Episode Transcript
[00:00:00] Welcome back to this Week in hrv. I'm Matt Bennett, co founder of Optimal hrv, and this is one of those weeks where I found myself going back through the papers a second time. Not because I missed anything on the first pass, but because the combination of what these seven studies are doing together is genuinely interesting. Individually, each of them is worthwhile on its own terms. Collectively, they reveal something about the current moment in HRV science. The field has moved past the foundational questions we established long ago that HRV is meaningful and that it reflects autonomic balance, that it correlates with fitness recovery, stress, resilience, and a range of clinical outcomes. And it is now operating in what I think of as a second generation of inquiry. That second generation is asking more demanding, more contextual, more epistemologically careful questions whether a finding applies equally to men and women, whether the physiological signal recorded in a laboratory setting still means the same thing in an immersive social environment. Whether the machine learning models we have come to rely on actually reveal something about biology, or whether they reveal patterns in data that look like biology but are not. These are harder questions, and answering them requires research that is more deliberately designed and more honest about what it can and cannot conclude. This week's seven papers are examples of that kind of work across an impressively wide range of topics and settings. The information in this podcast is for educational and informational purposes only and is not intended to diagnose, treat, cure, or prevent any medical condition. Please consult with a qualified healthcare provider before making any changes to your health regimen, training program, or clinical practice. Heart rate variability is a powerful and scientifically validated tool for understanding autonomic nervous system function, but it is one piece of a much larger picture, and context always matters. Nothing shared in this podcast should be taken as personalized medical advice and individual responses to interventions vary enormously. Always work with qualified professionals when making decisions about your health. Let me give you a preview of what we're covering. The first paper is a comprehensive review that synthesizes 15 years of research on the brain heart axis as measured by HRV and asking whether that axis operates the same way in males and females. The answer is more nuanced than a simple yes, and the implications are substantial for how HRV research is designed and interpreted. The second is a carefully controlled experiment with 67 early adolescents that asks what a brief, gentle touch to the forearm does to the autonomic nervous system of a young person who has just been socially excluded in a virtual reality game. The third is a large scale wearable data study that used over one and a half million nights of physiological recordings to ask whether co sleeping with a dog, a cat or a child a affects resting heart rate, breathing rate and hrv. The fourth is an exploratory pilot study examining whether five minutes of ear based vagus nerve stimulation changes baroreflex sensitivity, one of the most sensitive cardiovascular autonomic measures in healthy men. The fifth presents an open source computational framework for studying the coupling of brain oscillation patterns and heart rhythm during sleep in people with sleep disorders. The sixth is a multimodal environmental psychology study from Hong Kong that used eye tracking, eeg, skin conductance and HRV simultaneously to assess how six different types of urban tourism streetscapes affect physiological stress and restoration. And the seventh, the one I find most important for how the field needs to develop, is a methodological paper using a lifespan data set of nearly 4,000ambulatory cardiac recordings to ask whether the feature importance rankings produced by predictive machine learning models tell you anything trustworthy about the biology of heart rate variability. Let's take them one at a time. We begin with the systematic review published in Sex and Gender Specific Biomedicine. The lead author is so Yi Yoon and her co authors are Hee Sang, Yang, Maithre A Div Jihye, Li Na Jung, Yoon, Yuenhua Lee and Hyang Woon Lee. The paper's central question is whether the brain heart axis as reflected in HRV functions the same way in males and females. The short answer is that the evidence suggests it does not, and the longer answer involves some important nuance about what that means and why it matters. To understand what they were reviewing, it is worth establishing what we mean by the brain heart axis. The heart is not a free standing mechanical pump that happens to beat while the brain does something entirely separate. The two organs are in continuous bi directional communication through the autonomic nervous system. The sympathetic branch accelerates the heart and suppresses the variability of its rhythm. The parasympathetic branch, carried primarily through the vagus nerve, slows the heart and restores variability. The interplay between these two branches updating in real time in response to everything the body and brain are experiencing, generates the pattern of variation we measure as hrv. The components most tightly tied to parasympathetic activity. RMSSD root mean square of successive differences and and the high frequency spectral component serve as windows into how strongly the brain is exerting regulatory control over cardiac rhythm through vagal pathways. The brain regions that generate this top down regulation form what Researchers call the central autonomic network the medial prefrontal cortex, the anterior cingulate cortex, the insular cortex, and the amygdala, among others. When we use RMSSD or high frequency power as a biomarker of emotional regulation, stress resilience, or cognitive control, we we are implicitly using it as a downstream signal from this network. And when we ask whether the same HRV value means the same thing in different people, we are really asking whether the network itself is operating through the same pathways. Yoon and colleagues synthesized a large and methodologically diverse literature covering neuroimaging studies, autonomic assessments, and research that combined both. What they found consistent across multiple studies conducted by different research teams using different methods, was that women and men tend to engage the central autonomic network through partially different neural routes. Women consistently showed stronger activation of limbic structures, particularly the amygdala and insular cortex, during tasks that engaged HRV regulation. These are regions deeply involved in emotional processing, interoceptive awareness, the brain's representation of the body's internal state, and the integration of affective and somatic information. Men showed more consistent activation of lateral prefrontal regions associated with deliberate top down cognitive control and inhibitory function. The framework the authors built from this pattern is that men and women appear to regulate cardiac autonomic state through different structural weightings of the central autonomic network. Women generating or modulating that state more through pathways anchored in emotional and bodily awareness, men more through pathways anchored in deliberate cognitive modulation. These are tendencies across groups, not categorical rules about individuals. Variation within each sex is substantial, but the pattern is consistent enough across the evidence base to be taken seriously as a real biological difference rather than a statistical artifact. The practical implications of this finding are significant and often overlooked in discussions of HRV research design the standard methodological response to sex differences in a mixed sample study is to include sex as a covariate in the statistical model. This approach removes the variance attributable to sex, which can sharpen the precision of other estimates, but it also renders the model structurally incapable of detecting whether sex moderates the relationship being studied. If the relationship between HRV and emotional regulation or cognitive performance is actually mediated by the amygdala insular pathway in women and by the prefrontal pathway in men controlling for sex prevents the analysis from discovering that the mediated pathway differs and the different pathways might respond differently to interventions, correlate differently with clinical outcomes, and have different baseline trajectories across the lifespan, a finding reported as applicable to the general population may apply more cleanly to one sex than the other, but a study that controlled for sex has no way to reveal that asymmetry. Yoon and colleagues argue that this is not a minor methodological point. It is a structural limitation that has constrained what the HRV field has been able to learn about the biology it is trying to understand. Their recommendation is for a shift toward treating sex as a biological variable of primary interest rather than a confound to manage. Practically, this means recruiting samples large enough to test sex by outcome interactions with adequate statistical power, pre registering those tests rather than exploring them post hoc, building normative HRV databases that are stratified by sex rather than average to cross it, and critically examining whether clinical reference ranges derived from mixed sex samples are appropriate for making individual level clinical inferences.
[00:07:25] These are not trivial design changes they require larger samples and more explicit hypotheses. But they are necessary to answer questions that the current literature has been structurally prevented from addressing. The review has its own limitations. The literature it synthesizes is heterogeneous in its imaging modalities, task paradigms, populations, and HRV measures, and extracting consistent patterns from that heterogeneity requires judgment calls that may not always be reproducible. But the pattern is robust enough, and its implications consequential enough, that the field should take it seriously. For practitioners and applied researchers collecting and interpreting HRV data from individuals of different sexes and the immediate takeaway is that the sex of the person you are measuring is not background demographic information. It may be part of the biological story that the number is trying to tell. The second study takes us into a very different setting. A controlled laboratory experiment designed to test whether gentle human touch can shift the autonomic nervous system of a young person who has just experienced social rejection. The paper was published in the Pediatric Psychology section of Frontiers in Psychology. The authors are Letizia della Longa, Irene Guglielmanetti, Irene Valori, Francesco Rizzi, Gustavo Marfia, and Teresa Ferroni. Their 67 participants were early adolescents 11 to 14 years old, and the choice of this age range is scientifically meaningful. Early adolescence is the developmental period when the pure social world becomes genuinely important to the self in a way that it was not in childhood. Rejection and exclusion begin to feel existentially threatening rather than simply unpleasant. The autonomic nervous system is still maturing its adult regulatory architecture, which makes adolescents both more reactive to social stressors and and potentially more responsive to social regulatory interventions like touch, and the clinical context is relevant. Rates of anxiety, loneliness, and depression in early adolescence have risen substantially in recent decades, making research on autonomic regulation and social experience in this age group timely. The social stressor used in this study was a virtual reality version of the cyberball paradigm. Cyberball was developed in the late 1990s as a simple, standardized, and highly reliable way to induce feelings of social exclusion in a laboratory setting. Participants play a virtual ball tossing game with simulated other players who stop, including them partway through, leaving the participant out. The effect is robustly consistent even when participants are told in advance that the other players are simulated by software. The exclusion phase reliably produces reduced mood, lower sense of belonging, and cardiovascular changes consistent with a social threat response. The virtual reality implementation used in this study places the participant inside an immersive three dimensional social environment rather than observing a flat two dimensional screen, which is likely to increase the sense of social presence and produce a stronger autonomic response to the exclusion than the original flat screen version does. After the exclusion phase ended, each participant received one of two conditions. In the experimental group, a trained researcher applied gentle, slow stroking touch to the ventral forearm, the surface of the forearm that faces upward when the palm is open for a brief standardized period. In the control group, participants heard a neutral auditory tone. Both conditions were designed to involve receiving some form of external stimulus and in the aftermath of the stressor, so that any difference in outcome could be attributed to the specific content of the stimulus touch versus a neutral sound rather than simply to having something happen after the stressor. The results were clear across multiple outcome measures. Participants in the touch condition showed better recovery, heart rate returned toward baseline more quickly, autonomic balance shifted in a direction reflecting stronger parasympathetic and weaker sympathetic contribution, and self reported mood improved more in the touch group than in the control group. These findings fit within a substantial and growing literature on what researchers call effective touch the specific class of gentle, slow skin temperature appropriate stroking contact that activates a dedicated population of unmyelinated peripheral nerve fibers called C tactile afferents. These fibers, which are present in hairy skin but not in glabrous skin, are tuned to respond most strongly to stroking at velocities between 1 and 10 centimeters per second at skin temperature, and they project to the insular cortex and other brain regions associated with social processing and interoception rather than to the somatosensory cortex that processes ordinary tactile sensation. The downstream effects of C tactile activation, which include increased insular and anterior cingulate activity, reduced amygdala reactivity, and autonomic shifts toward parasympathetic dominance are consistent with what Delalonga and colleagues observed in their experimental condition. Whether C tactile afferents were the operative mechanism in this specific study cannot be determined from the data because the study did not include the neural or biochemical measurements that would verify that pathway, but the observed pattern matches what that mechanism would predict. The secondary exploratory finding that lonelier adolescents may have benefited more from touch is one of the more clinically interesting results and it deserves both attention and appropriate caution. It came from a subgroup analysis in which participants were divided by self reported loneliness scores and the touch effect was compared across those subgroups. This analysis was not pre registered as a primary hypothesis.
[00:11:53] It was an exploratory examination of secondary data, which means it carries a substantially higher risk of being a chance finding than the primary analysis did. The sample of 67 participants divided into loneliness stratified subgroups for this comparison does not leave much statistical precision to work with. The appropriate response to this finding is genuine curiosity in a purpose built study. A larger trial specifically designed to test whether loneliness moderates the autonomic response to touch with pre registered hypotheses and adequate power. If that hypothesis replicates under those conditions, it would have meaningful implications for understanding touch deprivation as a specific risk factor in lonely young people and for informing how touch based interventions might be targeted. For now, the primary contribution of this paper is a well designed demonstration that brief, gentle touch produces measurable and directionally meaningful shifts in the autonomic state of early adolescence following social exclusion. A finding that adds substance to the case for taking effective touch seriously is a physiological regulatory resource in adolescent settings. The third study moves entirely out of the controlled setting and into one of the largest real world wearable data sets that has been assembled to address a question of this type. The paper was published in Nature and Science of Sleep and the authors are Alexander Montoy, Dylan Curran and Gregory Grisicki. Their question is one that many people with a fitness tracker have probably already wondered informally, does sleeping with a pet or child actually change your physiology? Their answer is that it does change something, but what it changes and in what direction turns out to depend on which outcome you look at, and the autonomic picture and the sleep quality picture do not always tell the same story. The data set came from WHOOP users who logged their bed sharing arrangements in the app's daily journal feature over an extended period. Across 11,733 users, there were approximately 1.5 million recorded nights with physiological data. The methodological strength of this study is its within person design. Each participant's co sleeping nights are compared against their own non co sleeping nights, which automatically controls for all the stable characteristics of each individual their typical hrv, their habitual sleep patterns, their age, their fitness level, their underlying health status. The analysis is asking what changes within a person on the nights when they share a bed, rather than asking whether people who co sleep are systematically different from those who do not. This is a cleaner and more meaningful test of the question, and the data set is large enough that even small within person effects can emerge with statistical confidence. What the analysis found was a pattern that is simultaneously interesting and difficult to interpret cleanly. On co sleeping nights, resting heart rate was lower, breathing rate was lower and HRV was higher compared to the same participants solo nights. Those are the physiological signatures of a more parasympathetically dominant, less sympathetically activated state, the kind of autonomic profile associated with better cardiovascular recovery and lower physiological stress. If you looked only at those numbers, you might conclude that sharing a bed improves recovery. But the same data showed that co sleeping nights were associated with more fragmented sleep, more frequent waking events, low lower sleep efficiency and modestly lower subjective ratings of restfulness and sleep quality. Sleep was somewhat longer in total duration on co sleeping nights, but more interrupted and less consolidated. The autonomic profile looks better, the sleep architecture looks worse, that dissociation is not easily resolved into a simple conclusion about whether co sleeping is beneficial or detrimental. Several interpretive possibilities are worth considering. Sharing a bed with an animal or a child may activate social and affiliative neural circuits that reduce baseline sympathetic tone throughout the night, a sense of warmth, safety or connection that influences cardiac autonomic state even when the sleep itself is lighter and more frequently disrupted. It is also possible that people are more likely to co sleep on nights when they are already in a state associated with lower stress and stronger social connection, and that those states themselves drive the favorable autonomic profile independently of the bed companion. This is the fundamental challenge of observational with in person designs. Even when stable individual differences are controlled for time varying factors.
[00:15:29] The circumstances and emotional states that differ between co sleeping and solo nights cannot be fully ruled out as confounders. When broken down by companion type. Sleeping with a child produced the most disrupted sleep architecture of the three conditions. Dogs and cats produced relatively similar patterns to each other, both somewhat less disruptive than sleeping with a child next day. Physical activity levels were higher after pet co sleeping and lower after sleeping with a child, which likely reflects both genuine differences in physiological recovery and the behavioral demands of the following day.
[00:15:55] All of the effect sizes in this study are small in absolute terms, which is an important qualification. The statistical significance that emerges from one and a half million nights of data tells you that the effects are real and not noise, but it does not tell you that they are large enough to meaningfully change any individual's health or recovery. For a practitioner advising a client who is struggling with sleep quality, this paper would not provide strong grounds for recommending they remove their pet from the bed based on HRV or physiological recovery alone. What it provides is an honest empirical description of the trade offs that bed sharing involves. Measurable but modest shifts in autonomic state in one direction, measurable but modest shifts in sleep architecture in another, and no simple verdict on which set of effects dominates the individual's own preferences and what they value in their sleep environment belong in that conversation. Alongside the data, the fourth study examines transcutaneous auricular vagus nerve stimulation, a technique that has attracted substantial research interest over the past decade as a potential non invasive approach to modulating cardiac autonomic function. The paper was published in Scientific Reports and the authors are Yusuke Katahara, Atsuhiko Iijima, and Kazuki Taika. Their specific question was whether a brief session of ear based vagal stimulation produces detectable changes in baroreflex sensitivity in healthy men and how those changes compare to simultaneous changes in the HRV indices that researchers most commonly use to track autonomic responses to this kind of stimulation. To understand the rationale behind this work, it helps to know why the ear is interesting as a target for vagal stimulation. The vagus nerve is the primary channel of parasympathetic communication between the brain and most major organs, including the heart. In its standard trajectory, it runs deep inside the body and is inaccessible without surgery, which is why the implanted vagal nerve stimulators used in clinical medicine require a surgical procedure to place but the vagus nerve has a peripheral auricular branch that innervates part of the outer ear, particularly a region of the outer bowl called the simba conchae. This branch is accessible through the skin, meaning that a small electrode or clip applied there can potentially deliver electrical stimulation to the vagal fiber endings without any invasive procedure. If that stimulation propagates into the central vagal pathway and increases parasympathetic signaling to the heart, the results should be measurable as changes in HRV and in the baroreflex, the feedback system by which baroreceptors in the arteries signal blood pressure changes to the brain, which then adjusts heart rate and cardiac output through autonomic pathways to restore pressure to target levels. Baroreflex sensitivity how many milliseconds the RR interval changes per millimeter of mercury change in systolic pressure is a measure of how efficiently this feedback loop is operating, and it is considered a sensitive indicator of cardiovascular autonomic health. Kadahar and colleagues worked with 18 healthy men in an open label protocol with no control or sham condition. Each participant rested for five minutes in a seated position, received five minutes of left ear tavns targeting the simiconche, and then rested for five more minutes after stimulation ended. Throughout all three phases, the researchers continuously recorded blood pressure, heart rate, and the electrocardiogram from which they extracted rmssd, high frequency, HRV power, and Meyerband coherence. Mayer waves are low frequency oscillations in blood pressure and heart rate at approximately 0.1 Hz and and their coherence. The degree to which the heart rate and blood pressure rhythms at that frequency are synchronized reflects aspects of sympathovagal interaction that the standard HRV time domain and frequency domain measures do not fully capture together. This battery was designed to give a reasonably comprehensive picture of how TAVNS affects the cardiovascular autonomic system. The results were consistent across participants, but not in the direction that a simple TAVNS increases parasympathetic tone model would predict. Baroreflex sensitivity decreased in 17 of the 18 participants during the stimulation phase. Systolic blood pressure rose, RMSSD and high frequency power showed no significant change. Mayor band coherence did not change significantly either. The one consistent signal in the dataset was a decrease in baroreflex sensitivity, and that decrease coincides with an increase in blood pressure, which is mechanistically interpretable. When blood pressure rises, the baroreceptor reflex should activate the parasympathetic system to bring it back down, and if the baroreflex sensitivity is simultaneously decreasing during that blood pressure rise, it could reflect a reflex suppression of the baroreflex pathway or a complex interaction between the direct autonomic effects of the stimulation and the compensatory cardiovascular responses it triggers. The authors present several possible explanations without being able to definitively adjudicate between them, which is appropriate given the study's exploratory nature and the absence of a control condition. That absence of a sham condition is the most significant methodological limitation without a condition in which participants experience everything about the TAVNS procedure the electrode placement, the time sitting quietly, the the experimental context except the actual electrical stimulation, there is no way to attribute the observed changes to the stimulation specifically rather than to arousal, novelty, expectation or discomfort. The all male 18 person sample also limits generalization substantially and the stimulation parameters used intensity, pulse frequency, pulse width, duration were not optimized and may not represent the conditions under which TAVNS would produce its largest or most favorable autonomic effects. Despite these limitations, the paper is valuable for two reasons. First, it provides direct empirical evidence that the expected HRV changes that motivate much of the TAVNS research literature did not appear in a careful within session measurement, while a less commonly measured outcome, baroreflex sensitivity, did show a consistent response. Second, it does so with unusual methodological transparency, naming its limitations explicitly and not over claiming for anyone following the TAVNS research space. This paper is an honest data point that argues for more rigorous controlled study of what ear based stimulation actually does to the cardiovascular autonomic system under what conditions before we continue a quick word about Optimal HRV One of the recurring themes in this research is that what HRV reflects depends substantially on the context in which you measure it and the baseline you compare it against. A resting heart rate measurement taken on a co sleeping night tells a different story than one taken on a solo night, a baroreflex sensitivity measurement during vagal stimulation tells a different story than one taken at rest, and a single HRV number tells you almost nothing on its own. What it means becomes clear only in relation to your own history and your own patterns. This is exactly why the optimal HRV app centers its methodology on daily consistent measurement at the same time and in the same position. Each morning.
[00:21:49] Over days and weeks and months of that kind of consistent data collection, you build a personal baseline that makes your HRV values interpretable in a way that a single reading never can be. You learn what is normal for you, what a significant deviation looks and feels like, and what patterns tend to precede changes in recovery or performance. The app also includes guided paced breathing tools for actively engaging your parasympathetic nervous system, educational content that explains what the measurements mean, and tracking features that let you see your own data and context. If you're serious about using HRV as a genuine window into your autonomic health, not just collecting a number to look at, visit optimalhrv.com the fifth study shifts into methodological territory. It is a paper that builds infrastructure and tools for other researchers rather than testing a specific empirical hypothesis, but the practical implications for research on sleep and autonomic function are substantial. The paper was published in Frontiers in Physiology and the authors are Shaween Shoukir, Mostafa Moussa, Yahya, Al Zabi, Hassan Khandoker, and Zbigniew Struzik. What they built is an open source computational framework that allows researchers to simultaneously analyze brain oscillation patterns during sleep, specifically a phenomenon called cyclic alternating pattern, and a comprehensive battery of HRV features derived from the cardiac signal, with both streams synchronized in a common temporal reference frame. To appreciate why this matters, you need to understand what cyclic alternating pattern is and why it has become an important target for sleep research. During non rapid eye movement sleep, the majority of a typical night's sleep, the brain does not maintain a uniform level of activity. It cycles between brief periods of higher amplitude, more electrically activated states, which researchers call a phases, and intervening periods of quieter, lower amplitude background activity. This oscillation is visible in the electroencephalogram and constitutes cyclic alternating pattern. The clinical significance of this pattern is that its structure changes predictably in sleep disorders. The rate of aphase occurrence, the amplitude and morphology of individual aphases, and the distribution of aphase subtypes are all altered in characteristic ways in obstructive sleep apnea, insomnia, and periodic limb movement disorder, distinguishing affected patients from healthy sleepers in ways that standard sleep staging alone does not capture. What makes cyclic alternating pattern particularly interesting from the perspective of cardiac autonomic research is that the heart does not simply track these brain oscillations passively or ignore them. There's growing evidence that transitions into and out of a phases are accompanied by systematic changes in heart rate, respiratory pattern, and autonomic balance that the brain and heart are functionally coupled during sleep in ways that the oscillation events make visible. Studying that brain heart coupling requires measuring both the brain through EEG and the heart through ECG simultaneously and with accurate temporal synchronization between the two signals. This kind of research has historically been difficult to conduct for two related reasons. First, identifying cyclic alternating pattern events requires expert manual scoring of EEG recordings and which is slow, expensive, and difficult to scale to large research data sets. Second, the software tools that existed for cyclic alternating pattern analysis and the software tools that existed for HRV computation had not been integrated into a shared workflow. Researchers who wanted to study both had to write their own synchronization and alignment code, which created barriers to reproducibility. Different labs handling the synchronization differently and limited who could practically attempt this kind of work to groups with dedicated computational resources and expertise. The pipeline that Shukir and colleagues built addresses both problems. It is a MATLAB based framework that takes EEG and ECG recordings as inputs, applies an automated algorithm to detect cyclic alternating pattern A phases from the brain signal, extracts a comprehensive set of HRV features from the cardiac signal using validated toolboxes, and outputs a synchronized data set in which each brain oscillation event is temporally aligned with the cardiac autonomic state measured in its vicinity. The HRV feature set is broad, including nonlinear and geometric indices alongside the standard time domain and frequency domain measures, which is appropriate for capturing the full complexity of autonomic changes that occur around sleep oscillation events. The framework was applied and demonstrated using recordings from 18 individuals with obstructive sleep apnea, which is an appropriate test case because sleep apnea produces well characterized disturbances in both cyclic alternating pattern structure and cardiac autonomic function, giving a demonstration data set with known and interpretable characteristics. The framework also provides a normative HRV reference derived from healthy controls in the publicly available Cap Slide Sleep database, which gives future researchers a comparison benchmark when they want to characterize how a disordered population's brain heart coupling differs from typical patterns. The key limitation, the authors acknowledge, is that the automatic cyclic alternating pattern detection algorithm has not yet been validated against expert manual scoring, the gold standard for establishing how accurately an automated detector identifies the events it is supposed to identify. This is the element of the framework whose performance is most critical to the validity of research outputs, and it is the element that remains least well characterized until that validation study is conducted and published. Researchers using the framework need to hold their results with appropriate awareness that the detection step introduces uncertainty that has not been quantified. The framework is available as open source code, which means other researchers can inspect its methods, identify potential issues, and contribute to its validation. For a field where brain heart coupling during sleep is an important and underexplored territory, making working tools available at all, even tools that need further validation lowers real barriers to scientific progress. Sixth study takes us outdoors into one of the most heavily visited urban tourism environments in the world. The paper was published in Frontiers in Public Health, and the authors are Xin Liang, Wang, Yafei, Zhao, Xiangshu, Xiao Zhuoli, and Jingyuanqing. The research site was the network of tourism streets, promenades, and waterfront spaces surrounding Victoria harbor in Hong Kong. Their question was whether six distinct types of streetscape environments systematically identified from A survey of 66 tourism sites produce meaningfully different physiological and psychological responses in the people who experience them. The conceptual foundation for this work draws on two well established theories in environmental psychology. Attention restoration theory developed by Rachel and Stephen Kaplan in the 1980s and refined sense holds that natural environments are restorative because they engage what the Kaplan's called involuntary attention, the soft, effortless attention drawn by moving water, wind and leaves, varied natural patterns and other forms of gentle changing stimulation, allowing the voluntary or directed attentional system to recover from the depletion that sustained focused work causes. Stress recovery theory developed by Roger Ulrich and colleagues holds that the human physiological stress response evolved in a context where natural environments were associated with safety and resources and that the cardiovascular and autonomic stress response is therefore more readily downregulated by natural scenes than by built or man made environments because natural scenes activate an evolved orienting response that is incompatible with sustained threat vigilance. Both frameworks predict that environments with stronger natural elements, water features, greenery, open sky, varied natural textures will produce lower physiological stress, faster autonomic recovery and higher self reported psychological restoration compared to more built, dense or enclosed environments. Wang and colleagues designed their study to test these predictions comprehensively in a real world tourism context after surveying 66 tourism sites in the Victoria harbor area and characterizing each with 11 quantitative indicators including measures of sky visibility, water presence, greenery proportion, built surface density and spatial enclosure derived from photographs and geographic information system data that use statistical clustering to identify six representative scene types and these types were not hand picked but emerged from the data as the statistically natural groupings of sites ranging from open panoramic waterfront scenes with minimal build obstruction to dense canyon like urban streetscapes to intermediate environments. Combining greenery and built elements in various proportions, 46 participants then viewed images of each scene type in a controlled experimental setting while the researchers simultaneously measured HRV and heart rate skin conductance as an index of sympathetic nervous system activation, electroencephalogram measures of neural arousal and engagement, and eye tracking, capturing where participants directed their gaze and how long they dwelled on different elements within each scene. Subjective measures of perceived restoration, stress and pleasantness were also collected and a validated attention accuracy task gave a behavioral index of cognitive function that could be compared across scene types. The results broadly supported the theoretical predictions across most of the measured outcomes. Scene types with stronger natural elements produced lower skin conductance, indicating less sympathetic arousal, more favorable HRV profiles, and higher self reported restoration compared to more built and enclosed scene types. Eye tracking data showed that participants directed more visual attention to natural elements within mixed scenes and that dwell time on natural features was negatively correlated with sympathetic arousal, suggesting that the natural elements were specifically driving the physiological response rather than merely being part of a more pleasant scene. Overall, electroencephalogram measures showed lower neural arousal indices for nature dominant scenes. The one domain that did not show significant differences across the scene types was the attention accuracy task, which means the study found clear and consistent physiological and subjective differences between environment types without those differences producing measurable behavioral differences in cognitive performance on this particular task. Whether that represents a genuine null effect on cognition, natural scenes restore autonomic state but not measurable cognitive performance within the brief task duration or a sensitivity limitation of the specific task used is not clear from the available data. The limitations of this study are worth naming. The use of photographs rather than actual physical environments is the most fundamental. A substantial body of research shows that viewing images of natural or built environments produces physiological responses in the predicted direction, but those effects are typically smaller than what is observed with real world exposure. An image viewing in a controlled laboratory room introduces a confound of the viewing context itself. The sample of 46 participants, while adequate for primary comparisons, is modest for a study involving six scene types and multiple outcome domains. The cultural and geographic specificity of the Hong Kong harbor setting, characterized by a particular combination of high density built environment and dramatic waterfront, may limit how directly the specific findings translate to other urban contexts with different characteristic landscapes. Within these constraints, the paper is a methodologically ambitious and and thoughtfully designed demonstration that urban streetscape type leaves measurable and coherent traces across multiple physiological systems simultaneously, with practical implications for how designers, planners, and urban tourism developers think about the restorative value of different environmental configurations. The seventh and final study is the one I find most conceptually important for the direction the HRV research field needs to move. The paper was published in Applied Sciences and the authors are Junichiro, Hayano, and Amiyuda. Their data set is genuinely remarkable 24 hour ambulatory cardiac Holter recordings from 3854 individuals spanning ages from 0 to 102 years. This is a true lifespan sample in the most complete sense of that phrase. It captures the full developmental and aging arc of cardiac autonomic function from newborns through childhood, adolescence, adulthood, and old age, all the way to centenarians. The rarity of a data set with this scope is what makes the central empirical test the paper performs possible.
[00:31:54] The question Hayano and Yuta are asking is fundamentally about the epistemology of machine learning applied to biomedical data. Machine learning methods have become standard tools in biomedical research in general and in HRV science specifically over the past decade. Gradient boosting, random forest, deep learning, and Related approaches consistently achieve strong predictive performance on biological data sets, and they are being applied to an increasing range of predicting clinical outcomes, classifying physiological states estimating biological age identifying which participants will respond to a treatment. As these methods have proliferated, a particular interpretive habit has grown up alongside them. Researchers examine the feature importance rankings that predictive models produce, the model's own assessments of which input variables contributed most to prediction accuracy, and treat those rankings as rankings of biological relevance. If a gradient boosting model assigns high importance to a particular HRV measure for predicting some outcome, the implicit or explicit interpretation is that this measure is more biologically central to the outcome and that measures the model ranked lower. Hayano and Yuta set up a clean empirical test of whether that interpretive move is justified. They computed 12 HRV indices from each of the 24 hour recordings in their dataset, covering the standard time domain measures like RMSSD and sdnn, frequency domain measures including low frequency and high frequency, spectral power and geometric and nonlinear indices. They then built two classes of models. The first class used spline based regression, if interpretable, transparent models that describe how each HRV index actually changes across the lifespan, fitting smooth curves to the data that capture the genuine biological trajectory of each measure from birth through extreme old age. These are explanatory models. Their parameters directly represent biological patterns and are readable by any researcher who examines them. The second class used machine learning ensemble methods, specifically random forest and gradient boosting, which capture complex nonlinear relationships and interactions between variables without being told in advance what those relationships look like. These are predictive models. They are flexible and powerful, but their internal representations are opaque and what they have learned is not directly interpretable in biological terms. The first comparison the paper makes is whether the machine learning models were better than the spline models at characterizing how the 12 HRV indices actually change with age. This is an explanatory comparison, not asking which model predicts age from HRV most accurately, but asking which model most faithfully describes the biological trajectories of HRV measures across the human lifespan, the answer was essentially no. Advantage for the machine learning models. For the task of understanding how RMSSD or high frequency power or any of the other 12 indices genuinely changes from infancy through centenarian age. The flexible machine learning approaches offer no meaningful additional insight over the interpretable spline based models. If you want to understand the biology of how cardiac autonomic function changes across the lifespan, what the actual age trajectories look like when different measures peak and decline, how steep the changes are at different life stages. A well fitted spline model captures that just as well as a gradient boosting model, and the spline model has the substantial additional advantage of being directly interpretable. The second comparison was whether the machine learning models were better at predicting a participant's age from their 12 index HRV profile, using the HRV signature to estimate how old the person is. Here, gradient boosting did show a modest advantage. Its age predictions were somewhat more accurate on average than those from the linear models, demonstrating that the additional flexibility the ensemble methods bring does purchase some additional predictive power for this specific task, but the advantage was modest and was not uniform across the age range. Young adults and the very elderly were harder to predict accurately for all models, reflecting the known complexities of HRV patterns at those life stages. And now the central finding, which is the point the whole paper is building toward. Hayano and Yuta compared the feature importance rankings the gradient boosting model produced for the age prediction test, the model's assessment of which HRV measures were most useful for guessing a person's age against the actual patterns of age related change that the explanatory spline models identified. These two rankings did not match the measures the predictive model ranked as most important for age prediction were not reliably the measures that showed the strongest and most biologically coherent age related trajectories in the explanatory analysis and measures that were empirically the most robustly and consistently linked to age in a biological sense were not necessarily the ones that the predictive model ranked highly. The predictive importance ranking and the biological relevance ranking came apart. The explanation for this divergence is not mysterious, but it is important to understand clearly. Feature importance in a machine learning model reflects a variable's statistical contribution to predictive accuracy in that specific data set. Given that specific model architecture and given the specific set of other variables included in the model. A measure can rank highly in feature importance because it genuinely matters biologically, but it can also rank highly because it is correlated with the true biological cause without being that cause, or because it captures variants that other variables in the model happen not to capture even if that additional variance is physiologically peripheral. It can also rank highly for reasons that are specific to the training data, particular distributional features of the sample, correlations that exist in this data set but might not generalize to other populations, or interactions with other variables that are artifacts of the data rather than features of the biology. The machine learning model has no knowledge of anatomy, physiology, or mechanism. It has knowledge of statistical patterns. When statistical patterns and biological mechanisms align. Feature importance can look informative when they diverge, and Hayano and Yuta demonstrate empirically that they do diverge for HRV features. In this large and well characterized lifespan sample, the feature importance output can actively mislead researchers who treat it as a biological ranking. The implications of this finding extend well beyond the specific case of age and HRV researchers. Across the spectrum of HRV applications, clinical outcome prediction, physiological state classification, training, load monitoring, stress biomarker research are routinely publishing feature importance rankings and interpreting them as discoveries about which aspects of cardiac autonomic function matter most for the outcome of interest. Hayano and Yuta's paper provides empirical grounds for treating those interpretations with much greater caution. Feature importance outputs from predictive models can and should continue to be published. They are useful for the technical task of choosing features for prediction systems, and they can generate hypotheses about biological importance. But those hypotheses require independent empirical validation before they can be accepted as findings about mechanism. That validation means testing the hypothesis using traditional regression analysis, experimental manipulation, or causal inference methodology approaches specifically designed to evaluate explanatory claims rather than predictive claims. For anyone working in HRV research, and particularly for anyone who has been using machine learning feature importance as a guide to biological interpretation, this paper is essential reading. It makes its argument not as an abstract methodological concern, but as a concrete empirical demonstration using a data set of scale and quality that makes the finding hard to dismiss. Before moving to the synthesis, I want to spend a few moments with what each of these papers implies for how practitioners actually work with HRV data. Because the gap between what research findings mean in a scientific context and what they mean in an applied context is one that this show has always tried to bridge. The sex differences review has an immediate practical implication that is easy to state when you are working with a client on HRV based training or recovery monitoring. Knowing their sex is not a minor detail to note for demographics. The evidence suggests that the normative comparisons and the population level research underpinning your interpretation may have been generated from samples that mixed sexes without examining whether the relationships they found differed between them. If a piece of research you are drawing on established say that a particular HRV threshold is associated with overtraining risk and that research used a mixed sex sample and controlled for sex as a covariate, you do not know whether that threshold applies equally to men and women because the study was not designed to tell you. That is not a reason to abandon evidence based practice. It is a reason to hold population level HRV benchmarks with somewhat more humility and await individualized baseline tracking, which the optimal HRV methodology centers more heavily than cross sectional population comparisons when making inferences for an individual client. This the Touch study from Della Longa and colleagues raises a question for practitioners who work with young people in physical activity, sport, or therapeutic settings. How much systematic attention do we give to the role of appropriate, safe physical contact as part of the regulatory environment we create?
[00:39:42] Research on c tactile afferents has established that effective touch gentle, slow skin, appropriate contact activates autonomic and neural pathways that are specifically responsive to social safety cues and that these effects are measurable at the level of heart rate, autonomic balance, and subjective well being.
[00:39:56] The study with adolescents following virtual exclusion is a controlled demonstration of what practitioners in sport and therapeutic settings often observe in less controlled ways that a hand on the shoulder, a reassuring touch to the forearm, a brief physical contact that communicates care and presence can genuinely shift how a young person's nervous system responds to stress and perceived rejection. This is relevant for coaches, teachers, therapists, and anyone who designs environments for young people. And the HRV data makes the mechanism concrete rather than anecdotal. The Hayano and Yuta paper on machine learning and HRV has a direct implication for how any of us reads HRV research published in the last several years.
[00:40:29] When you encounter a study that used a machine learning model and reports which HRV features the model found most important the finding to attend to is the prediction accuracy, whether the model was better at its prediction tasks than simpler alternatives, whether the model generalizes to new data, what the actual performance numbers look like. The feature importance ranking is worth reading and thinking about as a hypothesis generator. These are the features that contributed most to prediction in this dataset, but it is not evidence that those features are more biologically fundamental than the ones ranked lower. The biological ranking which features are most causally and mechanistically linked to the outcome of interest requires a different kind of study to establish, and it may not look anything like the predictive ranking. Keeping that distinction in mind when reading the growing body of machine learning based HRV research will protect you from conclusions that sound more certain than the evidence can actually support. The Urban Streetscape study from Huang and colleagues is a useful reminder of something that applies directly to the design of measurement contexts in applied HRV work. If different types of visual and spatial environments can produce measurably different HRV profiles during brief passive exposure to photographs, then the environment in which a person is measured is not an irrelevant background variable A person who measures their resting morning HRV while sitting by a window with a view of a park may produce a systematically different measurement than they would while sitting in a bright built enclosed space independently of their actual recovery state. This is an extreme version of a broader point.
[00:41:42] The context shapes what HRV reflects, but the streetscape research gives it a specific and quantifiable form. Standardizing the measurement context, including the visual and physical environment, is one more reason why the consistent morning protocol that characterizes good HRV practice matters, and one more source of variability that practitioners should be aware of when interpreting daily data. Looking across the seven studies together, a few threads emerge with enough consistency to be worth naming explicitly. The most persistent is the theme of context and individuality, the recognition that HRV does not have a fixed context independent meaning, meaning that you can read off a number without knowing the circumstances around it. The sex differences review from Yoon and colleagues shows that the same numerical HRV value may reflect meaningfully different neural processes and regulatory pathways depending on the sex of the person producing it. The co sleeping study from Montoya and colleagues shows that the same person's HRV changes measurably depending on who is sharing their bed, and that the change in autonomic indices does not align cleanly with with the change in sleep quality measures. The context of the measurement matters for which story the data tells. The touch study from Della Longa and colleagues shows that a brief and gentle social interaction can shift the autonomic state of an adolescent, following social stress in a measurable and consistent direction. The social context of the person's experience shapes the autonomic state. The researcher then measures. The streetscape study from Wang and colleagues demonstrates that the visual and spatial environment surrounding a person leaves a physiological trace that can be detected in HRV and related measures even during a brief viewing experience. Taken together, these four papers reinforce something that practitioners who work closely with HRV data often know intuitively but that the research literature sometimes underweights. HRV is not a fixed trait but a continuously updated reflection of how an organism is relating to its biological, social, physical, and environmental circumstances, moment to moment. Interpreting it requires knowing something about those circumstances, not just reading the number. The second thread is methodological, and it runs through three of this week's seven papers in different ways. The Hayano and Yuta paper identifies a gap between how predictive machine learning tools are used in HRV research and what those tools can legitimately claim to demonstrate about biology. This is a gap that has probably been widening for several years as machine learning methods became more common and accessible without a corresponding increase in clarity about the difference between predictive and explanatory inference. The field needed a paper that made this argument empirically with a data set large enough to be convincing, and Hayano and Yuta have written it. The Shoe Gear framework paper makes a contribution of a different methodological kind. It removes a technical barrier to a class of research on brain heart coupling during sleep that was previously too labor intensive to scale, and in doing so it potentially widens the set of research groups that can contribute to that area. And the Kataharti AVNS pilot exemplifies the kind of honest, carefully bounded early stage empirical work that identifies a signal worth investigating while being transparent about the significant limitations on what can be inferred from it. These are all ways of improving the quality of the evidentiary foundation the field is building on, and they represent a kind of scientific maturity, a willingness to do the harder, slower work of establishing what we actually know rather than what we can make sound plausible. The co sleeping study from Montoya and colleagues also raises a point that extends beyond its specific findings. This paper used a within person design which is methodologically stronger than a between persons comparison for the question it was asking. But with in person observational designs have their own characteristic limitation. They can identify what is systematically different between two conditions that person experiences, but they cannot rule out the time varying confounders that tend to differ between those conditions along with the variable being studied. People who sleep with a companion on particular nights may do so for reasons emotional states, social circumstances, health conditions that are themselves associated with different autonomic profiles. And disentangling the effect of the companion from the effect of the surrounding circumstances is not possible from observational data alone. The study is honest about this. The right next step for this research would be a randomized experiment, difficult with bed companions but not impossible, or a quasi experimental approach that uses natural variation in co sleeping patterns to isolate the causal effect. For now, the study provides a well measured correlational picture that practitioners can hold with appropriate context. The brain heart coupling framework from Shoukir and colleagues similarly points forward. The value of the framework is not only in the specific findings from the 18 sleep apnea patients it was demonstrated on, but in the class of research questions it makes newly tractable. When that algorithm has been validated against expert scoring and the framework has been applied to larger and and more diverse populations, we will have a much richer empirical picture of how brain and heart coordinate during sleep and health and disease. What strikes me pulling back to look at the week's papers together is how far the HRV research enterprise has traveled from its origins as a measurement looking for applications. The question is no longer simply whether HRV tells us something useful. That question has been answered. By now the questions are more demanding. Useful to whom? Under what conditions, measured how, interpreted through which analytical framework, and compared against what baseline? Those are the questions that these seven papers are engaged with, and they are harder than the foundational questions were. Answering them well will require the kind of methodological rigor that the Hayano and Yuta paper calls for, the kind of systematic infrastructure building that Shoukir and colleagues modeled, the kind of careful attention to biological context that the sex differences review exemplifies, and the kind of honest limitation disclosure that the TAVNS pilot demonstrates. The field is capable of all of these things, and weeks like this one, with seven papers as diverse as these seven, each doing something substantive in its own domain, are evidence that it is doing them. Thank you for spending this time with me on this Week in hrv. If you found something in today's episode useful, a finding that shifts how you think about a client's data, a methodological point that is going to change how you design something, or simply a paper that you are now going to go read, I hope you will share it with someone who would value it. The audience for this research is still smaller than the research deserves, and word of mouth from practitioners who take this science seriously is a meaningful part of how that changes. I find this work genuinely fascinating and I am grateful every week for the researchers who are doing it with such care. I'm Matt Bennett, co founder of Optimal hrv, and I will see you next week.
