinical data, including comparator groups are summarised in Table 1. All articles reported cohort or case–control studies. The average age of patients was 60 ± 7 years. The majority of studies (22/31; 71%) explored the association between HRV measures, morbidity and mortality. Key clinical findings from these studies are summarised in Table 2. Due to significant differences in trial design, methodology, confounding, non-standardised comparator groups and inconsistent reporting of summary data, a meta-analysis could not be performed. However, there was consistency between studies in their findings that LF/HF ratio was inversely associated with increased disease severity or mortality. For illustrative purposes, the individual effect sizes across six studies reporting mean and standard deviation data looking at disease severity and mortality using the most commonly reported HRV parameter (LF/HF ratio) are shown (Fig. 1).

Table 1 Demographics and study design of studies
Full size table
Table 2 Study objectives and key findings
Full size table
Fig. 1
figure 1

Forest plot of individual effect sizes (Cohen’s d) across six studies investigating the relationship between LF/HF ratio and disease severity and mortality

Full size image

Quality of studies

No studies reported Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines. Two studies analysed data retrospectively. A minority of studies (n = 5; 16%) used individualised HRV data—i.e. patients serving as their own control, prior to an intervention. More than one third of studies (n = 13; 42%) did not describe any comparator group. The remainder of studies used non-age matched healthy volunteers, non-critically ill patients with established cardiovascular disease or HRV values derived from the literature. External validity (as adjudged by the Down and Black assessment tool) was poor, with the majority of studies achieving a score of 1.

Risk of bias assessment

We found recurring potential sources of bias in study design, with 19 (61%) studies failing to report whether HRV data analysers were masked to the patient condition/outcome (Additional file 1). Only one study performed a power calculation [41].

Data acquisition and preparation

Details on short-term recordings, including source of heart rate periods [54, 55], duration of recordings, epochs used for analysis and patient position [56] were variable or not reported. Fourteen (45%) studies did not describe the sampling frequency of recordings; four (13%) studies used sampling rates below the recommended 250 Hz [15].

ECG recording in the critically ill population is frequently contaminated by electrical and physiological artefacts. Thus, detailing methods to detect artefact (manual or automated) and its management (segment selection, deletion or interpolation) is important for data interpretation [57]. Fourteen (45%) studies reported automated and/or manual editing of the raw ECG to remove artefact by replacing the missing data with cubic spline or linear interpolation methods. In keeping with guidelines, the majority of studies used interpolation methods as opposed to deletion of abnormal beats to avoid a loss of information [15].

HRV analysis

Measurement protocols, processing and reporting of HRV data are summarised in Table 3.

Table 3 Procedures for measurement, processing and reporting of HRV
Full size table

A minority of studies (14; 45%) reported both frequency and time domain data (Table 3). A minority of studies (9; 29%) reported frequency data in normalised units together with absolute values, in keeping with established recommendations. Summary values for commonly reported HRV parameters revealed a wide range of non-normally distributed data for each (Additional file 1: Table S3). Reporting and/or adjustment for heart rate and respiratory rate, which dramatically alter both high and low frequency spectral components [58] was inconsistent between studies. A small majority of studies (17; 55%) reported average heart rate, whilst a minority (6; 19%) adjusted for, or reported, respiratory rate during data acquisition.

Pharmacologic and clinical interventions

Studies varied in their exclusion criteria and reporting of potential confounding factors including age, gender, body mass index [59], common comorbidities [60,61,62,63], drug therapy [64,65,66,67,68] and/or ICU interventions (Tables 4 and 5). Exclusion criteria used and comorbidities/drugs are summarised in Additional file 1. A minority of studies (12; 39%) excluded patients with chronic comorbidities that are commonly associated with autonomic dysfunction. Reporting of drugs that directly affect autonomic function was highly variable across studies. A majority of studies (25; 81%) did not detail drug therapy. Around 22% studies did not report the use of mechanical ventilation, and more than 25% failed to report whether sedation and/or vasoactive drugs were used at the time of HRV recordings.

Table 4 Reporting of potential clinical confounders
Full size table
Table 5 Reporting of potential confounders
Full size table

Discussion

This review is the first to systematically explore how HRV analyses are undertaken and/or reported in critically ill patients. Despite a wealth of laboratory and translational data suggesting that HRV may offer diagnostic and prognostic utility, significant heterogeneity in methodology between HRV articles precluded comparisons across studies and meta-analysis. Our review identifies several areas that require greater scrutiny in future, highlighting the need to develop consensus guidelines that are relevant and tailor-made for the challenges faced by researchers in critical care medicine.

Well-recognised technical, physiologic and clinical factors impact on the measurement, and interpretation of HRV [69, 70]. We found highly variable practice in three key technical areas. Low sampling rates (<250 Hz) impair the precise detection of the R wave fiducial point in the ECG waveform, which consequently affects the power spectrum [15]. This is particularly relevant for studies that derived R–R intervals from arterial pressure waveform analysis [20, 47], since non-neural respiratory influences (e.g. changes in ventricular mechanics) differentially affect mechanical pulse waves and electrical R waves [55]. Manual inspection of the raw ECG to identify artefact is preferred to automated methods to avoid the introduction of false frequency components into the power spectrum [57]. The variable (or unstated) masking of HRV analysers to clinical data also introduces potential significant bias.

From a physiologic perspective, reporting and/or adjustment for heart rate and respiratory rate was inconsistent between studies, with heart rate frequently not reported. Across species with highly variable heart rates, HRV appears to be largely attributable to incident heart rate. If heart rate is not taken into account, erroneous conclusions regarding HRV are likely since differences may merely reflect lower heart rate [12]. This is particularly of relevance to hemodynamically unstable critically ill patients, in whom heart rate may rapidly change. Similarly, increases in respiratory frequency and tidal volume affect both high and low frequency spectral components [58]. Hence, standardised criteria for ventilatory and heart rate reporting are required for the interpretation of HRV data between studies (and hence, potentially, meta-analysis).

From a clinical perspective, HRV parameters are influenced strongly by age, gender, functional capacity and chronic comorbidities. Whilst all studies estimated severity of illness, the most frequently employed—Acute Physiology and Chronic Health Evaluation II (APACHE-II)—are limited in capturing information about chronic comorbid disease that are over-represented in the critical care medicine population. For example, diabetes mellitus, a common condition associated with cardiac autonomic neuropathy, is not captured by this type of assessment [60]. Typically, chronic conditions at the severe end of the disease spectrum are included (e.g. APACHE-II score only includes severe heart failure (≥NYHA class 3). However, HRV parameters have been found to be abnormal in early cases of chronic disease, including preserved ejection fraction, coronary artery disease, chronic kidney disease and hypertension [60,61,62,63]. Although some studies have considered these factors, serial measures or dynamic autonomic challenges offer a potentially more insightful and individualised approach to assessing HRV. Novel HRV parameters that can be captured within the first few minutes of critical illness, such as deceleration capacity of heart rate [71], may mitigate the need for refining the use of more traditional time and frequency domain measures. For mechanistic studies investigating whether changes in autonomic parameters correlate with, or precede, pathologic events, targeting clinical scenarios where multiple, complementary baseline autonomic measures [72, 73] can be made before critical illness develops may be optimal [74]. Studies where basal autonomic function can be captured, including elective surgery [73,74,75,76] and oncologic sepsis [48, 49], may provide particularly powerful mechanistic insights since autonomic changes can be individualised and referenced to pre-insult normal, or pre-existing, dysfunction. Several studies have highlighted that HRV values in critical care medicine are similar to those found in common cardiovascular pathologic conditions [74, 75, 77]; this highlights the need for individualised patient data in order to rule out that autonomic dysfunction is not a precursor of critical illness, rather than merely a biomarker.

Commonly used anti-arrhythmic drugs, anti-hypertensive drugs, statins, metformin and inhaled bronchodilators have all been associated with changes in HRV parameters [60,61,62,63]. However, the lack of reporting on medications that critically ill patients received reduces the mechanistic insight afforded by this approach, particularly given the strong correlation between HRV and morbidity/mortality appears to be largely attributable to incident heart rate. Similarly, the majority of studies in this review failed to consistently report on the use of common critical care interventions. This may explain why conflicting conclusions over how variety of features of critical illness may affect HRV. Continuous enteral or parenteral nutrition are both associated with a reduction in time domain HRV measures indicative of parasympathetic cardiac modulation [67]. However, we did not find any studies that reported on the feeding or fasting status of patients. Although a significant limitation of our study was the lack of primary source data, in any event, we could not identify a single common HRV parameter measured in all studies that enabled comparison. A further limitation is that we did not consider newer nonlinear and multiscale approaches, since very few studies incorporating these analyses have been undertaken. These approaches are also likely to be affected by the same factors that influence traditional HRV parameters [78]. Thus, in a clinical setting, further work is required to establish whether these newer approaches reduce the impact of several confounding factors we have identified in this review.

Conclusions

Heart rate and derived heart rate variability offers a non-invasive, inexpensive tool that may add mechanistic insights to our understanding of critical illness and also assist clinical care. However, the current interpretation of generalizable and clinically relevant values to aid clinical decisions/research is hampered by a non-standardised methodologic approach and lack of adjustment for important confounding factors. For critical care medicine to exploit recent advances in translational autonomic physiology, further high-quality prospective HRV studies underpinned by the development of consensus reporting standards relevant for critical care medicine are needed.