Heart Rate Variability in Cognitive Modeling & the Cognitive Sciences

HEART RATE VARIABILITY IN COGNITIVE MODELING AND COGNITIVE SCIENCES

A Bibliometric Knowledge Base — Synthesis v1

1. HISTORY & TIMELINE

1.1 Narrative Overview

Era I — Founding Physiology (1847–1960)

The oldest recorded observation of what would become HRV science belongs to the mid-19th century: Carl Ludwig (1847) and Heinrich Ewald Hering (1869) independently documented that the heart accelerates during inspiration and decelerates during expiration — the phenomenon now called respiratory sinus arrhythmia (RSA). Their work established the central physiological fact: beat-to-beat interval length varies continuously with the respiratory cycle and this variation is neurally, not mechanically, mediated. The mechanism was eventually traced to phasic modulation of vagal efference by brainstem respiratory pattern generators, but the full mechanistic account awaited 20th-century neuroscience. For most of the intervening century, RSA remained a laboratory curiosity: recording beat-to-beat intervals with sufficient temporal precision required a matured clinical ECG, which was not widely available until the 1920s–1930s.

Era II — Clinical Entry: Fetal Monitoring and Cardiac Prognosis (1960s–1980s)

The decisive clinical entry point was obstetric fetal heart-rate monitoring. In the early 1960s, obstetricians established that reduced beat-to-beat variability in the fetal ECG predicted neonatal distress — providing the first actionable HRV biomarker and creating institutional incentives to automate beat-to-beat interval analysis. The paradigm extended to adult cardiology through two landmark findings: Ewing and colleagues (1985) demonstrated that autonomic neuropathy in diabetic patients suppressed HRV, and Kleiger and colleagues (1987) showed that post-myocardial-infarction patients with low SDNN had dramatically elevated mortality risk. Together, these established HRV as a quantitative prognostic index in cardiovascular medicine.

The parallel methodological breakthrough was power spectral analysis of the RR-interval tachogram. Akselrod and colleagues (1981) demonstrated in animal preparations — pharmacological blockade with atropine and propranolol — that distinct frequency bands of the RR spectrum could be assigned to autonomic sources: high-frequency power abolished with vagal blockade (atropine), low-frequency power reduced by both vagal and adrenergic blockade. This animal-pharmacology foundation directly established the autonomic-physiological interpretation of frequency-domain HRV metrics applied to humans. [Animal-model evidence.]

Era III — Metric Standardization (1980s–1996)

Proliferating metrics, incompatible recording protocols, and non-replicable cross-laboratory results produced a reproducibility crisis through the early 1990s. The European Society of Cardiology and the North American Society of Pacing and Electrophysiology convened a joint Task Force whose 1996 paper (Malik et al., Circulation) became the field's anchoring reference. It defined the canonical time-domain metrics (SDNN, RMSSD, pNN50), frequency-domain bands (VLF < 0.04 Hz; LF 0.04–0.15 Hz; HF 0.15–0.40 Hz) and their normalized variants, minimum recording-length requirements (5-minute short-term; 24-hour ambulatory), sampling rate standards, and guidelines for artifact handling. All subsequent HRV research is calibrated against these definitions, and open software tools implementing them continue to proliferate — including modern extensions such as HRnV-Calc, which adds scale-varying HR n-variability metrics alongside the 1996 standards (Niu et al. 2021).

Era IV — Psychophysiology, Theory, and Biofeedback (1990s–2010s)

The transition from cardiology to psychophysiology and cognitive science was driven by three developments. First, Berntson, Cacioppo, and Quigley (1991) proposed the autonomic space model, reframing the ANS as two semi-independent axes (sympathetic, parasympathetic) capable of reciprocal, coactivated, or coinhibited states — dissolving the folk assumption of simple sympathovagal balance and undermining the LF/HF ratio as a unitary index. Second, Stephen Porges (1995, 2001) articulated Polyvagal Theory, positing a phylogenetically ordered hierarchy of vagal systems in which myelinated fibers from nucleus ambiguus mediate social engagement and RSA amplitude indexes the "vagal brake." Third, Thayer and Lane (2000, 2009) proposed the Neurovisceral Integration Model (NIM), arguing that the prefrontal-CAN circuitry regulating autonomic outflow is the same circuitry mediating self-regulatory, executive, and emotional functions — making resting HRV an index of inhibitory control capacity.

These theoretical frameworks simultaneously generated the biofeedback intervention paradigm: Lehrer and Gevirtz showed that breathing at the individual's baroreflex resonance frequency (~0.1 Hz, ~6 breaths/min) produces maximal RSA amplitude and baroreflex engagement, forming the basis for the resonance-frequency HRV biofeedback protocol. Consumer translation of this protocol (HeartMath emWave, Unyte IOM2) introduced HRV coherence as a proprietary metric — a choice that subsequent formal analysis identifies as a systematic instance of reward misspecification, where the optimized proxy diverges from the intended psychological outcome (Bose 2026).

Era V — Computational, Nonlinear, and Machine-Learning Era (2010s–present)

Three converging developments define the current era: ubiquitous wearable sensing, open datasets (PhysioNet/MIT-BIH), and machine learning. Nonlinear complexity metrics — sample entropy, detrended fluctuation analysis (DFA), multiscale entropy — became standard in parallel with 1996 Task Force metrics, motivated by evidence that fractal scaling properties of RR-interval series carry information not captured by linear spectral analysis; these metrics are particularly sensitive in clinical populations including depression (Čukić et al. 2021). Deep learning architectures (1D CNNs, LSTMs, transformers) achieve state-of-art performance on ECG classification and, in lightweight form, on resource-constrained edge devices (Baig et al. 2025). PPG-to-ECG reconstruction via Vision Transformers (Li et al. 2025) and single-ear embedded systems (Santos et al. 2025) extend clinical-grade HRV monitoring to consumer wearables. Cross-modal frameworks mapping HRV features to EEG-based cognitive load markers open a new frontier: using the autonomic signal as a portable surrogate for central neural state (Pradeep et al. 2026; Sasi et al. 2026). Bayesian generative state-space models propose to infer latent sympathetic/parasympathetic outflow trajectories from RR intervals; they recapitulate the linear properties of conventional spectral estimators while improving discrimination of dynamical-complexity metrics across physiological states (Rosas et al. 2023).

1.2 Chronological Milestone List

2. BIOLOGICAL AND SYSTEMIC BASIS

2.1 Cardiac Autonomic Innervation

The sinoatrial (SA) node — the cardiac pacemaker — receives dual autonomic innervation whose interplay is the proximate source of measurable HRV. Sympathetic postganglionic fibers release norepinephrine, activating β1 adrenoceptors and increasing phase-4 SA node depolarization slope, thereby accelerating heart rate and shortening RR intervals. Parasympathetic (vagal) postganglionic fibers release acetylcholine, activating M2 muscarinic receptors via Gi protein, reducing the depolarization slope, slowing heart rate, and lengthening RR intervals. The two divisions differ critically in temporal kinetics: acetylcholine is hydrolyzed by acetylcholinesterase within one cardiac cycle, giving vagal modulation a beat-to-beat temporal resolution; sympathetic modulation acts over seconds. This asymmetry is the physiological basis for the frequency-domain interpretation of HRV: HF power (0.15–0.40 Hz) primarily reflects vagal modulation, LF power (0.04–0.15 Hz) reflects mixed sympathovagal contributions dominated by baroreflex oscillations near 0.1 Hz. The autonomic space model emphasizes that these two axes can be co-activated, co-inhibited, or reciprocally coupled — not merely balanced on a single dimension.

2.2 Respiratory Sinus Arrhythmia (RSA)

RSA is the rhythmic heart-rate oscillation synchronized with the respiratory cycle: acceleration during inspiration, deceleration during expiration. It arises from two coupled mechanisms: (a) central inhibition of cardiac vagal motor neurons by the brainstem respiratory pattern generator (ventrolateral medulla) during inspiration, reducing vagal tone; (b) pulmonary stretch-receptor afferents that further modulate vagal efference across the respiratory cycle. RSA amplitude, measured as HF HRV power or the respiratory-frequency component of the RR-interval spectrum, is therefore a non-invasive index of cardiac vagal tone. RSA diminishes with sympathetic activation, age, and autonomic neuropathy; it increases with parasympathetic activation, slow diaphragmatic breathing, and exercise training. The functional significance of RSA is debated: the efficiency hypothesis proposes it improves pulmonary gas exchange by matching cardiac output to ventilation phase; the vagal-brake hypothesis (Porges) emphasizes its role in enabling behavioral and social flexibility.

Animal-model evidence [flagged]: Pharmacological blockade studies in dogs (Akselrod et al., 1981) confirmed the autonomic source assignments of HRV frequency bands: atropine (vagal blockade) abolished HF power; propranolol (β-adrenergic blockade) reduced LF power. Vagotomy and VNS lesion studies in cats and rats established dose-response relationships between vagal nerve integrity and RSA amplitude. These animal experiments are the mechanistic foundation for interpreting human frequency-domain HRV.

2.3 The Baroreflex

The arterial baroreflex provides beat-to-beat blood-pressure homeostasis and is the primary generator of LF HRV oscillations. Baroreceptors in the aortic arch (vagal afferents) and carotid sinus (glossopharyngeal afferents) detect arterial wall stretch proportional to blood pressure, projecting to the NTS. From the NTS, a reflex arc modulates vagal efference (via nucleus ambiguus) and sympathetic outflow: rising blood pressure increases vagal tone and slows the heart; falling pressure reduces vagal tone and activates sympathetic drive. The resulting autonomous blood-pressure oscillation at ~0.1 Hz (the Mayer wave) is the physiological source of LF HRV power. This mechanistic detail explains why resonance-frequency biofeedback (breathing at ~0.1 Hz) produces maximal HRV amplitude: respiratory input at the baroreflex natural frequency resonates the feedback loop, entraining blood-pressure and cardiac-interval oscillations into high-amplitude synchrony [Bose 2026 context].

Baroreflex sensitivity (BRS, ms/mmHg) is a related but distinct metric of vagal responsiveness that predicts cardiac prognosis.

2.4 The Central Autonomic Network (CAN)

The CAN is the distributed cortical-subcortical network that generates, integrates, and modulates autonomic outflow. Its principal nodes are:

• Medial prefrontal cortex / ventromedial PFC: top-down inhibitory regulation of subcortical arousal systems; impaired in depression and PTSD, contributing to reduced HRV (field knowledge)
• Anterior cingulate cortex (ACC): conflict monitoring, emotional appraisal; contains Von Economo neurons (VENs), a specialized cell type modeled as implementing a "fast lane" projection pathway for rapid social-emotional decision-making (Keskin 2026); VENs are depleted in frontotemporal dementia and altered in autism. Keskin's model concerns social-decision speed-accuracy tradeoffs and makes no HRV or autonomic claim — the ACC/insula VEN system is offered here only as CAN-architecture background, structurally analogous but not empirically linked to the HRV-cognition relationship
• Anterior insula: interoceptive processing; monitors cardiac and visceral state; reciprocally connected to the NTS; also contains VENs (Keskin 2026)
• Amygdala: threat evaluation; drives sympathetic activation and suppresses vagal tone under perceived danger; the target of PFC inhibitory regulation in the NIM
• Hypothalamus: coordinates autonomic and neuroendocrine responses via descending projections to brainstem nuclei
• Periaqueductal gray (PAG): integrates amygdala and cortical inputs; organizes defensive responses
• Nucleus tractus solitarius (NTS): primary relay for cardiac and visceral afferents; convergence point for all descending CAN inputs; key site for vagal afferent-to-efferent integration
• Nucleus ambiguus: source of myelinated, fast-acting, cardioinhibitory vagal efferents to the SA node; primary substrate of RSA and the vagal brake
• Dorsal motor nucleus of vagus (DMNX): source of slower, unmyelinated vagal efferents mediating metabolic-vegetative functions

The CAN produces HRV through descending modulation of NTS and nucleus ambiguus activity. PFC and ACC inhibitory projections constrain amygdala-driven sympathetic activation, maintaining tonic vagal tone. Disruption of this circuit — in depression, anxiety, PTSD — reduces vagal tone and HRV (in depression, Čukić et al. 2021; the PTSD association is field knowledge). The VEN "fast lane" hypothesis (Keskin 2026) is a computational model of social-decision speed-accuracy tradeoffs in ACC and frontal-insula VENs; it makes no autonomic or HRV prediction and is included here only as CAN-architecture background.

2.5 Vagal Afferent/Efferent Pathways and Interoception

The vagus nerve is predominantly afferent (~80% of fibers): cardiac mechanoreceptors and chemoreceptors project to the NTS, relaying via parabrachial nucleus and PAG to hypothalamus and forebrain. This ascending pathway is the neural substrate of interoception — the continuous inference of internal bodily state. Impaired interoceptive processing (hypervigilant or blunted) characterizes PTSD and other conditions of autonomic dysregulation; mindfulness and MBSR are proposed to recalibrate interoceptive representations via sustained attention to afferent body signals (Kang et al. 2020). Efferent pathways divide into myelinated B-fibers from nucleus ambiguus (fast, RSA-producing, socially relevant per PVT) and unmyelinated C-fibers from DMNX (slow, metabolic/vegetative).

Bayesian outflow estimation: Standard spectral decomposition of HRV provides aggregated, noisy estimates of sympathetic and parasympathetic outflow. Bayesian generative state-space models fitted to RR-interval series propose to infer latent autonomic outflow trajectories as an alternative to spectral decomposition; the approach recapitulates the linear properties of conventional estimators and shows advantages mainly on dynamical-complexity metrics rather than a demonstrated gain in autonomic-outflow precision (Rosas et al. 2023).

3. THEORIES LINKING HRV, NERVOUS SYSTEM, COGNITION, AND BEHAVIOR

3.1 Neurovisceral Integration Model (NIM)

Core claims (Thayer & Lane, 2000, 2009): The prefrontal-CAN circuitry that regulates autonomic outflow is the same circuitry mediating self-regulatory, executive, and emotional functions. Resting HRV therefore indexes the functional integrity of this shared substrate — high resting HRV reflects efficient prefrontal inhibitory control over subcortical threat-response circuits; low HRV reflects disinhibited subcortical activation (the "default response" of chronic physiological threat engagement). Testable predictions: (1) higher resting HRV → better performance on executive function, attention, and working-memory tasks; (2) HRV reactivity (vagal withdrawal during challenge; recovery after) tracks adaptive self-regulatory engagement; (3) populations with impaired executive function or emotion regulation show reduced resting HRV.

Supporting evidence: Meta-analyses document a small-to-moderate positive correlation between resting HRV and executive function performance (attention, inhibition, working-memory). HRV is reduced in depression (Čukić et al. 2021) and, per the broader field literature, in PTSD and other conditions of autonomic dysregulation. Cross-modal frameworks report that HRV features carry cognitive-state information correlated with EEG-based load markers, though the cross-modal surrogacy result is modest (Pradeep et al. 2026; Sasi et al. 2026).

Challenges: The NIM has been criticized for circularity (HRV and executive function share PFC substrate by definition, making the correlation uninformative about causal direction), for not specifying the mechanism linking vagal tone to cognitive performance in sufficient mechanistic detail, and for inflated meta-analytic effect sizes due to publication bias. The LF/HF ratio widely used in NIM-adjacent studies is confounded by respiratory rate and carries ambiguous physiological meaning, a concern central to the autonomic space model.

3.2 Polyvagal Theory (PVT)

Core claims (Porges, 1995, 2001, 2007): The ANS is organized into a phylogenetically ordered hierarchy of three defense/engagement tiers: (1) myelinated ventral vagal complex (nucleus ambiguus → SA node) supports social engagement, calm-and-connect behavior, and RSA; (2) sympathetic nervous system mobilizes fight-or-flight; (3) unmyelinated dorsal vagal complex (DMNX) drives immobilization/freeze/shutdown. The "vagal brake" — tonic myelinated vagal inhibition of the SA node — regulates transitions between tiers. RSA amplitude is its index. The "vagal tank" metaphor captures individual capacity for vagal engagement, indexing regulatory reserve. A coordinated Social Engagement System (myelinated vagus, facial/middle-ear muscles via cranial nerves V, VII, IX, X, XI) enables prosocial and cognitive behavior contingent on perceived environmental safety.

Supporting evidence: The hierarchical behavioral response sequence (engagement → fight/flight → freeze) is broadly consistent with observed behavior across species. RSA as a myelinated vagal index is well-validated. PVT has generated productive clinical applications in trauma-informed therapy, somatic interventions, PTSD treatment, and MBSR (Kang et al. 2020), and has informed HRV biofeedback's therapeutic rationale (Bose 2026).

Challenges: PVT has attracted sustained methodological criticism. The claimed phylogenetic sequence of vagal myelination — central to the theory — has been challenged by comparative neuroanatomists who observe unmyelinated cardiac vagal fibers in mammals and question the clean nucleus ambiguus/DMNX functional segregation. Some predictions are not readily falsifiable. The theory draws causal inferences from correlational autonomic data. It remains highly clinically influential while remaining contested at the neuroanatomical and evolutionary levels.

3.3 Autonomic Space Model

Core claims (Berntson, Cacioppo & Quigley, 1991): Autonomic regulation occupies a two-dimensional space defined by semi-independent sympathetic and parasympathetic axes. Possible coupling modes include reciprocal (classical sympathovagal "balance"), coactivation (both increase simultaneously, as in exercise), coinhibition (both decrease, as in quiet sleep), and uncoupled (one changes independently). The LF/HF ratio is an intrinsically deficient summary because it collapses two independent dimensions.

Supporting evidence: Pharmacological dual-blockade studies (atropine + propranolol) confirm independent manipulability of both axes. Coactivation is documented in exercise and competitive stress; coinhibition in restorative sleep.

Relationship to other theories: The autonomic space model is primarily a framework correction rather than a theory of HRV-cognition. It provides the conceptual foundation for why single-index HRV interpretations are underdetermined and why measurement context (posture, breathing, activity level) must always be specified. It is implicit in the Bayesian outflow estimation approach (Rosas et al. 2023), which estimates two independent outflow dimensions rather than a single balance metric.

3.4 Interoceptive Predictive Processing Framework (Emerging)

An emerging framework — not yet dominant but rapidly gaining traction — proposes that interoception is implemented as Bayesian predictive processing: the brain continuously generates predictions about visceral state (including cardiac rhythm) and updates those predictions from ascending afferent signals, with prediction errors driving learning and arousal responses. Under this account, HRV reflects the brain-body informational loop rather than pure autonomic outflow. Hypervigilant interoceptive prediction errors produce anxiety-related autonomic dysregulation; PTSD involves locked-in high-threat priors that drive hyperarousal; mindfulness practice is proposed to recalibrate interoceptive priors through sustained present-moment attention to afferent signals (Kang et al. 2020). The Bayesian generative state-space model for autonomic outflow estimation (Rosas et al. 2023) can be read as a computational implementation of these principles at the signal level.

Current status: Theoretically productive but largely correlational; few direct tests of the predictive-processing interpretation against NIM or PVT predictions. The RCT showing MBSR reduces PTSD severity with EEG changes consistent with interoceptive mechanism offers indirect support (Kang et al. 2020), but the interoceptive interpretation depends on EEG proxies rather than cardiac/visceral measures.

4. ESTABLISHED PARADIGMS AND PROTOCOLS

4.1 Recording Standards and Signal Acquisition

Short-term recording: Minimum 5 minutes under standardized resting conditions (supine or seated; controlled ambient environment; paced or monitored respiration). Sufficient for all time-domain metrics and frequency-domain LF/HF power; not suitable for SDNN clinical interpretation or VLF. The standard in psychophysiological research.

24-hour ambulatory recording: Continuous Holter ECG. Required for SDNN (recording-length dependent by definition), VLF components, and circadian variability assessment. Standard in clinical cardiac risk assessment.

ECG vs. PPG: ECG remains gold standard; R-peak timing from standard lead achieves millisecond precision. Photoplethysmography (PPG) from wearable devices provides pulse peak intervals susceptible to motion artifact and waveform-dependent timing error. The gap between PPG and ECG-quality HRV is being addressed by deep-learning reconstruction: a multichannel ViT operating on four PPG-derived channels (raw signal, first-order difference, second-order difference, AUC) achieves superior PPG-to-ECG reconstruction (Li et al. 2025). Ear-based dry-electrode systems achieve real-time ECG reconstruction and HRV computation on embedded platforms without gel electrodes (Santos et al. 2025), extending continuous HRV monitoring to naturalistic cognitive contexts.

Sampling rate: Minimum 250 Hz for clinical R-peak precision; 500–1000 Hz preferred. PPG typically sampled at 64–256 Hz.

Artifact handling: Ectopic beats, movement artifacts, and signal dropouts create spurious RR intervals that distort all HRV metrics. Standard pipelines screen for physiologically implausible interval values and apply cubic-spline interpolation, deletion, or replacement before analysis. Automated artifact correction is implemented in HRnV-Calc (Niu et al. 2021), Kubios, pyHRV, and neurokit2.

Open-source software: HRnV-Calc (Niu et al. 2021) provides the first open implementation of multi-scale HR n-variability metrics alongside conventional time/frequency-domain HRV metrics; HRnV metrics (computed on sub-sampled or nth-interval series) showed improved triage-performance over standard HRV in pilot clinical studies. Standard ecosystem tools include Kubios HRV, pyHRV, HRV-Analysis (MATLAB/R), and neurokit2 (Python).

4.2 Paced Breathing and Resonance-Frequency Protocols

Spontaneous breathing rate varies from ~8–20 breaths/min; this variance shifts RSA energy within and across the HF band and — critically — into the LF band at rates below ~9 breaths/min, confounding LF/HF interpretation. Experimental protocols therefore either fix respiratory rate or measure it continuously:

• Paced breathing at spontaneous rate: simple rate stabilization; reduces inter-individual variance
• Resonance-frequency breathing: typically 0.1 Hz (~6 breaths/min, 5-second inhale / 5-second exhale); resonates with the baroreflex feedback loop, producing maximal RSA amplitude and what HeartMath terms "HRV coherence." Individual resonance frequencies vary (approximately 4.5–7 breaths/min) and can be identified via ascending-frequency search protocol
• Respiratory monitoring: concurrent pneumography or airflow measurement enables RSA computation independent of fixed-rate assumption

4.3 HRV Biofeedback Paradigms

Lehrer/Gevirtz protocol: Subjects receive real-time HRV oscillation feedback (amplitude and/or coherence display) while practicing resonance-frequency breathing. Typical format: 10 sessions × 20 minutes, plus daily home practice. Outcome measurements: within-session RSA amplitude gain; between-session resting HRV shift; self-reported and task-based self-regulatory outcomes.

Consumer devices and their limitations: HeartMath emWave, Unyte IOM2, and Muse (EEG-based) devices gamify this protocol. Formal analysis identifies three systematic failure modes from reward misspecification (Bose 2026):

1. Proxy mismatch: coherence scores measure a narrow-band spectral proxy, not autonomous function or wellbeing
2. Strategy shortcutting: users optimize the displayed metric (e.g., breath pacing) without producing the underlying physiological or psychological change
3. Transfer failure: improvements in in-session coherence do not reliably generalize to trait autonomic regulation or wellbeing

EEG biofeedback (neurofeedback) exhibits parallel misspecification between target EEG power bands and actual cognitive/clinical outcomes (Bose 2026). The implication for research is that clinical efficacy claims resting on device-defined metrics rather than pre-specified psychological endpoints carry substantial uncertainty.

4.4 Cognitive and Stressor Reactivity Paradigms

Standard paradigms for probing HRV reactivity to cognitive demand and psychosocial stress:

• Stroop Color-Word Task: cognitive interference/inhibition; reliable HR increase and vagal withdrawal (HRV suppression)
• Mental Arithmetic Task (MAT): sustained mental effort with mild stress; robust HRV reduction
• Trier Social Stress Test (TSST): public speech + mental arithmetic before an evaluative panel; strong psychosocial stressor; produces cortisol, HR, and HRV reactivity
• N-back working memory tasks: parametric load manipulation (0-back → 2-back → 3-back); typically associated with graded vagal withdrawal under increasing load (field knowledge; no direct HRV/n-back study in the available cards)
• Sustained attention/vigilance tasks: prolonged target-detection monitoring; resting HRV predicts vigilance performance consistent with NIM
• Passive listening / rest conditions: control comparators in cross-modal HRV-EEG paradigms (Pradeep et al. 2026; Sasi et al. 2026)

Workload-aware adaptive interfaces have been proposed in which real-time physiological signals drive automatic task-difficulty adjustment, notification deferral, or suppression of attention-demanding UI elements during peak cognitive demand (Kosch 2020 — a vision paper grounded in EEG and eye-tracking sensing rather than HRV).

4.5 Standard Metrics Toolkit

Time-domain (from 1996 Task Force):

• SDNN: SD of all NN intervals; total variability; recording-length dependent
• RMSSD: root mean square successive differences; beat-to-beat vagal modulation; preferred for short-term recordings; robust to non-stationarity
• pNN50: proportion of successive intervals differing >50 ms; vagal index; less sensitive than RMSSD in short recordings

Frequency-domain (via FFT or autoregressive modeling):

• VLF (< 0.04 Hz): long recordings only; composite sympathetic/vagal/thermoregulatory contributions; not interpretable in 5-minute segments
• LF (0.04–0.15 Hz): mixed sympathovagal; baroreflex oscillation dominant at ~0.1 Hz
• HF (0.15–0.40 Hz): predominantly vagal; directly corresponds to RSA when breathing is in-band
• LF/HF ratio: contested "sympathovagal balance" index; confounded by respiratory rate, posture, and recording conditions; inconsistent with the autonomic space model's orthogonality claim

Nonlinear and complexity metrics:

• Poincaré plot SD1/SD2: SD1 ≈ RMSSD/√2 (beat-to-beat vagal); SD2 reflects short- and long-term variability
• Sample entropy (SampEn): regularity measure; lower in disease states; reduced in depression (Čukić et al. 2021)
• Approximate entropy (ApEn): predecessor to SampEn; biased but widely cited
• DFA α1/α2: short- and long-range fractal scaling exponents; α1 (4–16 beats) is a nonlinear vagal/complexity index; altered in cardiac disease and mood disorders
• Multiscale entropy (MSE): entropy at multiple coarse-graining scales; captures complexity across timescales unavailable to single-scale metrics
• HRnV metrics: multi-scale extension of standard RR-interval metrics using nth-interval sequences; open implementation in HRnV-Calc (Niu et al. 2021)

Common pitfalls:

1. Breathing confound: uncontrolled respiratory rate shifts RSA across LF/HF boundary; LF/HF ratio becomes uninterpretable without concurrent respiratory monitoring
2. Recording-length dependence: SDNN derived from 5-minute recordings is not comparable to 24-hour SDNN by design
3. ECG vs. PPG artifact propagation: PPG peak-timing errors introduce systematic noise into entropy and DFA estimates beyond spectral metrics
4. Artifact-correction insufficiency: uncorrected ectopic beats dramatically inflate or suppress entropy, DFA, and spectral metrics; pipeline choice materially affects results
5. Non-stationarity: trending HR (stress induction, postural change) requires detrending or segment-stationarity handling before valid spectral estimation

5. COMPUTATIONAL / STATE-OF-THE-ART MODELING

5.1 Signal Processing Pipelines

R-peak/QRS detection: The Pan-Tompkins algorithm (1985) — bandpass filtering, differentiation, squaring, windowed integration, adaptive thresholding — remains the reference standard in research pipelines for clean ECG. Wavelet-based and matched-filter detectors address low-SNR or non-standard lead configurations. Lightweight 1D CNNs achieve competitive arrhythmia-class detection on resource-constrained edge hardware, enabling embedded deployment (Baig et al. 2025). Gradient-weighted class activation mapping (Grad-CAM) applied to these CNNs provides visual explanations of classification decisions, supporting interpretable clinical use (Baig et al. 2025).

Artifact correction: After R-peak detection, the inter-beat interval series is screened for physiologically implausible values (ectopics, missed beats, noise bursts). Standard methods include threshold-based exclusion, cubic-spline interpolation, and Poincaré-plot outlier rejection. HRnV-Calc (Niu et al. 2021) and Kubios implement automated pipelines.

PPG-to-ECG reconstruction: A multichannel ViT using raw PPG plus three derivative representations jointly captures temporal and waveform-morphological features via self-attention, achieving superior reconstruction over single-channel baselines (Li et al. 2025). Embedded real-time ECG reconstruction from ear-based dry-electrode biosignals enables continuous HRV monitoring in ambulatory cognitive contexts on the BioGAP platform (Santos et al. 2025).

5.2 Classical and Nonlinear-Dynamics Models

Point-process / integrate-and-fire models: The heartbeat generation process is modeled as a stochastic integrate-and-fire (IF) point process in which the SA node fires when an accumulated depolarization signal crosses threshold. RSA is captured as respiratory-frequency modulation of the input or threshold. Bayesian state-space formulations extend this: the latent instantaneous HR and autonomic outflow are treated as hidden states estimated via Bayesian filtering (Kalman or particle filter) from observed beat times (Rosas et al. 2023). This yields instantaneous HR estimates (vs. windowed averages), continuous adaptation to non-stationarity, and principled uncertainty quantification — advantages directly relevant to cognitive-state tracking in real time.

Nonlinear oscillator/coupling models: Coupled-oscillator models (e.g., DeBoer, van Leeuwen tradition) represent the cardiovascular system as interacting nonlinear oscillators (SA node, respiratory pattern generator, baroreflex) whose coupling generates the emergent HRV spectrum. These mechanistic models can be fit to individual data to infer coupling parameters.

Bayesian generative state-space model as an emerging approach: (Rosas et al. 2023) propose a Bayesian generative approach to RR-interval dynamics as a principled alternative to standard spectral decomposition; the paper recapitulates the linear properties of conventional estimators and demonstrates improved discrimination mainly on dynamical-complexity metrics, rather than a head-to-head gain in autonomic-outflow precision. This represents a methodological frontier for autonomic outflow estimation, not yet a validated replacement for spectral methods.

5.3 Machine Learning and Deep Learning

Task taxonomy and state-of-art by task:

Cross-modal cognitive monitoring (emerging frontier): Supervised ML frameworks train on simultaneous ECG and EEG recordings to learn mappings from HRV features to EEG-derived cognitive load markers. Both XGBoost-based (Sasi et al. 2026) and a dedicated cross-modal framework using HRV features plus Catch22 time-series descriptors (Pradeep et al. 2026) achieve meaningful correspondence with EEG cognitive load markers during working-memory tasks, using the publicly available OpenNeuro multimodal EEG+ECG dataset. Limitations: paradigm-specific validation; EEG ground-truth labels carry their own noise; generalization across task types unestablished.

Manifold learning for personalized classification: Nonlinear manifold learning preserves inter-individual ECG pattern variability overlooked by linear dimensionality reduction (PCA), enabling personalized label-free arrhythmia clustering with better cross-individual generalization (Vazifeh et al. 2025). While applied to arrhythmia detection rather than cognitive state, the methodology is transferable to personalized HRV-state classifiers.

Is there a consensus state of the art? For arrhythmia classification on clean ECG: deep learning (1D CNN/transformer) is current best practice; edge-deployable lightweight CNNs close the gap with resource constraints (Baig et al. 2025). For wearable HRV: multichannel transformer reconstruction is emerging SOTA (Li et al. 2025). For cognitive-state monitoring from HRV: no consensus; feature-based ML (XGBoost + HRV metrics + Catch22) is the validated practical approach with modest accuracy (Pradeep et al. 2026; Sasi et al. 2026). For autonomic outflow estimation: Bayesian state-space is a promising generative alternative that recapitulates conventional linear estimators and improves complexity-metric discrimination (Rosas et al. 2023), but not yet the field's adopted standard.

Open toolkits and benchmarks:

• Datasets: PhysioNet/MIT-BIH Arrhythmia Database; BIDMC 40-subject dataset; OpenNeuro multimodal EEG+ECG (Pradeep et al. 2026)
• Analysis software: HRnV-Calc (Niu et al. 2021); Kubios HRV; pyHRV; neurokit2; WFDB (PhysioNet toolbox)
• Consumer device specifications: HeartMath emWave; Unyte IOM2; Muse (Bose 2026)

6. COGNITIVE-SCIENCE APPLICATIONS

6a. ATTENTION AND EXECUTIVE FUNCTION

Framing: Per the NIM, resting HRV indexes the functional capacity of the prefrontal CAN circuitry shared with attentional control, working memory, and response inhibition. High resting HRV should predict better attentional performance; cognitive load should produce characteristic vagal withdrawal.

Reported relationships (evidence strength varies):

HRV as attentional-capacity index: Meta-analytic literature (not represented in available cards but established in the field) documents a small-to-moderate positive correlation between resting RMSSD/HF-HRV and performance on sustained attention, attention-switching, cognitive inhibition, and working-memory tasks. This is consistent across multiple paradigms and populations, though effect sizes are modest (r ~ 0.20–0.35).

HRV suppression under cognitive load: N-back and sustained-attention tasks are generally associated with vagal withdrawal under increasing demand (field knowledge; not established by the available cards). Adaptive cognitive-augmentation interfaces that would use such signals as input have been proposed (Kosch 2020 — a vision paper based on EEG and eye-tracking, not HRV). The HRV signal is plausibly informative for sustained-load monitoring, though latency constraints (windowed averaging) limit application to rapid-adaptation contexts.

HRV-EEG cognitive surrogate: ECG-derived HRV features (time-domain, nonlinear, Catch22 descriptors) trained to predict EEG spectral and Catch22 cognitive load markers achieve meaningful cross-modal correspondence during working-memory vs. passive-listening tasks (Pradeep et al. 2026; Sasi et al. 2026). XGBoost with HRV + Catch22 ECG features outperforms simpler classifiers (Sasi et al. 2026). These are early, single-dataset preprint results (both 2026, non-peer-reviewed, OpenNeuro / self-collected; single working-memory + listening paradigm). They suggest, but do not establish, that HRV could serve as a portable surrogate for EEG-based cognitive-load monitoring. This is the field's most immediately testable and most methodologically vulnerable claim: neither study measures or controls respiratory rate, and Sasi et al.'s feature set includes LF/HF metrics known to be respiration-confounded — so a shared respiratory artifact remains a live alternative explanation for the cross-modal correspondence. Reported accuracies are also inconsistent across the two papers (Sasi et al.'s near-ceiling 98–99% is a within-modality result, whereas Pradeep et al.'s actual cross-modal HRV→EEG projection reaches only ~40% against a 33% chance baseline), and neither reliably detects the resting baseline condition. See Final Hypothesis 1, which specifies the respiratory-deconfounding test that would decide the question.

Neural substrate (background): The ACC and anterior insula — both CAN nodes — are also the regions bearing Von Economo neurons, which Keskin 2026 models as a "fast lane" speed-accuracy mechanism for rapid social decisions (Keskin 2026); VENs are depleted in frontotemporal dementia. This VEN work is CAN-architecture background only: it addresses social-decision dynamics and makes no HRV or autonomic claim, so it is structurally analogous to — not empirical evidence for — an HRV–executive-function link.

Boundary conditions: The resting-HRV-cognition correlation is a consistent group-level effect, not a strong individual predictor. ECG-to-EEG surrogacy is paradigm-specific (Pradeep et al. 2026; Sasi et al. 2026). HRnV multi-scale metrics may capture attention-related variability missed by single-scale standard HRV (Niu et al. 2021).

6b. MINDFULNESS AND CONTEMPLATIVE PRACTICE

Framing: Mindfulness practice is hypothesized to increase parasympathetic tone and vagal flexibility through attention training, interoceptive awareness cultivation, and top-down emotion regulation — all mechanisms that should shift autonomic setpoint toward higher resting HRV.

Established relationships:

MBSR and PTSD: A randomized clinical trial (n = 98 veterans) demonstrated that MBSR produces greater PTSD symptom reduction than an active present-centred group therapy control; EEG changes during resting, meditation, and cognitive-task conditions are consistent with an interoceptive neural mechanism mediating the therapeutic effect (Kang et al. 2020). Direct HRV measurement was not a primary endpoint; the autonomic interpretation is inferred from EEG proxies. Nonetheless, this is the highest-quality evidence (RCT with active control) that mindfulness intervention produces neurophysiological changes in interoceptive-autonomic circuitry relevant to trauma recovery.

Deep meditation and central neural dynamics: LDA of MEG source-reconstructed time series during Samatha focused-attention meditation in experienced Buddhist monks detected dynamic changes in deep brain structures (beyond cortical surface activity), including regions consistent with subcortical CAN nodes (Calvetti et al. 2021). Direct HRV measurement was not reported; the relevance is the demonstration of meditation-induced dynamic change in autonomic-relevant deep structures in highly trained practitioners.

HRV biofeedback as mindfulness-adjacent intervention: Resonance-frequency HRV biofeedback produces physiological states (high-amplitude RSA, heightened interoceptive awareness of cardiac rhythm) overlapping with those cultivated by diaphragmatic breathing in mindfulness practice. Consumer implementations of this protocol exhibit reward misspecification — optimizing device-defined coherence scores rather than genuine autonomic or psychological outcomes — with three identified failure modes (proxy mismatch, strategy shortcutting, transfer failure) (Bose 2026). This substantially qualifies efficacy claims for consumer mindfulness-technology products.

Proposed mechanisms (supported by available evidence):

1. Repeated attention-to-breath training habituates interoceptive afferent pathways and expands their cortical representation
2. Reduced evaluative/ruminative cognition decreases amygdala-driven sympathetic activation and restores vagal tone
3. Diaphragmatic breathing inherent in most mindfulness traditions directly increases RSA amplitude
4. Long-term practice shifts autonomic setpoint (resting HRV increase), though most supporting studies are cross-sectional

Boundary conditions: The most rigorous mindfulness-HRV evidence (RCT) uses an indirect HRV proxy (Kang et al. 2020). Expert-practitioner MEG evidence comes from a small, highly specialized sample (Calvetti et al. 2021). Consumer biofeedback claims rest on device-specific metrics with formalized misspecification concerns (Bose 2026). Long-term longitudinal RCTs with HRV as a primary pre-registered endpoint remain rare.

6c. OTHER COGNITIVE-SCIENCE APPLICATIONS

Emotion regulation and depression: Reduced HRV and increased nonlinear-domain irregularity (lower sample entropy, altered DFA scaling) are found across depression cohort studies, suggesting ANS dysregulation is a transdiagnostic marker of impaired emotion regulation (Čukić et al. 2021). Critically, nonlinear/complexity HRV metrics are more sensitive to depression-related autonomic dysregulation than standard time/frequency-domain linear metrics — endorsing the clinical adoption of entropy and DFA measures beyond the 1996 Task Force standard set (Čukić et al. 2021). Heterogeneity across included studies (medication status, comorbidities, HRV computation methods) limits generalizability of specific effect estimates.

Stress and cognitive workload assessment: HRV is widely used as a non-invasive physiological indicator of cognitive workload and psychological stress, and cognitive load is generally associated with vagal withdrawal (field knowledge; no direct HRV/n-back study appears in the available cards). Adaptive-interface designs have been proposed in which real-time physiological signals adjust task difficulty, defer notifications, or suppress attention-demanding UI elements during peak demand periods (Kosch 2020 — a vision paper based on EEG and eye-tracking, not HRV). Latency constraints (windowed HRV requires seconds of data) challenge ultra-rapid adaptation but are addressable via instantaneous point-process methods [Rosas et al. 2023 context].

Social cognition and rapid decision-making: The VEN "fast lane" hypothesis (Keskin 2026) proposes a computational role for ACC and frontal-insula VENs as sparse, fast-projection neurons enabling rapid social-emotional decisions (τ = 5 ms, 8-afferent fan-in). Selective VEN depletion in FTD and alterations in autism predict specific impairments in rapid choice under social context, consistent with the CAN's role in integrating autonomic state with social cognition. This is an emerging computational-theoretical framework rather than established empirical literature; direct HRV predictions are not yet derived.

Animal welfare as cross-species context [perspective piece]: A canine-welfare perspective argues that multi-biomarker systems approaches incorporating HRV could provide more construct-valid welfare assessment than cortisol alone (Cobb et al. 2025), while noting the absence of standardized HRV norms specific to canine welfare science. This is an agenda-setting opinion/perspective paper rather than an empirical comparison, so it is broadly consistent with — but does not demonstrate — the general validity of HRV as a stress/self-regulatory index. Direct translation to human cognitive research requires caution.

Wearable continuous cognitive monitoring: The integration of ear-based (Santos et al. 2025) and PPG-reconstructed (Li et al. 2025) wearable HRV pipelines with cross-modal cognitive state classifiers (Pradeep et al. 2026; Sasi et al. 2026) defines a practical pathway from laboratory HRV-cognition findings to naturalistic continuous monitoring. Applications include operator fatigue tracking, clinical wearables for mood and attention management, and cognitive-load-aware adaptive computing environments (Kosch 2020).

7. RESEARCHERS AND INSTITUTES DIRECTORY

8. SEED-CORPUS COVERAGE NOTE

The synthesis request specifies a user-curated Zotero seed bibliography of 148 references spanning five clusters: Biofeedback, Modeling, CMW (Cognitive/Mindfulness/Workload), HeartRateLabOSS, and Mindfulness. Against this stated corpus, 15 knowledge cards were available for synthesis, several of which were explicitly flagged as supplemented to meet a minimum count with marginal HRV-cognition relevance. Coverage by cluster:

Four cards concern wearable/signal-engineering topics peripheral to HRV-cognition (Li et al. 2025, Santos et al. 2025, Baig et al. 2025, Vazifeh et al. 2025); one is animal welfare (Cobb et al. 2025); one is a VEN computational model with indirect CAN relevance (Keskin 2026). Together, these provide useful signal-processing coverage and CAN mechanistic detail but do not compensate for the absence of the theoretical primary sources.

Critical foundational references not present in any card: The 1996 Task Force paper (Malik et al., Circulation); Thayer & Lane 2000 and 2009 NIM papers; Porges 1995/2001/2007 PVT papers; Berntson, Cacioppo & Quigley 1991 autonomic space model; Kleiger et al. 1987 SDNN-mortality paper; Akselrod et al. 1981 spectral analysis; Pan-Tompkins 1985 QRS detector; Lehrer/Gevirtz biofeedback clinical trials; HRV-attention meta-analyses.

Practical implication: Sections 1 (History), 2 (Biological Basis), 3 (Theories), and 4 (Paradigms) of this knowledge base are primarily synthesized from established field knowledge rather than from available cards. The content is accurate to the field's consensus literature but readers requiring citable primary sources for those sections should retrieve the references listed in the Chronological Milestone List from bibliographic databases — they are high-citation, widely indexed papers. Sections 5 and 6 draw substantially from the available cards; coverage is most complete for the computational/ML dimension and most incomplete for the foundational theory-testing and biofeedback-efficacy literature.

9. OPEN QUESTIONS AND GAPS

9.1 Theoretical

• The causal direction and mediating mechanism of the resting-HRV–executive-function correlation remain underspecified in the NIM; shared prefrontal substrate creates potential circularity and does not distinguish vagal tone as cause vs. co-effect of executive capacity
• PVT's phylogenetic myelination claims require resolution with comparative neuroanatomical evidence; a revised account of dorsal-vs-ventral vagal functional differentiation is needed that either vindicates or reformulates the theory
• The interoceptive predictive-processing framework is theoretically productive but lacks systematic empirical adjudication against NIM and PVT predictions; bridging these frameworks remains an open theoretical problem
• The autonomic space model correctly predicts that LF/HF is insufficient, but replacement metrics for symmetric dual-axis characterization in short-term recordings are not yet standardized

9.2 Methodological

• Respiratory rate confounding of LF/HF is well-documented but solutions (continuous respiratory monitoring, HF-band respiratory normalization) are not uniformly adopted; methodological heterogeneity across studies persists
• ECG-to-EEG cross-modal surrogacy (Pradeep et al. 2026; Sasi et al. 2026) is validated only within single paradigms on limited datasets; generalization across cognitive task types, populations, and recording conditions is unknown
• Consumer HRV biofeedback devices optimize proxies with formalized misspecification (Bose 2026); rigorous RCTs with pre-specified clinical/psychological endpoints (not coherence scores) are lacking
• Nonlinear metrics (entropy, DFA) require longer recording segments than current cognitive paradigms typically provide; their validity in standard 5-minute laboratory windows requires formal assessment

9.3 Computational

• Bayesian state-space autonomic outflow models (Rosas et al. 2023) recapitulate the linear properties of conventional spectral estimators and show advantages mainly on dynamical-complexity metrics; whether their latent sympathetic/parasympathetic outflow estimates improve cognitive- or emotional-state decoding over standard HRV features has not yet been tested.

• Real-time HRV-based cognitive monitoring faces a latency-accuracy tradeoff: windowed HRV computation introduces lag incompatible with sub-second adaptive response; instantaneous point-process solutions [Rosas et al. 2023 context] are not yet integrated into practical adaptive-interface pipelines (Kosch 2020)

• Wearable HRV reconstruction pipelines (Li et al. 2025; Santos et al. 2025) lack independent clinical-grade validation of downstream HRV metric accuracy; benchmarks comparing reconstructed-ECG HRV to reference lead-II HRV across all standard metrics are absent

• Manifold learning for personalized ECG clustering (Vazifeh et al. 2025) has not been applied to cognitive or emotional state classification; transfer of the methodology from arrhythmia detection to autonomic-state inference is unvalidated

9.4 Application

• Long-term longitudinal RCTs measuring HRV as a primary pre-registered endpoint in mindfulness intervention studies are rare; most evidence is cross-sectional (expert vs. novice comparisons) or uses HRV as a secondary/indirect measure (Kang et al. 2020; Calvetti et al. 2021)

• The biofeedback-to-trait transfer problem (Bose 2026) — whether in-session coherence or RSA gains produce durable resting-HRV or wellbeing change — has no agreed measurement standard or minimum follow-up requirement

• The vision of HRV as a continuous cognitive workload monitor for adaptive interfaces (Kosch 2020 — a proposal based on EEG and eye-tracking rather than HRV) remains largely aspirational; no HRV-specific workload-monitoring validation appears in the available cards, and ecological validity in complex real-world tasks (driving, air traffic control, surgical performance) is assumed but not systematically established

• Canine HRV welfare research (Cobb et al. 2025) and the VEN fast-lane model (Keskin 2026) each open cross-species and cellular-level avenues for understanding autonomic-cognitive integration that remain methodologically disconnected from the core human HRV-cognition literature; integrative frameworks are lacking

• Nonlinear HRV metrics show greater sensitivity to depression-related dysregulation than linear metrics (Čukić et al. 2021), but clinical adoption lags because normative reference ranges, recording-length requirements, and software implementations for entropy and DFA are not included in current clinical reporting guidelines

Synthesis compiled from 15 available knowledge cards. Sections 1–4 draw primarily on established field consensus; sections 5–6 draw substantially from cards. See §8 for a detailed gap map between the 148-reference seed corpus and the available card set.

Synthesized Research Hypotheses

Synthesized Research Proposal: Final Hypotheses

Framing the Synthesis

The three perspectives converge on a single empirical flashpoint: the cross-modal HRV-to-EEG cognitive load frameworks (Pradeep et al. 2026, Sasi et al. 2026) are the field's most immediately testable computational claim and the most methodologically vulnerable. The contrarian correctly identifies that respiratory confounding may explain the entire correspondence; the innovator correctly identifies that phase information — not amplitude — may be the genuine cognitive signal; the pragmatist correctly supplies the pipeline. A decisive synthesis runs these as sequential hypotheses on the same dataset, in the same session of compute, forcing the perspectives into direct confrontation rather than parallel publication.

There are also genuine, unresolved disagreements that this synthesis does not flatten. They are noted explicitly at the end.

Final Hypothesis 1: Respiratory Rate Is the Primary Driver of HRV-EEG Cross-Modal Correspondence, and Controlling for It Eliminates Most of the Claimed Cognitive Load Signal

Rationale: The contrarian's strongest contribution is identifying that both ECG/HRV and EEG co-vary with respiratory pattern during n-back tasks through channels that are not cognitive-load encoding: respiratory rate changes alter HF-HRV directly (RSA shifts in and out of the HF band) and alter EEG indirectly (thoracic motion, CO2/O2-mediated neural excitability changes). The Pradeep et al. 2026 and Sasi et al. 2026 studies do not include respiratory rate as a covariate in their cross-modal models. The pragmatist confirms this is computable: ECG-derived respiratory rate (via RSA peak frequency, implementable in neurokit2) can be estimated without a separate respiratory channel and added to the cross-modal XGBoost pipeline in one additional feature column. The innovator's phase hypothesis is connected here: if phase carries genuine cognitive information orthogonal to amplitude, it should survive respiratory deconfounding where amplitude does not.

This hypothesis is empirically necessary before any positive claim about HRV's cognitive-monitoring utility can be made. It does not assume the contrarian is right — it tests the assumption.

Test design: On the OpenNeuro multimodal EEG+ECG dataset, estimate instantaneous respiratory rate from ECG (RSA peak frequency in a sliding window) for each subject. Train three cross-modal models predicting EEG cognitive load labels: (a) HRV features only [baseline, replicating published result]; (b) respiratory rate only; (c) HRV features + respiratory rate [full model]. Compare cross-validated AUC. The key quantity is: what fraction of HRV's predictive variance is shared with respiratory rate?

Measurable prediction: The reduction in HRV features' unique AUC after respiratory rate is controlled (i.e., AUC of model (c) minus AUC of model (b), compared to AUC of model (a) minus chance) will reveal whether HRV is adding cognitive-load information beyond respiratory modulation. The innovator's phase metric — cross-spectral phase angle between RR and ECG-derived respiratory estimate — is added as a candidate feature in model (c) to test whether phase survives where amplitude does not.

Failure conditions (two opposing falsifications): If respiratory rate alone achieves AUC within 0.05 of the full HRV-feature model, the contrarian's hypothesis is substantially confirmed and the cross-modal claims require re-evaluation. If HRV features retain > 0.10 unique AUC above the respiratory-rate-only baseline, the contrarian's respiratory-confound hypothesis is rejected and the cross-modal cognitive-monitoring program is validated. A partial result — respiratory rate explains 50–70% of HRV's predictive variance — is the most informative outcome: it means respiratory monitoring is the primary wearable cognitive monitor, with HRV providing modest additive signal.

Compute requirements: neurokit2, scipy.signal.csd, pycatch22, xgboost, sklearn; OpenNeuro dataset; CPU-only; approximately 25–35 minutes.

Final Hypothesis 2: HRV Features Predict Behavioral N-Back Accuracy Less Well Than They Predict EEG-Derived Load Labels — Quantifying the Double-Proxy Inflation

Rationale: The contrarian's meta-critique identifies a structural flaw in how the cross-modal frameworks define ground truth: EEG-derived cognitive load markers are themselves ML classifications on spectral EEG features, not behavioral performance. The field therefore validates HRV against a noisy proxy rather than the psychological construct of interest. The pragmatist's feature comparison pipeline can be re-run with behavioral n-back accuracy as the target rather than EEG labels — a trivial change in the outcome variable but a critical test. The innovator agrees that autonomic bandwidth metrics should predict behavioral performance; this provides the forum to test that claim honestly.

The synthesis asks: is there a gap between HRV's apparent cross-modal accuracy and its actual ability to predict what subjects do? If the gap is large, published cross-modal AUC values are inflated by proxy-chain noise and should be systematically deflated in interpretation. If the gap is small, the EEG labels are good behavioral proxies and the cross-modal program is methodologically sound.

A secondary test addresses the contrarian's fitness-confound critique at the available scale: resting heart rate (extractable from the pre-task ECG baseline in the OpenNeuro recording) is a fitness proxy. Including it as a covariate in the behavioral-prediction model tests whether the HRV-performance association survives a minimal fitness correction, without requiring VO2max measurement.

Test design: On the OpenNeuro dataset, train parallel XGBoost models with the same HRV + HRnV + Catch22 feature set predicting: (a) EEG-derived cognitive load labels [published comparison]; (b) actual n-back accuracy (proportion correct per subject per condition); (c) n-back accuracy residualized for resting heart rate. Compare cross-validated AUC/R² across all three. Report the gap between (a) and (b) as the double-proxy inflation estimate.

Measurable prediction: If EEG labels are meaningful proxies for behavioral performance, models (a) and (b) will produce comparable cross-validated performance (< 0.05 AUC difference). If the gap exceeds 0.10, the EEG-label-based validation literature is systematically inflated. Model (c) tests whether resting heart rate accounts for a fitness-mediated component of the HRV-behavior correlation; a drop of > 30% in R² when resting heart rate is controlled raises the fitness-confound concern to a threshold requiring full VO2max follow-up.

Failure conditions: If model (b) substantially outperforms model (a) for any HRV feature set, the EEG labels are noisier than behavioral accuracy — an unusual finding that suggests EEG feature engineering is degrading rather than clarifying the cognitive signal. If models (a), (b), and (c) produce comparable results with minimal fitness-correction impact, the pragmatist's feature pipeline is validated for behavioral cognitive monitoring and the contrarian's fitness concern is not supported at this dataset scale.

Compute requirements: Same pipeline as Hypothesis 1; adds one alternate outcome variable column; total compute across H1 + H2 approximately 45–55 minutes CPU.

Final Hypothesis 3: HRnV Multi-Scale Metrics Add Cognitive Load Information Beyond RMSSD, Particularly for the Behavioral-Performance Target That Frequency-Domain Metrics Cannot Reach

Rationale: The pragmatist's HRnV domain-transfer hypothesis is preserved here but reoriented by H1 and H2. After the respiratory deconfounding step in H1, HF-HRV is expected to lose predictive power (if the contrarian is right). RMSSD may survive better (it reflects beat-to-beat differences, less frequency-band specific). HRnV-2 and HRnV-3 metrics probe intermediate timescales via subsampling rather than frequency decomposition — making them less susceptible to respiratory rate confounding than HF-HRV, while capturing more temporal structure than RMSSD alone. The hypothesis is that HRnV metrics will show smaller percentage drop in predictive AUC after respiratory deconfounding than HF-HRV, and will show comparable or better performance than RMSSD for the behavioral n-back accuracy target from H2.

This directly arbitrates between the innovator's "unified bandwidth" claim and the pragmatist's "validated incremental improvement" approach. The innovator proposes permutation entropy as the respiratory-rate-robust measure; this hypothesis tests whether the simpler HRnV multi-scale approach achieves the same immunity with less computational overhead. It also tests whether the HRnV advantage demonstrated in clinical triage (Niu et al. 2021) transfers to cognitive monitoring when the ground truth is behavioral rather than EEG-derived.

Test design: Within the same pipeline as H1 and H2, compare percentage AUC reduction from H1's respiratory deconfounding step across three metric classes: HF-HRV (expected to drop most), RMSSD (expected to drop less), and HRnV-2/HRnV-3 (predicted to drop least). Then evaluate all metric classes against the behavioral n-back accuracy target from H2.

Measurable prediction: HRnV-2 RMSSD retains ≥ 70% of its pre-deconfounding predictive AUC after respiratory rate is controlled, compared to HF-HRV which retains < 50%. Additionally, the HRnV-augmented feature set achieves ≥ 0.05 higher AUC than RMSSD alone for predicting behavioral n-back accuracy (the H2 target), replicating the triage advantage in a cognitive-monitoring context.

Failure conditions: If HRnV metrics drop comparably to HF-HRV after respiratory deconfounding (all metrics retain < 50% of pre-deconfounding AUC), the respiratory confound is not frequency-specific and the contrarian's strongest version is supported. If HRnV and RMSSD perform equivalently for behavioral accuracy (< 0.02 AUC difference), the multi-scale extension offers no cognitive-monitoring advantage and the pragmatist's H3 original hypothesis is rejected.

Compute requirements: Same pipeline; HRnV extraction adds < 5 minutes to H1+H2 runtime. Total for all three hypotheses: approximately 50–60 minutes CPU.

Unresolved Disagreements — Not Flattened

1. Fitness confound in the NIM — genuinely undecidable on OpenNeuro data

The contrarian's most consequential hypothesis — that the resting HRV-cognition correlation is explained by cardiovascular fitness, not neurovisceral integration — cannot be settled on the OpenNeuro dataset or any dataset lacking VO2max measurement. The pragmatist's resting-heart-rate proxy (H2) provides a weak test but not a definitive one. The innovator implicitly accepts the NIM framework; the contrarian rejects it. This disagreement requires a purpose-built dataset with fitness assessment alongside cognitive performance and HRV, and it is the field's most important uncontrolled confound. It remains open.

2. Biofeedback: breathing mechanism vs. structural Goodhart trap

The contrarian argues that the active ingredient in HRV biofeedback is breathing rate, not feedback — a claim testable only by a metronome-only vs. HRV-feedback RCT that is beyond single-session compute. The innovator argues that even the breathing effect will exhibit structural transfer failure because the ANS optimizes context-specifically. These predictions differ: the contrarian predicts metronome-only = HRV feedback; the innovator predicts both conditions show in-session gains but neither transfers. Both predictions are compatible with Bose 2026's three failure modes but for different reasons. This disagreement is unresolvable computationally and requires a multi-session RCT with follow-up assessment at minimum 4 weeks post-intervention.

3. Whether RSA phase carries genuine cognitive load information

The innovator proposes that cross-spectral phase between RR and respiration is an untapped cognitive signal. H1 tests whether this phase survives respiratory deconfounding. But the contrarian would note that "RSA phase" computed from ECG-derived respiratory rate is partly circular (the respiratory signal itself is derived from the cardiac signal via RSA), and that a genuine test requires independent high-quality respiratory measurement. The phase hypothesis is the most innovative contribution of this synthesis and the most methodologically fragile. H1 tests a necessary condition (phase adds unique information) but not a sufficient one (the respiratory circularity problem may inflate the apparent effect). This disagreement can be partially addressed by comparing ECG-derived respiratory rate against a concurrent independent respiratory signal if such data becomes available.

Final Hypothesis Summary