Does the brain cope with anxiety by numbing out?
đ Read the Paper
đŒ View the Poster
TL;DR
I used EEG and unsupervised machine learning techniques to analyze how the brain processes internal signals (interoception) and its link to anxiety. I found that the brain is less sensitive to internal signals during high energy states, a pattern that was amplified during states of anxiety—insights from this project are relevant to stress detection and human-centered design.
Key Skills
- đ§č Data Wrangling â Cleaned high-dimensional EEG & physiological data
- đ Exploratory Data Analysis â Examined relationships between physiological and behavioral measures
- đ€ Machine learning â Applied signal separation for artifact removal and clustering to extract insights
- đ Visualization â Created clear, insight-driven plots
What I Learned
How to extract meaningful signals from noisy biological data, a skill transferable to real-world time-series analysis (wearables, health tech, user behavior modeling).
Sometimes life is angsty. We’ve all been there—you’re about to talk in front of a group, have a difficult conversation with someone, or do some task where a mistake would be really bad. While it might be healthier not to avoid anxiety, sometimes it’s just easier to reach for a distraction. What you may not know is that your brain might subconsciously be trying to avoid anxiety as well.
Figure 1: Distraction is an easy way to avoid feeling anxiety.
What we were interested in
We wanted to look at how the brain responds to heartbeats across different states of anxiety and energy
Previous research suggests that anxiety disrupts our brain’s subconsious sensitivity to bodily signals, a sensitivity referred to as interoception. 1 Interoception is our brain’s way of sensing and regulating many of the body’s functions, such as breathing and responding to heart rhythms.
Differences in interoception have also been linked to differences in one’s sense of energy or activation2,3—a sense we’ll refer to as subjective arousal. Subjective arousal—much like anxiety and interoception—is continually and spontaneously fluctuating, even when people are just at rest. But no previous study has looked at how subjective arousal, anxiety, and interoception spontaneously fluctuate with each other at rest.
We wanted to look at how the brain responds to bodily signals—specifically, heartbeats—during spontaneous fluctuations in subjective arousal when people are at rest, and how this interaction might be different for those low vs. high in anxiety. How these processes play out at rest might be important because certain brain networks are typically only active at rest, and these networks contribute to patterns like rumination and worry that feed into anxiety. 4
Understanding how the brain shifts interoception in response to changing subjective arousal states during rest might help us better understand processes that underlie anxiety.
How we did it
We measured electrical activity from the brain and heart during rest and occassionally asked people how they were feeling
A precise way to measure neural interoception is through what’s called the heartbeat evoked potential (HEP), which is basically the brain’s electrophysiological response to heartbeats.5 Your brain is constantly emitting electricity (emanating from the communication between neurons), and when the brain’s electrical response is high following a heartbeat, that’s thought to reflect higher levels of neural interoceptive sensitivity to the heart.
Participants (51) came into our lab and completed a survey indicating their level of state anxiety, or how much anxiety they were experiencing in that moment. We then fitted them with an EEG cap to measure their neural electrical signals. This EEG cap included 31 scalp electrodes plus an electrode that we placed on participants’ back (ECG) to measure their heartbeat activity (see Figure 2).
Figure 2: (A) Cartoonized illustration of EEG cap and (B) placement of EEG electrodes in the experiment according to the 10-20 coordinate system
Participants stared at a blank screen and were told to let their minds wander freely. Every so often, a set of questions (ie, a thought probe) appeared asking participants to report different aspects of their experience. Most important was the question asking about their level of subjective arousal, which includes feelings of energy linked to emotion, and participants gave ratings on a scale from 0 (completely deactivated) to 100 (completely activated; see Figure 3A).
Figure 3: Trial sequence and analysis approach. (A) Participants were instructed to let their minds wander freely while viewing a blank screen and occasionally reported on aspects of their experience (such as their level of subjective arousal). (B) Brain EEG signals before a thought probe and directly after heartbeats were averaged together, separated by a median split in subjective arousal, and submitted to a whole-brain analysis (ie, cluster-based permutation analysis).
To calculate the HEP, I isolated the 10 s period before the onset of a thought probe and extracted the EEG signals (from all 31 electrodes) that occurred directly after a heartbeat (see Figure 3B). I then separated the 10 s pre-probe periods according to whether the subjective arousal rating obtained during the thought probe was high or low (based on a within-participants median split on subjective arousal).
Finally, I averaged across the high vs. low levels of subjective arousal (collapsing across heartbeats and probes), which gave me one averaged brain signal value for timepoints immediately following a heartbeat, for each EEG channel, for each subjective arousal condition, across all participants. In other words, we had for each participant an array of shape (timepoints x channel x condition), or (100 x 31 x 2).
Ultimately, we wanted to know how interoceptive sensitivity to cardiac signals (as measured with HEPs) is different in periods leading up to self-reported high vs. low states of subjective arousal. To assess this, I took those arrays defined in the previous paragraph and conducted a whole brain analysis using what’s called a cluster-based permutation test.
Cluster-based permutation test is a super powerful and flexible data-driven method for finding differences between conditions in high-dimensional space. In our example, the conditions are a median split on subjective arousal ratings, each thought probe was labelled as “high” or “low” subjective arousal. The high dimensional space is the remaining (time x channels) array of EEG data for each participant. In short, we want to find whether EEG voltage is different across high vs. low levels of subjective arousal conditions at any timepoint(s) and for any EEG channel(s).
Expand this section to optionally read more about cluster-based permutation test and see how I implemented it with code.
Technical spotlight: Cluster-based permutation testing
What is a cluster-based permutation test?
We’re trying to find a meaningful pattern in brain data, but brain data is massive. In our case, we’re looking across 6,200 data points. If we tested each one individually, we’d almost certainly get false positives (random noise that looks meaningful).
A cluster-based permutation test helps solve this.
The core idea:
Instead of asking “is this single time point significant?”, we ask:
Are there clusters of nearby data points that are all showing an effect?
If a real difference exists, we expect it to show up not just as a blip, but as a spatially or temporally coherent cluster.
How it works:
- Compute the actual data difference between two groups (eg, high vs. low subjective arousal) at every point in time or space.
- Identify clusters of adjacent points that pass some basic threshold (eg, t-values > 2).
- Calculate a cluster-level statistic (like the sum of t-values) for each cluster.
- Randomly shuffle the group labels many times (eg, 5,000 permutations) and repeat steps 1-3 each time.
- Compare your original clusters to the shuffled ones. If a real cluster is bigger than 95% of the ones you’d see by chance, it’s considered statistically significant.
What makes this approch so powerful is that it doesn’t rely on strict assumptions about what the data would look like if there were no effect (ie, a null distribution). Instead, it builds a null distribution directly from the data. This allows us to compare many data points over an arbitrary number of dimensions to look for significant effects, making cluster-based permutation testing the perfect tool for finding differences in our brain data.
How I implemented it
To make the cluster-based permutation test flexible and reproducible, I wrote a collection of functions in both Python and R and coordinated them with R’s fancy reticulate package. The Python functions actually ran the cluster test using mne.stats.spatio_temporal_cluster_test, and the R functions were mostly for summarizing the result. I’m highlighting here the main Python function doing most of the work, but you can see the full set of functions on GitHub (for Python and R).
The whole main Python function:
from pathlib import Path
import pickle
from mne.stats import spatio_temporal_cluster_test
def permutation_cluster_test(item,
low_anchor,
bads,
time_window_min=0.25,
initial_alpha=0.01,
path=Path('analysis/data/derivatives/hep/06-evoked-clean')):
'''
Conducts a permutation-based clustering analysis across a median split of
item, analyzing time points from time_window_min to end of epoch.
--- PARAMETERS ---
------------------
item (str): Item name
low_anchor (str): Name of low anchor on the scale (all lowercase)
time_window_min (float): Analyze time points after this value
initial_alpha (float): Alpha for finding initial clusters
path (pathlib.Path): Path to directory containing data
--- RETURNS ---
---------------
out (dict) containing results of permutation test:
t_obs: (N_timepoints x M_channels) matrix with t values as elements
clusters: list of (array(time_idx, ...), array(channel_idx, ...)) tuples of all found clusters
p_values: np.array of shape (N_clusters,) where each element is a p value
'''
# Open dictionary
assert(isinstance(item, str))
file = 'eeg_dict_{}.pkl'.format(item)
full_path = path / Path(file)
with open(full_path, 'rb') as f:
dic = pickle.load(f)
# Ensure bads is a list of ints
if isinstance(bads, list):
bads = [int(x) for x in bads]
else:
bads = [int(bads)]
# Get numpy arrays of shape (subjects, time, chans) for each condition
low, high = _format_for_clustering(dic, low_anchor, bads)
# Get a sample evoked object for computing distances
probe_set = dic[list(dic.keys())[0]][low_anchor]
sample_evoked = probe_set[list(probe_set.keys())[0]]
adjacency, _ = find_ch_adjacency(sample_evoked.info, 'eeg')
# Get the first index of timepoint thats >= the min timepoint
time_window_idx = [i for i, e in enumerate(sample_evoked.times) if e >= time_window_min][0]
times = sample_evoked.times[time_window_idx:]
channels = sample_evoked.info['ch_names']
# Format data as list of arrays
X = [low[:, time_window_idx:, :],
high[:, time_window_idx:, :]]
# Configure parameters
df = low.shape[0] - 1
t_crit = stats.t.ppf(1 - initial_alpha, df)
tail = 0
# Run test
t_obs, clusters, p_values, _ = spatio_temporal_cluster_test(
X,
n_permutations=1000,
threshold=t_crit,
tail=tail,
n_jobs=None,
seed = 1510,
buffer_size=None,
adjacency=adjacency,
stat_fun=_my_t
)
out = {'t_obs': t_obs, 'clusters': clusters, 'p_values': p_values,
'times': times, 'channels': channels}
return out
Let’s zoom into the important stuff.
The actual cluster-based permutation test is being run with this code:
# Run test
t_obs, clusters, p_values, _ = spatio_temporal_cluster_test(
X,
n_permutations=1000,
threshold=t_crit,
tail=tail,
n_jobs=None,
seed = 1510,
buffer_size=None,
adjacency=adjacency,
stat_fun=_my_t
)
This function (spatio_temporal_cluster_test) comes from the MNE library, which is an excellent set of EEG/MEG analysis tools written for Python. Let’s step through the important input arguments.
The input data
The input data (X) needs to be formatted as a list of arrays, where each array in the list is data from one of the conditions—in our case, the two subjective arousal conditions (high vs. low). This formatting is done by the following lines of code:
# Get numpy arrays of shape (subjects, time, chans) for each condition
low, high = _format_for_clustering(dic, low_anchor, bads)
# Format data as list of arrays
X = [low[:, time_window_idx:, :],
high[:, time_window_idx:, :]]
I’m first calling one of my other functions (not shown above) to format the relevant data into two arrays split by subjective arousal condition. These arrays represent the data across each timepoint and channel for each subject. Next, I’m concatenating these two arrays into a list—taking only a subset of the time window with time_window_idx—which serves as our final input data X. We’re analyzing only a subset of the time window because we want to conduct our statistics only on those measurements that occur at least 250 ms after each heartbeat, which is to ensure that heartbeat activity doesn’t contaminate the neural activity we’re interested in analyzing.
The statistical hyperparameters
We need to configure a number of statistical hyperparameters that will influence how the cluster test algorithm runs. We first need to decide on what is a significant threshold for forming a cluster in the first place (ie, the threshold input argument). Since we’re only comparing two conditions within participants, the statistical test we use to compute this threshold is a paired-samples t test (which is defined in the _my_t function and passed to the stat_fun input argument of the cluster analysis).
For each of the many, many comparisons across data points, the algorithm computes one t test, and the threshold we set determines whether the resulting t value is considered significant. Significant t values that are adjacent to one another in data space are considered to be a “cluster”. For this type of test, it’s common to use an alpha value (ie, false positive rate) of 0.01 as a cluster forming threshold. But we need to convert that alpha to a t value, and for that we use the quantile function of the t distribution (ie, the function that maps probability values to t values, or the inverse cumulative distribution function):
# Configure parameters
df = low.shape[0] - 1
t_crit = stats.t.ppf(1 - initial_alpha, df)
tail = 0
The degrees of freedom (df) is the number of participants minus 1, and we
subtract alpha from 1 to calculate the cumulative density between the tail of
the t distribution and the alpha value. Finally, tail=0 tells the algorithm we
want to perform a two-tailed test, meaning either condition can have greater EEG
voltage than the other.
The algorithm returns the data points of all the clusters found, along with the t values across all data points, and cluster-specific p values. If a cluster has a cluster-level p value of less than .05, we know there was a significant difference across subjective arousal conditions during the cluster’s time period and for that cluster’s channels. Below (Figure 4) is a heat map of the t values across all data points (with a preview of a signifcant cluster):
# Read in data
m <- readRDS('post_data/cluster_result.rds')
# Parse time and channels, make data frame
channels <- colnames(m$eeg)[which(colnames(m$eeg)=='Fp1'):ncol(m$eeg)]
times <- unique(m$eeg$time)
times <- times[times >= .25]
ts <- as.data.frame(m$result$t_obs)
colnames(ts) <- channels
ts <- cbind(data.frame(time = times, ts))
# Configure colors and plot
blues <- brewer.pal(5, 'Blues')
blues_g <- colorRampPalette(c(blues[1], blues[5]))(100)
ts %>%
gather(channel, t, Fp1:POz) %>%
mutate(channel = factor(channel, levels = rev(channels))) %>%
ggplot(aes(x = time, y = channel)) +
geom_tile(aes(fill = t)) +
geom_rect(aes(ymin = 'C3', ymax = 31.5, xmin = .31, xmax = .38), fill = NA, color = 'green', linetype = 'dashed',
linewidth = 1.5) +
scale_fill_gradientn(colors = blues_g) +
labs(
x = 'Time post heartbeat (s)',
y = 'EEG Channels',
fill = 't value'
) +
theme_bw() +
theme(panel.grid = element_blank(),
axis.ticks = element_blank())
Figure 4: Heat map of t values from cluster-based permutation analysis across time and EEG channels. Time is relative to heartbeat onset. First letter of channel suggests it's position on the scalp (ie, F=frontal, C=central, T=temporal, P=parietal, O=occipital). Darker blues indicate higher t values, revealing a larger difference in the EEG signal between subjective arousal conditions. Green dashed lines highlight the significant cluster.
That’s it for the technical spotlight! Now back to the main post.
Research questions
In this post, I address three questions with the analyses:
- How was our measure of subjective arousal related to physiological arousal?
- How does interoception (as measured with HEP) vary across states of subjective arousal?
- How does the relationship between interoception and subjective arousal change depending on state anxiety?
What we found
1. Subjective arousal was unrelated to physiological arousal and somewhat related to other dimensions of experience
To first get a better sense of what we were really measuring with our subjective arousal scale, I looked at how subjective arousal ratings compared to (i) more traditional measures of objective, physiological arousal (such as heartrate), and (ii) how subjective arousal related to other dimensions of ongoing experience that we measured with our thought probes (such as tendency to think about the future). Figure 5 shows participant-specific correlations between subjective arousal and both heartrate and future thinking.
# Import data
d <- read.csv('post_data/MW_EEG_behavioral.csv')
bads <- c(10, 13, 14)
d <- d[!d$subject %in% bads,]
ecg <- read.csv('post_data/MW_ECG_summary.csv')
# Mean imputation on very few missing values
d <- d %>%
group_by(subject) %>%
# Mean imputation
mutate(trial = 1:(n()), arou = ifelse(!is.na(arou), arou, mean(arou, na.rm=TRUE))) %>%
ungroup() %>%
select(subject, trial, arou, fut)
# Summarize heart metric
heart <- ecg %>%
rename(trial = probe) %>%
mutate(trial = as.integer(gsub('Probe', '', trial)),
timepoint = sample / 250) %>%
mutate(hr = 60 / (timepoint - lag(timepoint))) %>%
group_by(subject, trial) %>%
summarize(hr = mean(hr, na.rm=TRUE)) %>%
inner_join(d)
## `summarise()` has grouped output by 'subject'. You can override using the
## `.groups` argument.
## Joining with `by = join_by(subject, trial)`
# Compute summaries for viz
sd <- heart %>%
gather(metric, measure, hr, fut) %>%
group_by(subject, metric) %>%
summarize(r = cor.test(measure, arou)$estimate,
ci_l = cor.test(measure, arou)$conf.int[1],
ci_h = cor.test(measure, arou)$conf.int[2],
p = cor.test(measure, arou)$p.value) %>%
mutate(sig = case_when(
p < .001 ~ 'p < .001',
p < .01 ~ 'p < .01',
p < .05 ~ 'p < .05',
p >= .05 ~ 'N.S.'
)) %>%
mutate(sig = factor(sig, levels = c('N.S.', 'p < .05', 'p < .01', 'p < .001')))
## `summarise()` has grouped output by 'subject'. You can override using the
## `.groups` argument.
# --- COMPUTE AGGREGATE STATS --- #
# Use Fisher Z to transform correlations to unbounded space
fisherz <- function(r) {
return(0.5 * log((1 + r) / (1 - r)))
}
sd$z <- fisherz(sd$r)
N <- length(unique(sd$subject))
# Aggregate correlation statistics
stats <- sd %>%
group_by(metric) %>%
summarize(m = mean(r), se = sd(r) / sqrt(N), t = t.test(r, mu=0)$statistic,
df = t.test(r, mu=0)$parameter, p = t.test(r, mu=0)$p.value,
bf = (extractBF(ttestBF(r, mu=0))$bf)) %>%
mutate(p_code = ifelse(p < .001, '< .001', paste0('= ', round(p, 3))),
bf = ifelse(metric=='fut', bf, 1 / bf)) %>%
mutate(r = -1, subject = c(1, 24),
label = paste0('p ', p_code, c('\nBF Alt = ', '\nBF Null = '), round(bf, 2)))
metric_order <- c('hr', 'fut')
stats$metric <- factor(stats$metric, levels = metric_order)
stats$metric <- recode(stats$metric, `hr` = 'Heart rate', `fut` = 'Future thinking')
colors <- brewer_pal(palette = 'RdBu')(10)[c(1, 10, 8, 6)]
# Visualize
p3 <- sd %>%
mutate(metric = factor(metric, levels = metric_order)) %>%
mutate(metric = recode(metric, `hr` = 'Heart rate', `fut` = 'Future thinking')) %>%
ggplot(aes(x = reorder_within(subject, r, metric), y = r)) +
geom_hline(yintercept = 0, linetype = 'dashed') +
geom_errorbar(aes(ymin = ci_l, ymax = ci_h, color = sig), width = .2, linewidth = .5) +
geom_point(aes(color = sig), size = .5) +
geom_text(data = stats, aes(label=label), size = 3, hjust = 0) +
scale_color_manual(values = colors) +
labs(
x = 'Participant',
y = 'Within-participant correlation with subjective arousal',
color = ''
) +
scale_x_reordered() +
scale_y_continuous(breaks = c(-1, 0, 1), labels = c(-1, 0, 1), limits = c(-1, 1)) +
coord_flip() +
facet_wrap(~metric, scales = 'free', ncol = 3) +
theme_bw() +
theme(axis.ticks = element_blank(),
axis.text.y = element_blank(),
strip.background = element_rect(fill = NA),
panel.grid = element_blank(),
legend.position = c(.39, .3),
legend.key.size = unit(.2, 'cm'),
legend.text = element_text(size = 8),
legend.spacing.y = unit(0, 'pt'),
legend.title = element_blank(),
text = element_text(size = 14))
## Warning: A numeric `legend.position` argument in `theme()` was deprecated in ggplot2
## 3.5.0.
## âč Please use the `legend.position.inside` argument of `theme()` instead.
## This warning is displayed once every 8 hours.
## Call `lifecycle::last_lifecycle_warnings()` to see where this warning was
## generated.
p3
Figure 5: Participant-level correlations between subjective arousal, tendency to think about the future, and heartrate. Bayes factors and p values characterize a one-sample t test against zero.
We can see that, for just about all participants, ratings of subjective arousal are largely unrelated to their average heart rate. We can also see that, for many participants, thinking about the future was associated with higher subjective arousal ratings. This makes sense if thinking about the future is energizing or stressful.
2. Interoceptive sensitivity to cardiac signals is stronger during states of lower subjective arousal
To assess whether interoception to cardiac signals changes based on subjective arousal, we examined the results of the cluster-based permutation test, and found one significant cluster. This cluster occurred largely over frontal electrodes about 350 ms post heartbeat.
library(ggpubr)
source('visualize_clusters.r')
##
## Attaching package: 'arrow'
## The following object is masked from 'package:lubridate':
##
## duration
## The following object is masked from 'package:utils':
##
## timestamp
##
## Attaching package: 'eegUtils'
## The following object is masked from 'package:stats':
##
## filter
## here() starts at /home/dave/Dropbox/post_doc/professional/career/davebraundotnet/blogdown
m <- readRDS('post_data/cluster_result.rds')
cluster_id <- which(m$result$p_values < .05)
topo <- plot_topo(m, cluster_id, n_breaks = 1, nrow = 1)
## Joining with `by = join_by(channel)`
## Warning in geom_topo(chan_markers = "text", aes(fill = t_statistic, label =
## electrode)): Ignoring unknown aesthetics: label
## Warning: Using `size` aesthetic for lines was deprecated in ggplot2 3.4.0.
## âč Please use `linewidth` instead.
## This warning is displayed once every 8 hours.
## Call `lifecycle::last_lifecycle_warnings()` to see where this warning was
## generated.
## Warning in geom_topo(chan_markers = "text", aes(fill = t_statistic, label = electrode)): Ignoring unknown aesthetics: fill and label
## Ignoring unknown aesthetics: fill and label
## Ignoring unknown aesthetics: fill and label
## Warning in geom_topo(chan_markers = "text", aes(fill = t_statistic, label =
## electrode)): Ignoring unknown aesthetics: fill
## Warning in geom_topo(chan_markers = "text", aes(fill = t_statistic, label =
## electrode)): Ignoring unknown aesthetics: fill and label
channels <- m$result$channels[m$result$clusters[[2]][[2]] + 1]
times <- m$result$times[m$result$clusters[[2]][[1]] + 1]
d <- m$eeg
N <- length(unique(d$subject))
pd <- d %>%
gather(channel, voltage, Fp1:POz) %>%
filter(channel %in% channels) %>%
group_by(subject, condition, time) %>%
summarize(voltage_ = mean(voltage)) %>%
group_by(condition, time) %>%
summarize(voltage = mean(voltage_), se = sd(voltage_) / sqrt(N))
## `summarise()` has grouped output by 'subject', 'condition'. You can override
## using the `.groups` argument.
## `summarise()` has grouped output by 'condition'. You can override using the
## `.groups` argument.
rib <- pd %>%
filter(time %in% times) %>%
group_by(time) %>%
summarize(ymin = min(voltage, na.rm = TRUE), ymax = max(voltage, na.rm = TRUE))
low <- '#FC8D59'
high <- '#91BFDB'
ts <- pd %>%
ggplot(aes(x = time, y = voltage)) +
geom_rect(aes(xmin = .25, xmax = .650, ymin = min(voltage), ymax = max(voltage)),
color = 'steelblue', fill = NA, linetype = 'dashed') +
geom_ribbon(data = rib, aes(ymin = ymin, ymax = ymax,
xmin = times[1], xmax = times[length(times)], y=1),
alpha = .6, fill = 'green') +
geom_ribbon(aes(ymin = voltage - se, ymax = voltage + se, fill = condition), alpha = .3) +
geom_line(aes(color = condition)) +
labs(
title = 'Heartbeat evoked potential (HEP)',
x = 'Time post heartbeat (s)',
y = latex2exp::TeX('$EEG Voltage~ (\\mu ~V)$'),
color = 'Subjective arousal',
fill = 'Subjective arousal'
) +
scale_color_manual(values = c(Deactivated = low, Activated = high)) +
scale_fill_manual(values = c(Deactivated = low, Activated = high)) +
theme_bw() +
theme(axis.ticks = element_blank(),
panel.grid = element_blank(),
legend.position = 'bottom')
g <- ggarrange(ts, topo, labels = c('A.', 'B.'), nrow=2)
## Warning: The following aesthetics were dropped during statistical transformation: label.
## âč This can happen when ggplot fails to infer the correct grouping structure in
## the data.
## âč Did you forget to specify a `group` aesthetic or to convert a numerical
## variable into a factor?
## Warning: The following aesthetics were dropped during statistical transformation: fill
## and label.
## âč This can happen when ggplot fails to infer the correct grouping structure in
## the data.
## âč Did you forget to specify a `group` aesthetic or to convert a numerical
## variable into a factor?
## The following aesthetics were dropped during statistical transformation: fill
## and label.
## âč This can happen when ggplot fails to infer the correct grouping structure in
## the data.
## âč Did you forget to specify a `group` aesthetic or to convert a numerical
## variable into a factor?
## The following aesthetics were dropped during statistical transformation: fill
## and label.
## âč This can happen when ggplot fails to infer the correct grouping structure in
## the data.
## âč Did you forget to specify a `group` aesthetic or to convert a numerical
## variable into a factor?
## Warning: The following aesthetics were dropped during statistical transformation: fill.
## âč This can happen when ggplot fails to infer the correct grouping structure in
## the data.
## âč Did you forget to specify a `group` aesthetic or to convert a numerical
## variable into a factor?
## Warning: The following aesthetics were dropped during statistical transformation: fill
## and label.
## âč This can happen when ggplot fails to infer the correct grouping structure in
## the data.
## âč Did you forget to specify a `group` aesthetic or to convert a numerical
## variable into a factor?
print(g)
Figure 6: Main result of cluster-based permutation analysis. (A) The grand averaged timeseries of EEG voltage timelocked to the heartbeat (ie, the HEP). Blue dashed rectangle highlights the time points that were analyzed in the cluster test. Green shaded area highlights the significant cluster. (B) Topographical heat map of t values during the significant time window (at time point 0.328 s post heartbeat). The significant cluster centers on frontal electrodes.
Figure 6 highlights the significant cluster from the cluster-based permutation analysis. We can see that, on average, EEG voltage was higher (ie, stronger interoception) prior to lower ratings of subjective arousal. This suggests that participants were more sensitive to cardiac signals during periods of lower subjective arousal.
3. The link between interoceptive sensitivity and subjective arousal was amplified for those high in state anxiety
Given the strong links between anxiety and interoception1, we next wanted to look at how this HEP effect we found might be different for those with low vs. high anxiety. At the beginning of the experiment, we collected measures of both trait anxiety (ie, how much anxiety one experiences generally) and state anxiety (ie, how much anxiety one is experience while completing the survey).
If I were to plot that significant green window in Figure @ref(fig:cluster_result) for each participant, we would see that each participant is different with respect to how strong their link is between interoception and subjective arousal (ie, their HEP effect). We wanted to know whether these differences could be explained by differences in state or trait anxiety.
So, for each participant, I calculated their average HEP effect and correlated it with their anxiety measures:
# --- TOP PANEL --- #
# Individual-level correlations between HEP effect size and survey response
# Select correct data
channels_s <- m$result$channels[m$result$clusters[[2]][[2]] + 1]
times_s <- m$result$times[m$result$clusters[[2]][[1]] + 1]
cluster <- data.frame(time = times_s, channel = channels_s)
cluster$string <- paste(cluster$time, cluster$channel, sep='_')
survey <- read.csv('post_data/MW_EEG_survey.csv')
survey <- survey[, c('subj_id', 'GAD7_Score', 'STAI_Score')]
colnames(survey) <- c('subject', 'gad', 'stai')
d <- m$eeg
# Select appropriate (significant) cluster
hep_s <- d %>%
gather(channel, voltage, Fp1:POz) %>%
mutate(string = paste(time, channel, sep='_')) %>%
filter(string %in% cluster$string, channel %in% channels_s) %>%
group_by(subject, condition) %>%
summarize(voltage = mean(voltage)) %>%
mutate(condition = ifelse(condition == 'Activated', 'High Arousal', 'Low Arousal')) %>%
inner_join(survey)
## `summarise()` has grouped output by 'subject'. You can override using the
## `.groups` argument.
## Joining with `by = join_by(subject)`
# Make general summarized data
t <- hep_s %>%
spread(condition, voltage) %>%
mutate(hep = `Low Arousal` - `High Arousal`) %>%
select(-`High Arousal`, -`Low Arousal`) %>%
gather(survey, response, gad,stai) %>%
mutate(survey = recode(survey, `gad` = 'Trait anxiety', `stai` = 'State anxiety'))
# Compute group-level correlations
cr <- t %>%
group_by(survey) %>%
summarize(r = cor.test(hep, response)$estimate,
df = cor.test(hep, response)$parameter,
p = cor.test(hep, response)$p.value,
ci_h = cor.test(hep, response)$conf.int[1],
ci_l = cor.test(hep, response)$conf.int[2]) %>%
mutate(label = paste0('r(', df, ') = ', round(r, 3), '\n95% CI = [', round(ci_l, 3), ', ',
round(ci_h, 3), ']\np = ', round(p, 3)),
hep = c(-.5, -.5), response = c(39, 10), survey = c('State anxiety', 'Trait anxiety'))
# Visualize correlations
p1 <- t %>%
filter(!is.na(response)) %>%
ggplot(aes(x = response, hep)) +
geom_point() +
geom_smooth(method = 'lm') +
facet_wrap(~survey, scales = 'free_x') +
geom_text(data = cr, aes(label = label), hjust = 0, size = 3) +
labs(
x = 'Survey response',
y = 'Participant-specific\nHEP effect size'
) +
ylim(-1, 1) +
theme_bw() +
theme(panel.grid = element_blank(),
strip.background = element_rect(fill = NA, color = 'black'),
axis.ticks = element_blank())
# --- PANEL 2 --- #
# Visualize median split on EEG timeseries
# Keep relevant columns
channels_s <- unique(channels_s)
d <- m$eeg
mask1 <- colnames(d) %in% c('subject', 'condition', 'time')
mask2 <- colnames(d) %in% channels_s
d <- d[, (mask1 | mask2)]
# Plot
p2 <- d %>%
inner_join(survey[,c('subject', 'stai')]) %>%
filter(!is.na(stai), time %in% times) %>%
mutate(stai_d = ifelse(stai > median(stai), 'High state anxiety', 'Low state anxiety')) %>%
gather(channel, voltage, Fp1:FC5) %>%
group_by(subject, time, condition, stai_d) %>%
summarize(voltage_ = mean(voltage)) %>%
group_by(time, condition, stai_d) %>%
summarize(voltage = mean(voltage_), se = sd(voltage_) / sqrt(n())) %>%
mutate(stai_d = factor(stai_d, levels = c('Low state anxiety', 'High state anxiety'))) %>%
ggplot(aes(x = time, y = voltage)) +
geom_ribbon(aes(ymin = voltage - se, ymax = voltage + se, fill = condition), alpha = .3) +
geom_line(aes(color = condition)) +
facet_wrap(~stai_d) +
labs(
x = 'Time post heartbeat (s)',
y = latex2exp::TeX('$EEG~voltage~(\\mu ~V)$'),
fill = 'Subjective arousal',
color = 'Subjective arousal'
) +
scale_color_manual(values = c(`Activated` = high, `Deactivated` = low)) +
scale_fill_manual(values = c(`Activated` = high, `Deactivated` = low)) +
theme_bw() +
theme(panel.grid = element_blank(),
strip.background = element_rect(fill = NA, color = 'black'),
axis.ticks = element_blank(),
legend.position = 'bottom')
## Joining with `by = join_by(subject)`
## `summarise()` has grouped output by 'subject', 'time', 'condition'. You can
## override using the `.groups` argument.
## `summarise()` has grouped output by 'time', 'condition'. You can override using
## the `.groups` argument.
g <- ggarrange(p1, p2, labels = c('A.', 'B.'), nrow = 2)
## `geom_smooth()` using formula = 'y ~ x'
## Warning: Removed 2 rows containing non-finite outside the scale range
## (`stat_smooth()`).
## Warning: Removed 2 rows containing missing values or values outside the scale range
## (`geom_point()`).
print(g)
Figure 7: Participant-level associations between HEP effect size and self-reported anxiety ratings. (A) Correlations with HEP effect size for both state (left) and trait (right) anxiety. (B) EEG grand averaged timeseries broken by a median split on state anxiety.
What it means
As your subjective arousal increases, your brain becomes more sensitive to your heart signals—but if you’re higher in anxiety, that boost in sensitivity is even stronger
We found that the brain’s response to heartbeats changes depending on how energized you feelâand that this effect is different for people with higher anxiety. When people were in a more amped-up state (what weâre calling high subjective arousal), their brains showed a stronger reaction to heartbeats. And for folks with high anxiety, this pattern was amplified.
So what does that mean, practically?
It suggests that when your energy level shiftsâeven just quietly, while youâre restingâyour brain is tuning in to your body in a different way. And if youâre someone who struggles with anxiety, your brain might be tuning in differently altogether.
This matters because a lot of anxiety happens at rest, when weâre left alone with our thoughts. The brain networks that light up during rest are also the ones tied to rumination, worry, and âwhat ifâ spirals. Our results suggest that these moments of rest may not be as neutral as they seem: your brain is actively tracking how your body feels, and if you’re anxious, it might be tracking differently.
In the future, this kind of research could inform better mental health tools. Imagine a wearable that doesnât just track your heart rate, but notices when your brainâs response to your body changes in a way that predicts anxiety. Or an app that could nudge you toward a calming activity right when your internal signals suggest you’re about to start spiraling.
More broadly, this work shows that our internal worldâhow we feel insideâhas a real, measurable footprint in the brain. It reinforces the idea that emotions arenât just things we talk about in therapyâtheyâre deeply embedded in how our body and brain talk to each other. And that connection might be key to understanding not just anxiety, but how we can better support mental health in everyday life.
-
Pang, J., Tang, X., Li, H., Hu, Q., Cui, H., Zhang, L., Li, W., Zhu, Z., Wang, J., & Li, C. (2019). Altered Interoceptive Processing in Generalized Anxiety DisorderâA Heartbeat-Evoked Potential Research. Frontiers in Psychiatry, 10, 616. https://doi.org/10.3389/fpsyt.2019.00616 ↩︎ ↩︎
-
Coll, M.-P., Hobson, H., Bird, G., & Murphy, J. (2021). Systematic review and meta-analysis of the relationship between the heartbeat-evoked potential and interoception. Neuroscience & Biobehavioral Reviews, 122, 190â200. https://doi.org/10.1016/j.neubiorev.2020.12.012 ↩︎
-
Fourcade, A., Klotzsche, F., Hofmann, S. M., Mariola, A., Nikulin, V. V., Villringer, A., & Gaebler, M. (2024). Linking brainâheart interactions to emotional arousal in immersive virtual reality. Psychophysiology, e14696. https://doi.org/10.1111/psyp.14696 ↩︎
-
Kucyi, A., Kam, J. W. Y., Andrews-Hanna, J. R., Christoff, K., & Whitfield-Gabrieli, S. (2023). Recent advances in the neuroscience of spontaneous and off-task thought: Implications for mental health. Nature Mental Health, 1(11), 827â840. https://doi.org/10.1038/s44220-023-00133-w ↩︎
-
Park, H.-D., & Blanke, O. (2019). Heartbeat-evoked cortical responses: Underlying mechanisms, functional roles, and methodological considerations. NeuroImage, 197, 502â511. https://doi.org/10.1016/j.neuroimage.2019.04.081 ↩︎