In trying to make deductions about our patients’ state of health, doctors have moved to looking more at changes in physiological variables than at absolute values. Cyclical changes can reflect the patient’s current state, whereas time trends, because they require a longer observation period, can be affected by changes in the patient’s condition during measurement. Furthermore, cyclical changes can be captured during a single visit to the patient’s bedside. A familiar example is the fluctuation in the amplitude of the direct arterial pressure trace that occurs in time with respiration (the ‘respiratory wave’) and the fact that this fluctuation increases in hypovolaemia. Automated analysis of the trace can report a number of measures, including variation in systolic pressure and pulse pressure. This editorial primarily discusses heart rate variability (HRV) because automated analysis of this has simultaneously become both easier to do and much more mathematically complex, but many of the comments apply equally to studies of blood pressure variation (BPV), baroreceptor sensitivity (alterations in heart rate in response to changes in blood pressure), heart rate turbulence (transient rise and then fall in heart rate after a premature ventricular contraction) and QT-interval variability. Also of interest is the Surgical Stress Index (SSI) [1], a composite of heart rate and pulse pressure (measured non-invasively, as optical plethysmography). This is a non-invasive way to measure responses to surgical stimulation that can give the user some insight into autonomic function. The umbrella term ‘cardiovascular variability’ (CVV) encompasses all of these. Another umbrella term, ‘autonomic variability’, implies measurement of changes in the autonomic nervous system, either non-invasively, such as measurement of skin conductance [2] or that part of CVV that is attributable to autonomic function, or invasively, such as with intra-neural electrodes. In the case of CVV studies, I suggest that authors should be discouraged from using the term ‘autonomic variability’ unless their work differentiates between autonomic and non-autonomic causes of CVV (see below). Here, discussion of ‘variability’ refers predominantly to changes that are repeated regularly so that the frequency of the change, as well as the change itself, can be analysed. Some comments will also apply to SSI and skin conductance studies, even though time-trend analysis of these parameters is the norm. In 1965, Hon and Lee reported that in early fetal distress, changes in variability in the fetal heartbeat (recorded electronically and continuously) preceded changes in overall rate (sampled stethoscopically and intermittently) [3]. In that example, the variability was cyclical in time with uterine contraction. Cyclical changes in heart rate can be investigated using an ECG recording taken over a number of minutes, from which instantaneous heart rate can be derived by measuring the R-R interval. The rate of change of heart rate can be calculated for each beat by comparing it with the preceding beat. Machines that automatically analyse HRV have become readily available and easy to use. This has resulted in a huge addition to the scientific literature, starting in the world of physiology, and spreading into internal medicine, and now increasingly into the anaesthetic literature. Early techniques of HRV analysis used statistical methods such as calculation of the mean and the standard deviation of the instantaneous heart rate or R-R interval, or simply plotted a graph of the heart rate, so that regular changes in rate could be seen. These approaches do not yield results according to the relative contributions at different frequencies, and are called time-domain techniques. Many time-domain techniques, of varying complexity, have been developed [4]. Frequency-domain techniques yield the relative contributions of influences at different frequencies to the overall variability. Parametric frequency-domain techniques make the assumption of Gaussian distributions of variability, but since this should not be assumed, non-parametric frequency-domain techniques have also been developed. Recently-developed mathematical concepts such as non-linear dynamics and chaos theory have spawned yet more techniques of HRV analysis, but there is no evidence that these are superior to older techniques. Even before the analysis stage, there are multiple ways to ‘pre-process’ an ECG trace into a tachogram, a graph or table of the timing of heartbeats or the interval between them [5]. Figure 1 shows a tachogram, adapted from a study by Scheffer et al. [6], with instantaneous heart rate on the vertical axis and ‘beat number’ (labelled ‘intervals’) on the horizontal axis (this is not the same as using time on the horizontal axis: different beat-to-beat intervals are ascribed the same horizontal distance). Heart rate tachogram (adapted from Scheffer et al. [6]). Non-parametric frequency-domain methods that use Fourier transformation have come to predominate, because it was the technique that was achieving maturity at the time that many manufacturers realised there was a great market for automated HRV analysis machines. The output is often termed a ‘power spectrum’ because the technique originated from the analysis of electrical signals where the output was expressed in watts. Heart rate variability gives the user two measurements: ‘high frequency’, around 0.25 Hz; and ‘low frequency’, around 0.1 Hz. The high-frequency measurement represents parasympathetic activity, whilst the low-frequency measurement is taken to represent sympathetic activity. These correlations are based on human and animal studies using more invasive measurements of autonomic activity, such as intra-neural needle electrodes to measure activity in the vagus nerve and the sympathetic fibres of nerves (often the peroneal nerve), comparisons with and without maximal blockade of either the sympathetic or parasympathetic system, measurements of plasma catecholamines, or tracking of radiolabelled catecholamines. It is widely accepted that high frequency HRV is influenced by parasympathetic activity, but the evidence that low frequency HRV represents sympathetic activity is less strong [4, 7, 8]. There is, however, a resonance with clinical experience. High-frequency power seems to provide automated quantification of ‘sinus arrhythmia’, the frequently-seen vagally-mediated phenomenon of a rising and falling heart rate in time with respiration. Low-frequency power seems to be automated quantification of the heart rate response to variations in blood pressure that occur with a cycling time of around 10 seconds. These oscillations in blood pressure are called Mayer waves, and are thought to represent control of blood pressure with a slight delay between detection and response, a ‘hunting’ phenomenon comparable to fluctuation of a diesel engine’s speed around its set idling speed, or the fluctuating temperature of a room as a thermostat clicks on and off. Most (but not all) of this blood pressure control is neurological and involves the sympathetic system [9]. The simplicity of performing HRV analysis using modern machines belies the complexity of the mathematical processing within. We are accustomed to the ability of computers to do fantastic tasks on our behalf, but the problem is that many steps along the processing path require user choices driven by an understanding of the technology and its limitations. Many machines are programmed with default choices that may or may not be appropriate for a given situation, and they still return to the user what appears to be a simple measurement. In 1996, a Task Force, set up by the European Society of Cardiology and the North American Society of Pacing and Electrophysiology, defined the terminology and methods of HRV and usefully summarised its many areas of uncertainty [4]. One criticism of using HRV to study the balance between sympathetic and parasympathetic tone is that the same information may be contained in the heart rate [10]. However, it is known that the rate of firing of the vagus nerve varies with respiration, and supporters of the technique argue that it is this variability that is being measured, whereas heart rate reflects average autonomic tone and not moment-by-moment fluctuation [11]. Another problem is that some studies have demonstrated high-frequency effects of sympathetic manipulations, and low-frequency effects of parasympathetic manipulations. This has led to further mathematical manipulations (‘normalisation’), for example calculating the ratio of the absolute values to each other, or to ‘total power’ (which is not total in a strict sense as its measurement requires an arbitrarily chosen, low-frequency cut-off). In judging the reliability of a study based on HRV analysis, important considerations include the techniques that were chosen to filter out artefacts and ectopic beats, and for the analysis itself. If these choices were made on the basis of availability and ease of use, rather than consideration of the theoretical advantages of one technique over the others, the authors will have put themselves at risk of both type-1 and type-2 errors. In many cases, even the experts cannot advise about which analysis technique to use. Studies that have used two disparate analysis methods, and obtained the same results for each, will carry more weight than those that only used one method. In research in this area, two groups of studies seem to predominate: those seeking to measure activity in the sympathetic and parasympathetic systems from a physiological/pathophysiological point of view; and those (mainly CVV studies and some skin conductance studies) seeking to identify patients at higher risk of complications and/or predict their outcomes. The first of these groups uses non-invasive measurement of autonomic variability to explore the functions of the sympathetic and parasympathetic systems, and also to investigate suspected pathological states of these in various physical and psychological diseases. However, in clinical and experimental situations where autonomic influence has been partly or completely removed (denervated dog models, patients with high spinal cord transection or who have received transplanted hearts, human volunteers and animals that have been given double autonomic blockade with maximal doses of anticholinergic and adrenergic-blocking drugs), CVV still occurs. With regard to BPV, as mentioned above, Mayer waves can still occur [9], and pronounced respiratory waves are common: in anaesthetic practice they are frequently seen in tetraplegic patients receiving positive pressure ventilation. With regard to HRV, the heart’s intrinsic pacemaker is sensitive to changes in right atrial load, which fluctuates with respiration [12]. Studies that truly measure autonomic variability should be designed to distinguish it from other causes of CVV. There are few examples of this. One study measured HRV before and after double autonomic blockade in patients with atrial flutter [13], and there are studies using simultaneous measurement of SSI and skin conductance [2], and of BPV and plasma noradrenaline [14]. The second group of studies comprises many papers. Correlations have been reported between various measures of CVV and severity, risk of complications, or prognosis, in a large number of diseases. These include ischaemic heart disease, heart failure (with normal or reduced ejection fraction), diabetes mellitus, chronic renal failure, irritable bowel syndrome, postural tachycardia syndrome (formerly called vasovagal syndrome), panic disorder, depression, tobacco smoking, and various forms of critical illness including multiple organ dysfunction syndrome, septic shock and severe neurosurgical diseases. Examples published recently in this Journal have looked at prediction of fluid responsiveness [15, 16] and the use of CVV in the anaesthetic pre-operative assessment clinic [17]. Importantly, these studies need not concern themselves with the mechanism of the variation. However, the indicators they report are in competition with a large number of existing prognostic indicators and scoring systems. It is scientifically valid to develop new indicators and to show that they reliably predict outcomes. Subsequent comparisons with more established indicators, and considerations of cost and convenience, will determine whether they find their way into clinical practice. It is easy to confuse the two types of study, fallaciously concluding that an association between CVV and outcome demonstrates that abnormalities in the autonomic nervous system are worsening disease states. Papers appearing in the anaesthetic literature can be expected mostly to fall into the same two groups, the first perhaps offering insights into changes in autonomic function with particular anaesthetic techniques, and the second seeking to identify patients at particular risk of complications or bad outcomes from anaesthesia and surgery. One study, comparing SSI-guided analgesia with standard anaesthetic practice, used unwanted changes in heart rate and blood pressure as outcomes [1], these being surrogates for the risk of complications such as stroke, myocardial infarction or death. Several conclusions can be drawn: it is preferable to measure HRV in more than one way in studies of autonomic function and perhaps also in studies looking at outcomes; with the current uncertainty about the ability of HRV to predict clinically relevant outcomes, ‘improvement’ in HRV should not be used as a surrogate outcome measure; studies purporting to show that CVV effects were caused by autonomic influences should include additional supporting evidence (e.g. intra-neural electrode measurements, plasma catecholamine levels or CVV measurements taken before and after abolishing autonomic influences); and authors using CVV to identify at-risk groups of patients should be careful not to make inferences about autonomic influences unless their study design justifies these. No external funding and no competing interests declared.