Traditionally, heart rate (HR) and metrics calculated from it (e.g. heart rate variability; HRV) have been collected via an electrocardiograph (ECG) within a laboratory. This environment, or a similar clinical one, is characterised by good control over the parameters of recording; movement is managed or minimised, equipment can be adjusted, multiple devices or ECG leads can be observed, feedback on signal quality is usually immediately available from a trained technician and analytical standards for recording, filtering, beat-detection, etc. are well established (PAN TOMPKINS).
Recently, advances in miniaturisation have resulted in researchers having access to compact, portable and wireless hardware for physiological recording. These devices offer several advantages over their laboratory-based cousins. First, they allow naturalistic observation of the real phenomena which laboratory observations either attempt to model or provoke (CITE GOODWIN). Second, they allow fieldwork and remote collection to take place with limited support (i.e. without access to wired power). Third, they are inexpensive and accessible. Fourth, and perhaps most importantly, due to all of the above they greatly increase the scientific fidelity of observations which are typically underpowered; instead of taking single laboratory measurements, experimental participants can populate their own within-subject recordings over long periods of time. A typical laboratory recording is 5 minutes, which is a well accepted analytical standard for short-term HRV (TASK FORCE), while a whole day's uninterrupted recording contains 288 5-minute observations.
Take, for instance the example of an experimenter recording single before-and-after measurements of an eight-week intervention. Instead of 2 bookend measurements, a researcher with a portable or lightweight device may receive daily measurements from participants over eight weeks (i.e. 56 within-subjects measurements). Happily, the other traditional benefit of within-subjects design also applies (that measurements are internally consistent without depending on random assortment) while the most commonly cited drawback (that within-subject measurements may be subject to order or learning effects) is minimised in physiological measurement. Simple Monte Carlo models (e.g. Bellemare 2014) show an increase in experimental power in the proportion of an order of magnitude for within- versus between-subjects with as few as 6 internal observations.
The use of the portable/ambulatory electrocardiograph also confers several problems which are not yet fully resolved. Generally well recognised is the fact that baseline noise is increased due to movement, and generally not observed at the time of recording - signal quality is generally not checked in real time, and errors are not detected and/or corrected until recording is completed. However, the issues conferred by the electrocardiogram being recorded between two electrodes which are extremely close together (sometimes no further apart than a few centimeters), termed here the short dipole problem, are unresolved. ECG is generally recorded from two or more electrodes attached to the chest wall or thorax, whose points of comparison generally cross the whole thorax and are on both sides of the myocardium. On a compact device, especially one without attached electrode wires where electrodes attach directly to the device body, this is not possible. The cardiac axis (the general direction of ventricular depolarisation, with reference to the QRS complex) changes quite substantially between individuals, and the orientation of the dipole of observation may change dramatically with the position and orientation of the device. In addition, a compact device may be attached or adjusted by an experimental participant to minimise contact, or may be detached and reattached during longer term out-of-laboratory recording.
The sum total of the above is that the traditionally reliable of the laboratory ECG shows significant heterogeneity not previously recognised in ambulatory devices. The normally dominant QRS complex treated as the canonical representation of each heart beat (and thereby the difference between QRS complexes the representation of the inter-beat interval) is affected. QRS complexes which are generally peaked and easily obtained are instead noise-prone and heterogeneous both within and between people. The typical QRS complex is the representation of the ventricular depolarisation due to the changed relative angle between the cardiac axis and the short dipole of observation means the most reliable feature within the ECG is a non-specific R-wave of variable height and morphology. This is highly problematic to detect or define using traditional measurements which rely entirely on the measurement of isolated peak height.
This paper presents a general analytical pathway we have developed to be effective between subjects when ECG data is noisy and R-waves are heterogeneous.
METHOD
Principles of analysis
semi-automated
peak detection of template match
where is the 'r wave peak' as much as 'what is the time difference between R-wave templates'
1. windowed derivative
As can be seen in Figure 1, R-waves may show significantly heterogeneity but are generally robust to noise.
2. ensemble averaged to a canonical curve
median
3. dynamic time warping (canonical curve vs. enveloped signal)
4. peak detection
5. non-continuous HRV methods
the best supported are RR, SDNN, RMSSD, and spectral analyses. all can be calculated without the assumption of continuous data.
DISCUSSION
This algorithm is not suitable for real-time heart rate processing and therefore, by extension,