Heartbeats

A plugin for Alakazam: heartbeat-evoked potentials

Tools > 1. Preprocessing > Heartbeats · continuous → continuous · Heartbeats

Every heartbeat as an event: the R-peaks of an ECG, found exactly as spectHR finds them (rpeak.py), CARSPAN’s R-peak detection (Mulder et al. 2009), each becoming an event at its own sub-sample latency. The ECG can be a channel of the recording, a stream of the XDF file it was read from, or a recording made by another device that received the same triggers (Where the ECG comes from); only the beats come over, never the ECG itself. A check on the heart’s own field in the EEG confirms that the beats are on time (The check). The data are not changed. With the heartbeats as events, the rest of an analysis of heartbeat-evoked potentials is Alakazam’s own: DefineBins chooses the heartbeats per condition, Average gives one heartbeat-evoked potential per bin, and ERP Measure, grand averages and the statistics take it from there.

The heartbeat-evoked potential is the EEG averaged time-locked to the R-wave of the ECG (Schandry et al. 1986), studied as a cortical response to the heartbeat and a measure of interoception (Park and Blanke 2019; Coll et al. 2021).

Table 1: The options of Heartbeats.

Option	Default	Meaning
ECG from	A channel of this recording	Where the ECG is: A channel of this recording, A stream in this recording’s XDF file, or Another recording, on shared triggers.
ECG channel	the channel typed ECG, else one labelled ECG or EKG	For a channel of this recording: the channel the R-peaks are found in.
Stream	blank	For an XDF stream: the ECG stream’s name; blank for the first stream typed ECG, else one whose name says ECG or EKG.
Its channel (stream)	blank	The stream’s ECG channel, by label or number; blank for one labelled ECG or EKG, else its first.
Recording	blank	For another recording: its file, a path or a name in this recording’s folder; blank for the one file beside this one with its name and another format Alakazam reads.
Its ECG channel (recording)	blank	The other recording’s ECG channel, by label; blank for one typed or labelled ECG or EKG, else its only channel.
Shared triggers	blank	The codes of the triggers both recordings received, e.g. 1 2 3; blank for every code both have.
Largest trigger misfit	2 ms	How far a paired trigger may lie from the line that carries one clock onto the other (at least one EEG sample).
ECG polarity	Automatic	Automatic decides as spectHR does (below); Upright or Inverted says so instead. An inverted ECG is turned upright to search it, not in the data.
Shortest interval	300 ms	Of two peaks closer than this, the higher stays: spectHR’s refractory period, at most 200 beats a minute.
High-pass before detection	0.5 Hz	Removes baseline wander before the search, as spectHR does (0 for none). Only the copy searched is filtered.
Event code	heartbeat	The code the heartbeats get.
Check on the heart’s field in the EEG	on	Finds where the heart’s field in the EEG peaks relative to the beats (The check).
Move the beats by the lag found	off	Moves every beat onto the heart’s field by the lag the check found.

How the R-peaks are found

CARSPAN’s R-peak detection (Mulder et al. 2009), as spectHR implements it, step by step, ported to MATLAB:

1. Polarity. The middle third of the channel, less its median, is band-passed from 5 to 20 Hz, where the QRS complexes dominate. Tall upward R-waves make its skewness positive, so a negative skewness means an inverted ECG. Only where the skewness is too small to trust (below 0.1 either way) the summed prominence of the upward peaks is weighed against that of the downward ones.

2. High-pass. A zero-phase Butterworth high-pass of order 4 at 0.5 Hz removes the baseline wander.

3. Peaks. A peak is a local maximum at least the median plus 1.5 standard deviations of the signal. Of peaks closer than the shortest interval, the higher one stays.

4. Timing between samples. The top of an R-wave seldom falls exactly on a sample. With $x_0$ the peak sample and $x_{-1}$, $x_{+1}$ the samples before and after it, the rise into the peak is $a = x_0 - x_{-1}$ and the fall after it $b = x_0 - x_{+1}$ (neither is negative at a local maximum). The peak’s time is moved by

   $$\delta = \frac{a - b}{2\,\max(a, b)}$$

   sample periods. This is the apex of a symmetric triangle through the three samples: a straight line through the peak sample and its lower neighbour, along the steeper flank, and a line of the opposite slope through the higher neighbour meet at $\delta$. A peak with equal neighbours stays where it is ($a = b$, $\delta = 0$); the higher the next sample is relative to the previous one, the later the peak, up to half a sample when the next sample is as high as the peak itself ($b = 0$), and likewise earlier. Where both flanks are flat, spectHR divides by $10^{-12}$ instead of 0 and the peak stays put.

The same R-peaks as spectHR. On spectHR’s own example ECG (five minutes at 1000 Hz), spectHR and this plugin find the same 451 beats on the same samples, their times within a ten-thousandth of a microsecond; the same holds at 500 Hz, the rate of most EEG. The detector itself, given the same signal, gives spectHR’s threshold and times to the last bit, its sums added in numpy’s order. The shortest interval is counted in samples as spectHR counts it: the rate is derived from the mean time step, and 300 ms times that rate is truncated to whole samples. spectHR reads its example recording’s rate as 999.9999998 Hz, a hair under 1000, so there 300 ms is 299 samples, here as in spectHR; at a rate of exactly 1000 Hz it is 300. The filters agree with scipy’s to a few parts in ten million, not to the bit: numpy multiplies complex numbers with fused multiply-adds where the processor has them, which MATLAB cannot copy, and which makes spectHR’s own filters differ in the last bit from one machine to another. The plugin’s tests hold every step to spectHR’s own results (SpectHRParityTest).

Where the ECG comes from

Only the R-peaks are needed, never the ECG itself, so an ECG recorded apart from the EEG serves as well as a channel of it, once its clock is lined up with the EEG’s. Each R-peak is found in the ECG’s own samples, at its own rate, and placed in the EEG at the fraction of a sample its time falls at.

A channel of this recording. The ECG recorded by the EEG amplifier, on the EEG’s own samples: nothing to line up.

A stream in this recording’s XDF file. With Lab Streaming Layer (Kothe et al. 2025) the ECG often comes as a stream of its own, from a chest strap or a separate amplifier, saved in the same XDF file as the EEG; Alakazam reads only the EEG stream. Heartbeats reads the file again, as EEGLAB’s xdfimport reads it (load_xdf with its defaults): every stream’s time stamps on the recording computer’s clock, each stream’s clock offsets fitted with their drift, the jitter removed. The R-peaks are found in the ECG stream, timed by its own time stamps, and each is placed between the two EEG samples whose time stamps surround it. (EEGLAB’s import puts a marker of another stream on the nearest sample; a beat keeps its fraction.) A beat outside the EEG stream, or in a gap in it where samples were lost, has no place: it is left out and counted.

The recording must still be the EEG stream as it was read, every sample at its rate, so run Heartbeats before Resample or SelectData.

Another recording, on shared triggers. When the ECG was recorded by another device that received the same triggers as the EEG, the two clocks are lined up on those triggers, as eye tracking and EEG are (Dimigen et al. 2011). The other recording is the file named, or else the one file beside this one with the same name and another format Alakazam reads (sub01.vhdr’s sub01.edf). Codes are compared without spaces, without BrainVision’s S and without leading zeros, so S 1 is 1.

Two quartz clocks drift apart by tens of parts per million, tens of milliseconds an hour, so the line that carries the other recording’s time onto the EEG’s, $t_\text{EEG} = a + b\,t_\text{other}$, has both an offset $a$ and a rate ratio $b$. It is found in steps that triggers missing from either recording, or extra in one, do not upset:

1. A vote. Every pair of same-code triggers, one from each recording, proposes an offset for each rate ratio tried (within 300 parts per million of 1). The true line gathers the right pairs in one narrow bin; the wrong pairs scatter. Trying the ratios is what lines up triggers far apart, a block’s onset every ten minutes, say, over which the drift adds up to more than the pairing allows.
2. Pairs. Each trigger of the EEG is paired with the trigger of the same code that the line puts nearest it, within 20 ms, one to one.
3. The line, fitted to the pairs with Huber’s loss (Huber 1964), so that one trigger timed late counts for less.
4. Again from step 2 with the new line, the tolerance narrowed to six robust standard deviations of the misfit (2 ms at least), until the pairs no longer change.

Heartbeats refuses, rather than place beats on a doubtful line, when fewer than three triggers pair; when the triggers are too regular to tell one offset from another (one code at a fixed interval fits an offset by a whole period nearly as well: the second best vote is more than 0.8 of the best); or when a paired trigger lies further from the line than Largest trigger misfit. The R-peaks are found in the other recording and carried along the line, each to its fraction of a sample, their intervals on the EEG’s clock.

One line carries one clock onto another, so neither recording may have cuts: run Heartbeats before SelectData.

What it records. EEG.etc.alz.heartbeats.sync holds how the clocks were lined up: for an XDF stream, the file, the streams, the channel and rates, and the beats found and left without a place; for another recording, the file, the codes, the pairs and the unpaired triggers on either side, the misfit (RMS and largest), the drift and the offset. The console says the same.

The check on the heart’s field

Clocks lined up are not yet beats on time. A device can deliver its samples late, a wireless one especially, and the delay within a device is one Lab Streaming Layer cannot measure or correct (Kothe et al. 2025).

The EEG itself shows when the heart beat. The heart’s electrical field reaches every scalp electrode, largest during the QRS complex (Dirlich et al. 1997), at the moment of the R-wave. Heartbeats averages the scalp channels at the beats, from 200 ms before each to 200 ms after, less each channel’s mean from 200 to 120 ms before, and finds where the field’s strength (its global field power: the standard deviation across the channels) peaks within 100 ms of the beats, to a fraction of a sample. With the beats on time, that is at 0 ms; where it is instead is the timing’s error from end to end, whatever its cause. The field counts as seen when that peak is at least three times the median power before the beats, from at least 20 beats and two scalp channels; otherwise nothing is concluded.

The lag is recorded in EEG.etc.alz.heartbeats.check and reported, with a warning when it is more than 10 ms. Move the beats by the lag found moves every beat onto the field; a beat moved past either end of the recording is left out and counted. The QRS of the scalp field and the R-wave of one ECG lead need not peak at quite the same moment, so the check confirms the timing to within a few milliseconds: it catches what is wrong by more.

The check needs the heart’s field in the EEG, so run Heartbeats before the heart components are removed (below).

The events

Each heartbeat becomes an event of the chosen code at its R-peak, in time order among the other events, with two fields:

• ibi, the interval to the next beat in ms (spectHR’s interbeat interval series), and
• hr, the heart rate it makes, in beats per minute.

Both are NaN for the last beat before a cut, a rejected stretch, a gap in the ECG or the end of the ECG. As numeric fields of the events they can sort an ERP image of the heartbeats (The plot area). Heartbeats of the same code already in the recording, from an earlier run, are replaced.

Cuts and rejected stretches. An ECG channel cut by boundary events, or with rejected (NaN) samples, is searched one stretch at a time, each as spectHR would search a recording of its own, so no beat is found across a join and no interval spans one; so is an ECG stream with a gap in its time stamps. The polarity is decided once, on all the valid samples.

The series is taken as it is found: there is no check for missed or extra beats. The R-peaks of a clean ECG are found reliably; where the ECG is noisy, look at the beats in the recording’s view before going on.

From heartbeats to heartbeat-evoked potentials

1. Heartbeats, on the continuous recording as it was read, with the check on: the heart’s field is still in the EEG to check the timing on.

2. Remove the cardiac field artefact. The heart’s own electrical field reaches the scalp, largest during the QRS complex and the T wave (Dirlich et al. 1997), and it is locked to the R-peak just as the heartbeat-evoked potential is, so averaging does not remove it. How studies handle it varies (Coll et al. 2021; Park and Blanke 2019). One way is independent component analysis: in ICA (ICA), sort the components by ICLabel’s Heart probability (Pion-Tonachini et al. 2019) and remove the heart components. The heartbeat events stay.

3. DefineBins (DefineBins) chooses the heartbeats per condition, by what happens around them. The heartbeats in the second after a face (code 1) and after a house (code 2):

   epoch [-200, 600] ms
   bin 1 "Heartbeats after faces"  "heartbeat" any(1) within [-1000, 0] ms
   bin 2 "Heartbeats after houses" "heartbeat" any(2) within [-1000, 0] ms

   A heartbeat near a stimulus carries that stimulus’s own evoked response; to leave those out, exclude the heartbeats with a stimulus inside their epoch:

   bin 3 "Heartbeats away from stimuli" "heartbeat" not any(1|2) within [-200, 600] ms

4. Average (Average) gives one heartbeat-evoked potential per bin, and ERP Measure (ERP Measure) measures it.

Troubleshooting

• “No heartbeat was found.” The channel chosen holds no ECG, or a flat one. Choose the ECG channel.
• Twice as many beats as there should be. Tall T waves can pass the threshold; a longer Shortest interval keeps them out, at the cost of missing beats above the rate it allows.
• The beats sit on the S wave. The polarity was judged wrongly, which a noisy ECG can cause: set ECG polarity to Upright or Inverted.
• “This recording no longer has the time base of its XDF stream.” It was resampled or cut after it was read. Run Heartbeats on the recording as it was read, and Resample or SelectData after.
• “The XDF file has no stream typed ECG or named so.” The message lists the file’s streams: name the ECG stream under Stream.
• “The shared triggers are too regular.” One code at a fixed interval fits an offset by a whole period nearly as well. Choose codes that come at irregular times under Shared triggers, or several codes.
• “The two recordings do not line up well enough.” A paired trigger lies further from the line than Largest trigger misfit: one device timed a trigger late, or a clock did not run steadily. If one device is known to time its triggers less precisely, allow for it there.
• “The heart’s field in the EEG peaks more than 10 ms from the beats.” The ECG lags the EEG (or leads it), most often by the ECG device’s own delay. Move the beats by the lag found puts them on the field.
• “The heart’s field is not clear enough in the EEG.” Too few beats, too few scalp channels, or the heart components already removed: run Heartbeats before ICA removes them.

References

Coll, Michel-Pierre, Hannah Hobson, Geoffrey Bird, and Jennifer Murphy. 2021. ‘Systematic Review and Meta-Analysis of the Relationship Between the Heartbeat-Evoked Potential and Interoception’. Neuroscience and Biobehavioral Reviews 122: 190–200. https://doi.org/10.1016/j.neubiorev.2020.12.012.

Dimigen, Olaf, Werner Sommer, Annette Hohlfeld, Arthur M. Jacobs, and Reinhold Kliegl. 2011. ‘Coregistration of Eye Movements and EEG in Natural Reading: Analyses and Review’. Journal of Experimental Psychology: General 140 (4): 552–72. https://doi.org/10.1037/a0023885.

Dirlich, G., L. Vogl, M. Plaschke, and F. Strian. 1997. ‘Cardiac Field Effects on the EEG’. Electroencephalography and Clinical Neurophysiology 102 (4): 307–15. https://doi.org/10.1016/s0013-4694(96)96506-2.

Huber, Peter J. 1964. ‘Robust Estimation of a Location Parameter’. The Annals of Mathematical Statistics 35 (1): 73–101. https://doi.org/10.1214/aoms/1177703732.

Kothe, Christian, Seyed Yahya Shirazi, Tristan Stenner, et al. 2025. ‘The Lab Streaming Layer for Synchronized Multimodal Recording’. Imaging Neuroscience 3: IMAG.a.136. https://doi.org/10.1162/IMAG.a.136.

Mulder, B., H. Hofstetter, and A. van Roon. 2009. CARSPAN for Windows: User’s Manual. University of Groningen.

Park, Hyeong-Dong, and Olaf Blanke. 2019. ‘Heartbeat-Evoked Cortical Responses: Underlying Mechanisms, Functional Roles, and Methodological Considerations’. NeuroImage 197: 502–11. https://doi.org/10.1016/j.neuroimage.2019.04.081.

Pion-Tonachini, Luca, Ken Kreutz-Delgado, and Scott Makeig. 2019. ‘ICLabel: An Automated Electroencephalographic Independent Component Classifier, Dataset, and Website’. NeuroImage 198: 181–97. https://doi.org/10.1016/j.neuroimage.2019.05.026.

Schandry, R., B. Sparrer, and R. Weitkunat. 1986. ‘From the Heart to the Brain: A Study of Heartbeat Contingent Scalp Potentials’. International Journal of Neuroscience 30 (4): 261–75. https://doi.org/10.3109/00207458608985677.