HRT assessment reviewed: a systematic review of heart rate turbulence methodology

Heart rate turbulence is the naturally occurring fluctuation of heart rate (HR) after a ventricular premature contraction (VPC) (see figure 1) (Schmidt et al 1999). While re-establishing the former blood pressure, which fluctuates due to a VPC, a characteristic heart rate pattern occurs (see figure 2): as the VPC is a premature beat, it leads to a shortened interval, called the coupling interval (couplI). This interval is followed by a compensatory interval (compI) that is much longer than a normal cycle duration, because the ectopic beat suppresses one contraction that would regularly have been triggered by the sinus signal. After these two irregular intervals, a characteristic pattern can be observed which consists of an initial fast increase of HR followed by a smooth HR decrease and a latter return to the baseline. The term heart rate turbulence (HRT) refers to this fluctuation of the HR after the compI.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

The turbulence depends mainly on the baroreflex and thus is an indirect marker of the condition of the autonomic nervous system (Cygankiewicz et al 2013). Two HRT parameters, turbulence onset (TO) and turbulence slope (TS), were developed to define HRT.

TO quantifies the initial increase of the heart rate after the compI and is suggested to reflect vagal inhibition (Lombardi et al 2011). The VPC is hemodynamically inefficient compared to a normal sinus beat: It has lower contractility strength due to the missing atrial contraction, incomplete electrical recovery, and thus less synchronization and it moves less blood volume due to less diastolic filling and higher afterload. Hence, the systolic blood pressure drops. This lack of afferent baroreflex input results in vagal inhibition (Bauer et al 2008). To quantify this fast reaction, the relative difference between the arithmetic mean of the two beats after compI ($RR_1 \& RR_2$) and the arithmetic mean of the last two beats before couplI ($RR_{-1} \& RR_{-2}$) is calculated. TO is given as a percentage:

$$\begin{equation}((RR_1 + RR_2) - (RR_{-1} + RR_{-2}) / (RR_{-1} + RR_{-2}) * 100 [\%].\end{equation} \tag{ 1 }$$

It is first calculated for each VPC and averaged afterwards to get a subject's overall TO value.

TS is the steepest slope of the increase of interval length following the compI and reflects vagal activation (Lombardi et al 2011): after the initial inhibition vagal activity recovers with increasing interval lengths (Bauer et al 2007). The slope is measured as the steepest regression line over any 5 succeeding intervals within the first 20 intervals after the compI and is given as ms/RR. In contrast to TO, an averaged tachogram of all suitable extrasystoles is calculated first, before TS is assessed once for a subject.

The workflow of HRT assessment begins with a 24-h Holter electrocardiogram (ECG)-recording. The record has to be cleaned from erratic data, meaning artefacts and noise. Afterwards, all VPC snippets (VPCSs) have to be singled out. A VPCS contains several intervals: the couplI between a sinus beat and the early occurring VPC, the long compI between the VPC and the following sinus beat, and a number of regular intervals preceding (preRRs) and following (postRRs) these two VPC intervals (VPCIs). The intervals of the VPCS are filtered based on a set of established criteria (Grimm et al 2003). These criteria regard the length of the intervals to discard ectopy within the regular intervals and ensure a sufficient prematurity of the VPC. Afterwards, HRT can be calculated from these suitable VPCS as mentioned before. In healthy volunteers TO ranges from −2.7% to −2.3%, TS ranges from 11.0 to 19.2 ms/RR interval (Diaz et al 2002, Grimm et al 2003, Tuomainen et al 2005). Finally, patients can be classified into risk groups. Established as representing a high risk are TO values higher than 0% and TS values less than 2.5% as suggested by Schmidt et al (1999).

In many studies the two HRT parameters have been identified as feasible markers for all-cause mortality after myocardial infarction or chronic heart failure as reviewed by Cygankiewicz et al (2013). In recent studies HRT has been shown to significantly contribute as predictive factor in different clinical settings. HRT combined with traditional heart rate variability increases the sensitivity of the diagnosis of early cardiac autonomic neuropathy in diabetic patients to 98% (Lin et al 2017). Including HRT to a predictive model of death associated with chronic heart failure improved the accuracy of the model Ramírez et al( 2017). HRT increased the predictive sensitivity and specificity in life-threatening ventricular tachyarrhythmias (VTA) (Frolov et al 2017). The predictive value of HRT has also been shown in subjects with Marfan syndrome (Schaeffer et al 2015), chronic obstructive pulmonary disease (Gunduz et al 2009), metabolic syndrome (Erdem et al 2012) or obstructive sleep apnea syndrome (D'Addio et al 2013).

In contrast to heart rate variability (HRV) that shows autonomic activity during sinus rhythm, HRT reflects autonomic responses to endogenous interference. Cardiac diseases can significantly depress autonomic functions and thus attenuate HRV (Malik et al 1996). Since they can also increase VPC occurrence (Gorenek et al 2020), HRT may be able to detect residual autonomic activity that cannot be detected by HRV. Therefore, it can be a valuable addition to risk prediction, that usually incorporates many different indicators.

In 2008 the International Society for Holter and Noninvasive Electrophysiology Consensus was published (Bauer et al 2008). Though it is referred to as a standard for HRT calculation, it did not define norms but rather summarised commonly used methods. Despite the common ground established in this paper, HRT methodology varies widely regarding data recording, filtering of measurements and calculation itself. Different methodologies lead to different results. This lack of comparability hinder the understanding of HRT, e.g. the intensity of influences on this phenomenon. It also complicates finding suitable standards for HRT application in the medical field.

We systematically assess the methodology of determining HRT parameter values in research since its original description. The scope of the study is not to review the clinical use of HRT but the assessment itself. The aim is to outline the different types of HRT assessment as well as their usefulness in order to create a basis for determining precise guidelines that will lead to higher comparability between studies.

When defining guidelines the aim of ECG recording and HRT assessment have to be considered: Recording ECGs in order to assess HRT is rare and limited to scientific research. In most cases HRT is calculated from records made during clinical practice and without focus on HRT. Therefore, guidelines have to take into account what is possible to implement in the daily clinical routine.

Our approach followed the Preferred Reporting Items for Systematic reviews and Meta-Analyses( PRISMA) statement (Liberati et al 2009) wherever possible. The PRISMA statement provides guidelines for systematic reviews and meta-studies in the medical field (Liberati et al 2009). Since the scope of our study differs from common reviews not all guidelines were applicable.

In a PubMed search made on 12.10.2018 we gathered all articles containing the term 'heart rate turbulence'. After checking for HRT assessment and language we examined the remaining 240 articles regarding the exact methodology of HRT that was used in the respective study. As baseline methods we used the filtering criteria of Grimm et al (2003), the calculation rules by Schmidt et al (1999) and the minimum number of needed VPCS as reviewed by Bauer et al (2008). A detailed description of our methods with all examined aspects, baseline methods and resulting number of articles can be found in the appendix Methodological details.

We want to give an overview about the most common practices regarding HRT assessment as well as a critical evaluation to provide the reader with guidelines. Therefore we structured our findings into chapters following a customary data assessment:

• Section 2 'Data collection' summarises the variety of the length of the ECG measurement, different approaches to assess HRT after other arrhythmias and the technique of inducing VPCs to circumvent the difficulty of getting enough suitable VPCSs. It also includes the approach to adapt TS to address several biasing influences due to different recording lengths.
• Section 3 'Data preparation' assesses the application of filtering criteria for VPCSs by Grimm et al (2003). Furthermore, it covers the number of VPCSs needed to calculate HRT and the different approaches to handle measurements with lesser VPCSs.
• Section 4 'Data analysis' covers variations of the calculation techniques by Schmidt et al (1999) and a third established HRT parameter.
• Section 5 'Classification' discusses cut-off values for risk stratification and classification systems.
• Section 6 'Description of HRT assessment' covers the methodological details that need to be given in publications.

Although most inspected articles state the planned recording duration, the actual recording duration (i.e. given as mean of the population) is mostly missing: in 198 articles either Holter monitoring or 24 h are reported as the planned recording duration. Some studies used short recordings with a planned duration of less than 1 h (Raj et al 2005, Berkowitsch et al 2004, Davos et al 2009, Malberg et al 2004, Voss et al 2006, Davies et al 2001, Jeron et al 2003, Wichterle et al 2006, Goernig et al 2006, Segerson et al 2007, Watanabe et al 2006, Tekelioglu et al 2013) or between 2 and 8 h (Solem et al 2008, Rich et al 2012, Gil et al 2013, Tekelioglu et al 2013, Solem et al 2006, D'Addio et al 2013, D'Addio et al 2014, Yang et al 2005, Watanabe et al 2008, Tuomainen et al 2005). In several studies 24-h Holter recordings were cut into 1, 2, 4 or 6 h snippets (Cygankiewicz et al 2004, Hallstrom et al 2004, Hallstrom et al 2005, Lewis et al 2011, Watanabe et al 2007). In contrast, two studies used longer recordings, namely 28 h (Ovreiu et al 2008) and 7 days (Manzano-Fernández et al 2011).

Naturally, HRT can only be calculated when recordings contain VPCSs. Based on the researched population, obtaining a sufficient number of suitable VPCSs can be challenging, since the prevalence of VPCs is positively correlated with increasing age (Kostis et al 1981) and higher cardiovascular risk (Latchamsetty et al 2015, Lown et al 1971). The amount of VPCSs is likely to decrease even more after filtering. Options to address the problem of having a low incidence of VPCs for HRT assessment are: extending the recording time, inducing VPCs or using other heart beats to calculate HRT.

2.2.1. Extending the recording duration

2.2.2. Induced VPCs

2.2.3. HRT based on other arrhythmias

HRT can be influenced by various factors. Apart from diseases associated with the impairment of the autonomic nervous system, HRT is biased by age, HR and the number of VPCSs used for the calculation as reviewed by Bauer et al (2008) and Cygankiewicz et al Cygankiewicz et al (2013). The latter two influences have to be taken into account during data collection, because they can have implications on the desired ECGs recording time.

2.3.1. Influences related to recording duration

2.3.2. Adjustment of turbulence slope

An approach to bypass these biases on TS is to normalise it with regard to HR and the number of VPCSs used. This parameter would be independent of the circadian pattern and the length of the recording. Several adaptions have been proposed that are discussed below. To reduce confusion we adopted the nomenclature of the parameters used in the respective original papers.

Cygankiewicz et al adapt TS to HR (Cygankiewicz et al 2004, Cygankiewicz et al 2004) by assuming a relationship of the form:

$$\begin{equation} TS = TS_c \cdot RR^x \end{equation} \tag{ 2 }$$

where TS_c is the 'corrected' TS, that is the part of TS that is independent of HR. The parameter x is population-specific. It can be determined by a linear regression of log RR against logTS, since

$$\begin{equation} \log TS = x \log RR + \log TS_c. \end{equation} \tag{ 3 }$$

For their population, Cygankiewicz et al find a value of x = 3.4. With this, TS_c becomes

$$\begin{equation} TS_c = \frac{TS}{RR^{3.4}}. \end{equation} \tag{ 4 }$$

Hallstrom et al used two normalisation steps for HR and the number of VPCSs. First, they rescaled the tachogram to an average HR of 75 bpm before calculation which yields nTS. In a second step, they also address the fact that TS (and also nTS) is biased towards higher values for a low number of VPCSs. For a set of intervals that should yield a TS of zero, the maximising step in the calculation actually introduces a slope proportional to the standard deviation of the local variation in the intervals. The averaging step before the calculation of this slope reduces this standard deviation, which can be assumed to be equal to RMSSD ⁶ , by a factor of $1/\sqrt{\#VPCSs}$. This yields the following formula:

$$\begin{equation} TS_{k} = 0.024\,75 \cdot (k-2)^{0.944\,9} \frac{RMSSD}{\sqrt{\#VPCSs}}. \end{equation} \tag{ 5 }$$

Here k is equal to the number of intervals in which TS is calculated (#TSRR). The numerical parameters were determined by Hallstrom et al by a fit of the formula to simulated data. The new unbiased estimator vnTS is then defined as

$$\begin{equation} vnTS = TS - TS_{k} .\end{equation} \tag{ 6 }$$

This approach was reused by Yang et al( 2005) and D'Addio et al( 2014).

The adjustments of TS to HR and the number of VPCSs abolished the dependency on circadian patterns for both parameters as well as the dependency on recording duration for vnTS Hallstrom et al( 2004). Like TS the new parameter vnTS is correlated with age Hallstrom et al (2004).

Hallstrom's vnTS showed the same clinical value as the standard TS: D'Addio et al found a significant difference of TS as well as vnTS ⁷ in obstructive sleep apnea syndrome (OSAS) patients during normal and obstructive apnea breathing (D'Addio et al 2014). Similarly, Yang et al used vTS (only adapted to the number of VPCSs) and found a significant decrease of both TS and vTS in patients with moderate or severe OSAS compared to patients with mild OSAS with both parameters identically inversely correlating with apnea-hypnea index (Yang et al 2005). But the positive and negative predictive value of TS was higher than of vTS (Yang et al 2005). Still, TS as well as vnTS were both univariate predictors of survival after MI (Hallstrom et al 2005). Cygankiewicz' adjusted TS_c , too, showed the same results as TS and significantly decreased after CABG surgery (Cygankiewicz et al 2004).

Lin et al also normalised TS to HR, but used 1000 ms as normalisation interval length (Lin et al 2004). Malberg et al introduced a number of other variations of TS (Malberg et al 2004). It was adjusted to the length of the couplI and compI and several blood pressure indices. Both Malberg et al and Lin et al did not compare their adapted TS to standard TS (Malberg et al 2004, Lin et al 2004).

Almost all comparisons between the standard TS and the respective adjusted TS show equal results regarding their usability (D'Addio et al 2014, Yang et al 2005, Hallstrom et al 2005, Cygankiewicz et al 2004). However, an TS adjusted to the HR as well as the number of VPCSs used is less variable and therefore much more comparable.

Most articles use 24-h recordings for HRT assessment. However, a minimal number of VPCSs has to be recorded for parameter calculation, which may be a challenge in different populations. Therefore, longer recordings could increase the amount of suitable data similarly to HRT induced by electrical stimuli and HRT calculated from other types of contractions. While extending the recording duration and induced VPCs are not feasible for clinical routine, calculating HRT based on other contractions might increase the accessible data. Best studied is HRT after APCs, but further research is needed to determine the prognostic value of these types of HRT. Particularly, we suggest whether a shifted onset of HRT after a APCs in comparison to VPCs can be verified and eventually adapt TS calculation to take this shifted onset into account. For HRT assessment, where longer recordings are possible, the extent of HRT variability between days remains to be studied and should be taken into account when determining the optimal recording duration.

Furthermore, HR and the number of VPCSs bias TS. This leads to a dependency of this parameter on the recording duration as well as on circadian rhythm. To cancel this influence we suggest to use a TS adjusted to HR and the number of VPCSs. Nevertheless, to improve comparability we suggest to give the arithmetic mean of the ECG measurement duration and the time of day if the recording durations are no multiple of 24 h as well as the mean of VPCSs used for HRT calculation per patient.

Research questions:

• Can HRT after other arrhythmias be used with a similar prognostic value?
• Is there a day-to-day variability of HRT?
• How long is the optimal recording duration to calculate valid HRT parameter values?

Only snippets that are as free as possible of bias, e.g. other arrhythmias, and which contain effective VPCs with a sufficient prematurity should be included in HRT calculation. Therefore, filtering of the VPCs and their surrounding intervals is necessary. Since filtering criteria of the ECG data determine the input for the calculation, the exact procedure is crucial. In the original publication of HRT no criteria for filtering were suggested (Schmidt et al 1999). The first set of criteria was introduced by Grimm et al (2003) (see the filtering criteria summarised in the section 1.1 or detailed in the appendix). These criteria were repeated almost identically in the HRT guidelines (Bauer et al 2008). Nevertheless, there are many different approaches to filter the interval series before parameter calculation. Some studies report excluding VPCSs containing abnormal beats (arrhythmia or ectopy) or erratic data (artefacts and noise). Commonly, a quantitative filtering is made regarding the length and the range of the inspected intervals.

Length of intervals in a VPCS: The minimal prematurity as well as the minimal length of compI varies in different studies and is sometimes not explicitly stated. The surrounding preRRs and the regular intervals in a VPCS after the compI (postRRs) are either filtered via absolute values or proportionally in comparison to preceding intervals or a calculated reference interval (refI) (see table 1).

Table 1. Filtering criteria that differ from the Grimm et al criteria Grimm et al (2003) regarding the length of all intervals in a VPC snippet (VPCS). How many intervals are included as surrounding intervals varies (see paragraph Range of checked intervals). NS: not specified in article.

Criterion	RefI	References
Length of couplI
20-80%	Preceding interval	Fazio et al (2010)
88%	NS	Lindgren et al (2003)
100 ms shorter	Preceding interval	Watanabe et al (2007)
'Less'	Preceding interval	Bauer et al (2006)
Not '(too) long'	-	Manzano-Fernández et al (2011), Casella et al (2006)
'Distinct'	-	Vikman et al (2005)
Length of compI
110%	Mean of five intervals preceding the couplI a	Tuomainen et al (2005), Tuomainen et al (2005)
100 ms longer	Preceding interval	Watanabe et al (2007)
Not '(too) short'	-	Manzano-Fernández et al (2011), Casella et al (2006)
'Distinct'	-	Lindgren et al (2003)
'Full'	-	Goernig et al (2006)
Length of surrounding intervals
110%	Mean of five intervals preceding the couplI b	Tuomainen et al (2005), Tuomainen et al (2005)
$\lt$ 2000 ms	-	Watanabe et al (2007)
200-2000 ms	-	Park et al (2014), Malberg et al (2004), Kowalewski et al (2007)
200-2500 ms	-	Kossaify et al (2014)
80-120%	Preceding interval	Bauer et al (2006)
80-120%	'5 consecutive sinus intervals'	Golukhova et al (2016)
80-120%	Mean c of all intervals in recording	Yoshihisa et al (2014), Walker et al (2016), Moore et al (2006)
Not 'very short or long'	-	Cano et al (2008)

^a Notice: The phrasing in the papers is 'mean', which probably refers to the arithmetic mean. ^b Notice: The phrasing in the papers is 'mean', which probably refers to the arithmetic mean. ^c The exact type of mean is not specified in the articles.

Range of checked intervals: The first range in which intervals were checked was introduced by Davies et al consisting of 20 intervals before and after the VPCIs, respectively (Davies et al 2001). The range of 2 preRRs and 15 postRRs, which was proposed by Bauer et al, was first defined by Grimm et al (2003) and propagated on www.h-r-t.org/.com ⁸ . Most authors do not explicitly specify the checked range before and/or after the VPCIs (176 articles), though some cite Bauer's guidelines (Bauer et al 2008) or use the algorithm from www.h-r-t.org/.com.

Most authors that state the range explicitly use the ranges of 2 preRRs and 15 postRRs, 5 preRRs and 15 postRRs, or 20 intervals before and after, respectively (see table 2).

Although only two intervals are needed for HRT calculation, many studies filter more intervals before the VPC. This is in accordance with the results of Chen showing a negative and positive correlation of TO and TS, respectively, with the number of VPCs within the 2 minutes before the VPC used for HRT calculation (Chen 2009). However, the results may be biased for several reasons: different sample sizes of HRTs with or without preceding VPCs in the study (Chen 2009), the incidence of VPCs due to circadian rhythm (Lown et al 1973, Chen 2011) and the correlation of diminished HRT with a high incidence of VPCs (Cygankiewicz et al 2004).

Number of VPCSs needed for HRT analysis: The minimal number of suitable VPCSs needed for HRT analysis (#minVPCS) is not explicitly stated in most articles (173). Sometimes a minimal number of VPCs is given as a criterion for a patient's inclusion in the study; however, these VPCs are not explicitly defined to be part of VPCSs suitable for analysis. While Bauer et al suggested 5 VPCSs for HRT calculation (Bauer et al 2008), various numbers between 1 and 100 have been used, mostly 5, 1 and 2 (in descending order of the frequency of usage, see table 3).

Only two articles used a maximal number of VPCs, namely 2400 (Szydlo et al 2011) and 200 (Chen 2009).

All numbers used for #minVPCS are merely suggestions, because the optimal #minVPCS has yet to be systematically identified. By now there have only been two studies focusing on #minVPCS: Both did not find any difference of the predicitve value of HRT when comparing the whole dataset or just patients with a minimum of 2 (Osman et al 2004) or 4 (Berkowitsch et al 2004) VPCs. However, their approach compares two groups where one is the subset of another, which inherently makes finding a difference less probable. Cygankiewicz et al found that TS may be only correlated to the number of VPCSs for low absolute numbers of VPCSs, while the dependency of HRT on the number of VPCs vanished in patients with more than 10 VPCSs (Cygankiewicz et al 2004). A systematic study to find #minVPCS for reliable risk stratification should not only focus on significant differences of HRT parameters between study groups, but on their variances and should be done with a larger range of numbers of VPCSs.

Data with insufficient number of VPCs: Most authors do not specify the procedure regarding measurements not meeting the #minVPCS criterion. In most cases it can only be assumed that the subjects were excluded from analysis. The second most common approach is categorising subjects with too few VPCs in a low-risk group. Some studies use the category (HRT0 Kop et al 2010, Harris et al 2013, Maeda et al 2009, Erdem et al 2013, Marynissen et al 2014, Exner et al 2007, Poliwczak et al 2014) or define these subjects as having 'normal' (HRT Stein et al 2011, Hayano et al 2011, Stein et al 2008, Stein et al 2009, Bauer et al 2009, Zuern et al 2012, Huikuri et al 2010, Stein et al 2009, Rizas et al 2017) or being HRT 'negative' (Kawasaki et al 2015, Miwa et al 2012, Hoshida et al 2013, Miwa et al 2011). Other approaches use an additional category with indeterminate patients (Carney et al 2007, Manzano-Fernández et al 2011, Barthel et al 2003, Zuern et al 2012, Golukhova et al 2016). For detailed information about classification systems see section 5.

Some authors differentiate between too few VPCSs and a sufficient number of VPCSs, that, however, are not measurable (Carney et al 2007, Stein et al 2009, Stein et al 2008, Stein et al 2011).

It is reasonable to categorise subjects with too few #minVPCS as having low risk, since a high number of VPCs is correlated with high risk of SCD in chronic heart disease (Lown et al 1971) and poor prognosis after MI as reviewed by Latchamsetty et al (2015). The data situation however is insufficient and contradictory: Barthel et al found no difference between postinfarction patients in groups HRT0 (low risk) and HRT$\oslash$ (noncalculable) (Barthel et al 2003), while Manzano-Fernández et al show that 5 of 12 subjects classified as HRTx (noncalculable) moved to HRT1 (intermediate risk) or HRT2 (high risk) within 6 days of recording (Manzano-Fernández et al 2011). However, in contrast to Barthel et al, Manzano-Fernández et al made no survival analysis so that patients moving to higher risk groups may be false positives (Manzano-Fernández et al 2011). Additionally, since TS decreases with increasing number of VPCs, 'worse' values should be expected with longer measurements. For more information about this influences see section 2.3.1.

For risk stratification in the clinical practice a standardised workflow needs to be established. It must be investigated whether an unsuitable number of VPCSs displays a low risk regarding HRT or whether HRT should be removed from the risk model in that case.

After the common step of ensuring the quality of a recording, the VPCSs need to be sorted out. Therefore, filtering criteria specify the length of the intervals and the range that has to be checked. These criteria are mostly not mentioned in the literature and if mentioned vary widely. We suggest the usage of Grimm's filtering criteria for the time being (Grimm et al 2003). One exception from these criteria should be the number of preRRs being 5 intervals instead of just 2, because these intervals are used as reference for other filtering criteria. It should be analysed how many intervals before the couplI are suitable to exclude VPCSs that are directly preceded by a VPC and are therefore biased. It is noteworthy, that the number of postRRs should be longer or equal to #TSRR.

The #minVPCS mostly used is 1, 2 or 5. The optimal #minVPCS still needs to be determined systematically in a large sample considering the difficulty of low incidence of VPCs on one hand and statistical validity of a sufficient amount of data points on the other. Subjects with less VPCSs are either excluded from the studies or classified as having low risk. It must be investigated which procedure is optimal for risk stratification in the clinical setting.

Research questions:

• What is the minimal and optimal number of VPCSs to calculate valid HRT parameter values?
• How many preRRs are needed to avoid bias by preceding arrhythmias?
• What is the optimal handling of subjects with too few VPCSs?

Most studies (203) follow the calculation process suggested by Schmidt et al (1999). In 136 articles the method is explicitly stated, in 67 a calculation reference is given. In 31 studies TO is used, but no calculation or reference is described. TO is consistently calculated with the two intervals before and after the VPCIs, respectively, and is hardly calculated with a different approach.

In most studies #TSRR is either 15 as first suggested by Barthel et al (2003) or 20 intervals as suggested by Schmidt et al (1999). In 129 articles the range is explicitly stated; 77 provide only a calculation reference. Apart from this the calculation process remains the same. Some studies examine the use of TS that was adjusted to several biases, which we describe in 2.3.2. For higher comparability we suggest to use this adjusted TS as defined by Hallstrom et al (2004) instead of the original TS.

The most commonly calculated HRT parameter apart from Sc hmidt's originals is turbulence timing (TT). It was first defined by Watanabe et al as the index of the first interval in the interval series, whose regression line has the steepest slope and is therefore used for the TS calculation (Watanabe et al 2002). It has been used since then in nine of the studies inspected here (Schwab et al 2004, Schwab et al 2005, Bonnemeier et al 2005, Ortak et al 2005, Schwab et al 2005, Watanabe et al 2006, Średniawa et al 2010, Cebula et al 2012, Bonnemeier et al 2003). Its value ranges from 1 to #TSRR-4.

TT was positively correlated with HR (Watanabe et al 2002, Średniawa et al 2010). While Schwab et al found no gender-dependent difference of TT (Schwab et al 2004), Średniawa et al found TT to be significantly higher in women than in men (Średniawa et al 2016). No conclusion is possible, since women have significantly higher HR in the study population of Schwab et al and Średniawa et al do not mention HR as a covariate (Schwab et al 2004, Średniawa et al 2016). While no differences of TT could be found between postinfarction patients 10 days and 1 year after MI (Ortak et al 2005), TT was an independent risk stratifier for major adverse cardiac event (Cebula et al 2012) and all-cause mortality (Średniawa et al 2016) after MI as well as in patients before and after percutaneous coronary intervention (Bonnemeier et al 2003) or carvedilol and metroprolol treatment of acute myocardial infarction (AMI) (Bonnemeier et al 2005). Additionally, TT was identified as the most powerful risk predictor for development of end-stage heart failure in patients with stable congestive heart failure (CHF) (Średniawa et al 2010).

The only cut-off for TT used by now is 10. It was determined by Średniawa et al, because it was the ninth decile in another study. Unfortunately, this result could not be found in any of the references given for TT (Średniawa et al 2010). Nevertheless, with this cut-off abnormal values were found in 30% of investigated patients with stable CHF (n = 110) (Średniawa et al 2010), in 28.6% MI patients (n = 500) (Cebula et al 2012) and 28.6% MI patients (n = 489) (Średniawa et al 2016). To this day, no study used this cut-off on a healthy study group to evaluate its specificity.

Until now the optimal #TSRR has not been studied. One standardised range is crucial for comparability, because differing #TSRR can lead to different filters and therefore different VPCS sets as well as diverging TS values for the same VPCs. The optimal range can possibly be deducted from TT, since it describes the first interval of the sequence used to calculate TS. An analysis of the statistical keypoints of this parameter might indicate the optimal value of #TSRR.

Table 4 shows values for TT that have been obtained by articles covered in this work.

The arithmetic mean TT ranges from 3.6 ± 1.0 (Watanabe et al 2006) to 7.3 ± 3.0 (Średniawa et al 2010), meaning that in most cases 15 intervals seem to be suitable as #TSRR. It must be noticed that the choice of #TSRR limits the maximal value of TT and can thus bias the results. Apart from that, a short #TSRR is favourable to increase the number of available VPCSs. Since a turbulence appears in proximity to the VPC, we hypothesise that TS with a TT of 7 or more, meaning a #TSRR of more than 10, just presents random fluctuation rather than HRT. Watanabe et al found 6 to be the maximum TT of all subjects, so 10 was defined to be the minimum number of intervals needed after a VPC to calculate HRT (Watanabe et al 2006). In theory, the intra-subject variability in TT could also be used to distinguish between TS occurring randomly and based on autonomic regulation. For this, TT must be calculated before the tachogram is averaged. If the turbulence is based on a regulated mechanism, it is likely to be similar for each VPCS, thus having little intra-subject variance. If autonomic regulation is failing, however, and TT is calculated from random fluctuations, the variance should be considerably higher, while the TS can still lie in the normal range. However, no results of correlation analysis between TT and TS have been published by now apart from Cebula et al, who found no correlation between all HRT parameters (Cebula et al 2012). Thus, the optimal #TSRR remains to be determined.

TO as well as TS are influenced by the order of their calculation: Schmidt et al propose to first calculate TO in each VPCS and average the results afterwards while TS should be calculated from the averaged tachogram (Schmidt et al 1999). In a few studies TO was calculated from the averaged tachogram instead (Solem et al 2008, Wichterle et al 2006, D'Addio et al 2013, Huikuri et al 2010, Schwab et al 2004, Lin et al 2002, Yang et al 2005, Wichterle et al 2007). More commonly the calculation order of TS is reversed, calculating individual values for each VPCS which are then averaged (Wichterle et al 2003, Chen et al 2011, Savelieva et al 2003, D'Addio et al 2010, Davies et al 2001, Davos et al 2009, Yalta et al 2007, Erdem et al 2015, Golukhova et al 2016, Chen et al 2009, Chen et al 2010, Roach et al 2002, Erdem et al 2012, Erdem et al 2012, Lin et al 2004, Watanabe et al 2006, Erdem et al 2013).

Chen et al showed, that the order of the workflow has an impact on the parameter values: Though the results from both methods were significantly correlated for both parameters, the arithmetic mean of TS after Schmidt's method was considerably lower than TS with averaging afterwards. Since averaging smooths the slope due to different TT, it decreases the overall slope in the tachogram and therefore decreases TS. This difference is most notable when TS values are very low (Chen et al 2011). Thus, averaging the TS results afterwards leads to classification in lower risk groups compared to Schmidt's method as shown by Chen et al (2011), Soguero et al (2013). Notice, that Soguero et al used a different approach to assess HRT categories, thus results may change when the data were assessed as usual (Soguero et al 2013). Nevertheless, the order of calculating and averaging of the HRT parameters affects the results and should therefore be performed uniformly. While averaging the tachogram first reduces noise, assessing TS from the seperate tachograms takes into account the variable onset of the HR acceleration.

In the original article by Schmidt et al, the method description uses the general term 'average', but the numbers given in the article fit the definition of the arithmetic mean (Schmidt et al 1999). Accordingly, the arithmetic mean is mostly used as descriptor for HRT. Few studies use the median (Savelieva et al 2003, Witham et al 2012) or a trimmed mean (Marine et al 2002) instead.

TO is almost consistently calculated as described by Schmidt et al. TS is calculated differently with either 15 or 20 intervals in which the slope is measured. The new parameter TT might help to determine the optimal range for #TSRR. It is defined as the index of the first interval from which TS is calculated and therefore provides information about the occurrence on TS after the VPC. A first analysis leads to #TSRR of 10, but this needs in-depth investigation. However, TT is a new parameter that is easy to assess without additional cost and has been shown to be a feasible risk stratifier. Therefore, we suggest to further investigate its usefulness and possible cut-offs. Another variation of TO and TS calculation is the order of the calculation steps. In some studies they are reversed to Schmidt's original approach. Whether the tachogram itself or the parameter values from single VPCSs should be averaged is not comprehensively studied yet. If parameters are calculated from single VPCSs, they should be given as arithmetic mean.

Research questions:

• How many intervals are sufficient for #TSRR?
• In which order should TO and TS be calculated?

Alternative Cut-off Values: Some authors suggest other cut-off values (see tables 5 and 6). These values are either statistical descriptors like the median or determined with receiver operating characteristics (ROC) analysis. Accordingly, Cygankiewicz et al used the quartiles of their coronary artery disease (CAD) population –0.37% for TO and 4.25 ms/RR for TS, because Schmidt's cut-offs were determined for postinfarctional patients (Cygankiewicz et al 2004, Cygankiewicz et al 2004, Cygankiewicz et al 2003, Cygankiewicz et al 2003). Likewise, quartiles were used in stable CAD patients by Sestito et al (2004). Medians were also used as cut-offs, namely 0.1% (TO) and 2.0 ms/RR (TS), for patients with cardiomyopathy after implantable cardioverter defibrillator implantation (Seegers et al 2016) and 0.005% (TO) for risk-stratification of disease deterioration in patients with liver cirrhosis (Jansen et al 2018).

The following studies performed ROC analysis to determine optimal cutoffs: Schaeffer et al identified 3.95 ms/RR to be the optimal cut-off to stratify risk of cardiac events in patients with Marfan syndrome (Schaeffer et al 2015). Perki et al reported TS to predict HF hospitalization in postinfarctional patients with the cut-off 1.29 ms/RR (Perki et al 2010), while TS failed to predict cardiac events in the same population with the cut-off 1.3 ms/RR (Perki et al 2008). Huikuri et al found 2.0 ms/RR as cut-off, with which TS was predictive for arrhythmic events in postinfarctional patients (Huikuri et al 2010). Balcioğlu et al suggested 3.32 ms/RR to detect cardiac autonomic neuropathy in patients with diabetes mellitus (Balcioğlu et al 2007). Karakurt et al used 1.2 ms/RR for TS and found a significant correlation between TS and mortality rate in children with DCM (Karakurt et al 2007). In patients with acute decompensated HF TS with a cut-off of 1.695 ms/RR was predictive for cardiac events (Yamada et al 2018). Yuan et al did not find any prognostic power of TO and TS with the cut-off values -1.17% and 12.1 ms/RR, respectively, in patients with acute coronary syndrome (ACS) (Yuan et al 2015).

Some studies compared their suggested cut-offs with the original values and showed a similar performance of both cut-off types: Schmidt's cut-off values as well as quartiles (0.025% and 1.27 ms/RR) and continuous values were all useful to predict all-cause mortality in patients with CHF caused by ischemic cardiomyopathy (Cygankiewicz et al 2006). The quartiles 0.22% for TO and 1.42 ms/RR for TS were used for CHF patients, but no difference was found compared to the results using Schmidt's cut-off values (Cygankiewicz et al 2008). Quartiles 0.31% and 1.5 ms/RR were used for risk stratification in MI patients, but showed no significant results (Berkowitsch et al 2004). A cut-off of 3.2 ms/RR for TS was proposed in postinfarctional patients with malignant ventricular arrhythmias, though its performance was similar to Schmidt's cut-off regarding sensitivity and specificity (Szydlo et al 2011).

Some new cut-off values yielded better results than Schmidt's values. In the Cardiovascular Health Study, investigating cardiovascular disease in older adults, the cut-off 3.0 ms/RR for TS was used (Patel et al 2017, Kop et al 2010, Stein et al 2011, Stein et al 2010, Stein et al 2008). In contrast to Schmidt's settings, this value showed a significant correlation between cardiac death and TS (Stein et al 2011). The predicitve capability of TS with the new cut-off has also been shown in other studies as well, but without comparison to the traditional values (Stein et al 2010, Koyama et al 2002). A comparison of the cut-off values by Stein et al and Schmidt et al with data from the Eplerenone Post-Acute Myocardial Infarction Heart Failure Efficacy and Survival Study showed that Stein's TS cut-off had a slightly higher sensitivity but slightly lower specificity than Schmidt's value (Stein et al 2009). Other cut-offs showing improved performance were -1.52% (4th decile) and 4.9 ms/RR (6th decile). In contrast to Schmidt's cut-offs, these values allowed HRT parameters to be independent predictors for all cause and cardiac mortality in patients with unstable angina pectoris (Lanza et al 2009).

It is difficult to find a single cut-off value which is suitable for different datasets, even with the same patient background. Only three new cut-off values have been used in more than one study (see tables 5 and 6). However, most of these studies are either based on the same dataset (Stein et al 2008, Patel et al 2017, Stein et al 2011, Kop et al 2010) or do not compare the new cut-off with Schmidt's original values (Cygankiewicz et al 2004, Cygankiewicz et al 2004, Cygankiewicz et al 2003, Cygankiewicz et al 2003).

Different populations: It is important to notice that cut-off values have first been set for postinfarction patients (Schmidt et al 1999). In the following, HRT has been used for patients with other pathological background without setting new cut-off values. For example, the standard cut-offs were used in chronic obstructive pulmonary disease (Gunduz et al 2009), diabetes mellitus (Lin et al 2017, Balcioğlu et al 2007) or metabolic syndrome (Erdem et al 2012, Yilmaz et al 2006, Balcioğlu et al 2016), even apparently healthy subjects (Schwab et al 2004, Poręba et al 2011, Grimm et al 2003).

Additionally, Yilmaz et al analysed children, while Schmidt et al analysed elderly patients (Yilmaz et al 2006, Schmidt et al 1999). Due to the age-dependency of HRT, however, it can be expected that other values should be used for different age groups (Schwab et al 2005). This is done by Stein et al, because the population's age was higher than the population used for Schmidt's cut-offs. A new cut-off 3.0 ms/RR for TS was determined (Stein et al 2008). This is interesting, since heart rate turbulence is decreased with increasing age (Schwab et al 2005), thus a lower cut-off would have been expected in elderly people.

The cut-off values are also used in the context of medical treatments, e.g. to show an effect of elective gynecologic surgery on autonomic functions (Tekelioglu et al 2013) or compare different types of hemodialysis (Kaplan et al 2016). Since HRT is a feasible marker for autonomic functionality, it might be used for these purposes, but specialised cut-offs should be determined for every population.

Number of false-positives: Although the standard cut-off values for high-risk patients after infarction were established on the basis of a large population, there are indications that these values still need optimisation. Some studies have reported false-positives, when examining patients indicated for assessment of supraventricular arrhythmias (Vukajlovic et al 2006) or apparently healthy subjects (Grimm et al 2003, Schwab et al 2004). This especially relates to TO. Therefore, Grimm et al even recommended that only TS is to be used, because TO is not specific enough (Grimm et al 2003).

Customised cut-offs: Instead of using a fixed value, a formula could determine an adaptive cut-off value. This would adjust the cut-off value to the background of the given dataset, avoiding the necessity to determine a great amount of cut-off values for different populations and use cases. However, this approach has not been implemented in the literature and is impractical for risk stratification in the medical routine.

Based on the cut-offs, subjects are categorised as having distinctive HRT (with low risk of adverse outcome) or having impaired HRT (with high risk). The most common categorisation system is defined by Ghuran et al composed of the three categories: (1) TO and TS normal, (2) either parameter abnormal, and (3) both parameters abnormal (Ghuran et al 2002). The categories are generally called HRT0-2 and are used for dichotomisation in 23 inspected articles (Maeda et al 2009, Li et al 2012, Yin et al 2014, Kaplan et al 2016, Soguero et al 2018, Uzna et al 2018, Lewek et al 2009, Balcioğlu et al 2015, Schaeffer et al 2015, Yoshihisa et al 2014, Cygankiewicz et al 2008, Cygankiewicz et al 2013, Park et al 2014, Trzos et al 2008, Bauer et al 2006, Poręba et al 2014, Gać et al 2014, Suzuki et al 2012, Bauer et al 2005, Berkowitsch et al 2004, Secemsky et al 2011, Stein et al 2009, Erdem et al 2012).

Simpler classifications are: 'HRT positive' (both parameters abnormal) or 'HRT negative' (one or both normal) (Miwa et al 2009, Miwa et al 2011, Miwa et al 2012, Ikeda et al 2011, Kawasaki et al 2015); 'HRT' (both parameters normal) and 'non-HRT' (one or both abnormal) (Roach et al 2002); 'HRT abnormal' and 'HRT normal' with either the same definition (Hayano et al 2011) or already classifying subjects with one abnormal parameter as 'HRT abnormal' (Cetin et al 2014, Park et al 2014).

Stein et al and D'Addio et al used categories HRT0-2, but differentiated HRT1 into 'TO abnormal' and 'TS abnormal' (Stein et al 2011) or HRT1a (TO abnormal) and HRT1b (TS abnormal) (D'Addio et al 2013), respectively. One system was introduced that considers the number of VPCs: Carney et al suggested the two new classes 'a' with less than 5 VPCs and 'b' with a sufficient number of VPCs but not calculable because of artifacts etc The standard classes HRT0-2 are called c-e, respectively (Carney et al 2007). It should be investigated, whether a simplification of the classification system yields sufficient results or a differentiation provides added value. Only Barthel et al reported analysing the discriminative power of the used categories, resulting in combining HRT0 and HRT$\oslash$ (Barthel et al 2003).

When not only TO and TS but also TT are analysed, the three categories A (all normal), B (at least one abnormal) and C (all abnormal) have been used (Średniawa et al 2016, Cebula et al 2012). The classification into HRT0-2 displays HRT at one point in time. To show HRT changes, Kurpesa et al grouped patients into the categories type I (improvement of both parameters), II (worsening of both) and III (improvement and worsening of one, respectively) (Kurpesa et al 2007).

If subjects are dichotomised, Schmidt's cut-off values are mostly used. Different cut-offs for several populations have been proposed with equal or better results than the original cut-offs, but none can clearly be recommended. The cut-off 0% of TO leads to more values being classified abnormal and therefore a higher false-positive rate than the cut-off 2.5 ms/RR of TS. Cut-off values should be reviewed with the result of either new cut-off values for different population backgrounds like medical condition and age or mathematical formulas for adaptive cut-off calculations.

On the basis of these cut-offs subjects can be categorised into different risk classification systems. The most common is a classification into the three groups HRT0 (both parameters normal, low risk), HRT1 (one parameter normal, intermediate risk) and HRT2 (both parameters abnormal, high risk.) Other systems consider subjects that fall short of suitable VPCSs, the new parameter TT or HRT development between two points in time. The usefulness of these classification systems with more or less groups than HRT0-2 should be studied. If subjects are categorised into groups, the handling of subjects with an insufficient number of VPCSs should be described.

Research questions:

• Which cut-off values are optimal for which population and question?
• Is there an alternative for hard cut-off values?
• Are other classification systems more suitable than the standard HRT0-2?
• Are the prognoses of subjects in categories HRT0 and HRT$\oslash$ comparable?

Some articles provide no recording duration at all (Yang et al 2013, Sandberg et al 2014, Lenis et al 2013, Lenis et al 2013). As mentioned, most articles only state the planned recording duration (see section 2.1), not the actual recording duration that could help estimate the bias on TS. Some authors report the average of the actual recording duration as the arithmetic mean (Tuomainen et al 2005, Balcioğlu et al 2007, Lammers et al 2006, Kowalewski et al 2007, Bienias et al 2015, Liu et al 2014, Secemsky et al 2011, Balcioğlu et al 2015, Lindgren et al 2003), others as the median (Balcioğlu et al 2016, Balcioğlu et al 2016, Bienias et al 2010a, Bienias et al 2010b, Wichterle et al 2004, Bienias et al 2010c, Bonnemeier et al 2003, Ortak et al 2005, Zhong et al 2007, Wichterle et al 2004, Bonnemeier et al 2005). Some articles of the same first authors report different types of means (Balcioğlu et al 2015, Balcioğlu et al 2016, Balcioğlu et al 2007, Balcioğlu et al 2016, Bienias et al 2010a, Bienias et al 2010b, Bienias et al 2010c, Bienias et al 2015).

As with recording duration, the arithmetic mean of the number of VPCSs actually used for HRT calculation is rarely provided. If a number is mentioned, most of the times it is not transparent whether the number includes different kinds of ectopic beats, all VPCs or just VPCs used for analysis. Additionally, the number may be absolute or extrapolated for 24 h. 46 articles give the mean of VPCs per hour or day, partly even before filtering out patients without suitable VPCs, while 129 articles give no number of VPCs at all. Only in 27 of the 240 articles studied the mean of VPCs in the recording per patient is stated absolutely as either arithmetic mean (17), median (9) or both (1). Again, the majority of these articles report the overall number of VPCs, not the number of VPCSs suitable for HRT calculation, or the number is not clearly described. In contrast, Wichterle et al stated the numbers of VPCs and APCs, both recorded and suitable for analysis, respectively, listing four medians (Wichterle et al 2004).

In contrast to the calculation methods, most articles explain their filtering criteria either very briefly or not at all. Examples include vague descriptions such as 'very short or long cycle lengths' (Cano et al 2008), no 'too long' couplI or 'too short' compI (Manzano-Fernández et al 2011) or no 'inappropriate RR intervals' (Manzano-Fernández et al 2011). Some articles describe filtering only as 'manual review' (Balcioğlu et al 2007, Arslan et al 2008, Ozdemir et al 2007) or 'visual examination' (Stein et al 2010).

The refI is often defined ambiguously, for example as '5 consecutive sinus intervals' (Golukhova et al 2016), 'normal RR-interval' (Jochum et al 2012), 'the sinus RR interval' (Jeron et al 2003) or 'normal interval' (La et al 2011, La et al 2012), not giving the range on which basis refI is calculated.

Some articles give as filtering reference the HRT analysis software HRT View or the algorithm from www.h-r-t.org/.com ⁹ , but their reported values of the number of intervals used for filtering differ from the default values of the software tools (2 preRRs and 15 postRRs) (Wustmann et al 2009, Dursun et al 2015, Celik et al 2011, Karakurt et al 2007, Yorgun et al 2012, Celik et al 2011b, Schwab et al 2011a, Celik et al 2012, Grimm et al 2003, Iwasa et al 2005).

As mentioned, #TSRR is often explicitly stated or given with a calculation reference. However, of the articles stating this parameter explicitly, 43 studies use a different range than the respective calculation reference. In three articles the range 15 to 20 intervals is given without further explanation which number is used in which case (Szymanowska et al 2008, Soguero et al 2013, Średniawa et al 2016).

Mostly the actual recording duration is not provided in the literature, but only the planned recording duration. Descriptions differ between the arithmetic mean and the median. Similarly, the number of VPCSs used for calculation is mostly missing. Since TS is correlated with the number of VPCs and therefore the length of the recording, comparability between studies is only given if the mean of the actual recording duration – or better the number of actually used VPCSs – is stated.

Filtering criteria are usually not or only partially given. The #TSRR as all calculation steps is usually provided, however, many studies use a different range than the respective reference for the calculation workflow. Unclear definitions and therefore varying filtering criteria lead to a selection of different sets of VPCSs out of the same data and consequently to different results. Of course references can be given for the used workflow, which creates the need to accurately carry out HRT assessment as stated there.