Applications of Heart Rate Variability Metrics in Wearable Sensor Technologies: A Comprehensive Review

Abstract

Heart rate variability (HRV) has emerged as a key biomarker for assessing autonomic nervous system activity, stress, fatigue, and emotional states. With the rapid development of wearable sensor technologies, HRV analysis has expanded from clinical environments to real-world, continuous monitoring. This review summarizes current applications of HRV metrics in wearable devices, including fitness tracking, mental stress assessment, sleep quality evaluation, and early detection of physiological or psychological disorders. Recent advances in photoplethysmography (PPG)-based HRV estimation have enabled noninvasive and user-friendly measurement, though challenges remain in accuracy under motion and variable environmental conditions. We also discuss methodological considerations, such as artifact correction, data segmentation, and the integration of HRV with other biosignals for multimodal analysis. Emerging research suggests that combining HRV with metrics such as respiration rate, skin conductance, and accelerometry can enhance robustness and interpretability in dynamic settings. Finally, future directions are proposed toward personalized health analytics, emotion-aware computing, and real-time adaptive feedback systems. This review highlights the growing potential of wearable HRV analysis as a foundation for preventive healthcare and human–machine symbiosis.

1. Introduction

1.1. Background: HRV as a Physiological Biomarker

Heart rate variability (HRV) is a widely used physiological biomarker reflecting autonomic nervous system regulation. Foundational studies have established HRV as a measure of sympathovagal balance and extended its analytical framework beyond conventional frequency-domain indices toward long-term and integrative approaches [1,2,3,4,5]. Furthermore, the neurovisceral integration framework highlights the bidirectional relationship between cardiac dynamics and brain function, linking HRV to cognitive and emotional regulation [6,7]. Subsequent research has emphasized the effects of aging and the importance of nonlinear dynamics in characterizing complex heart rate fluctuations [8,9,10]. In addition, sex differences in autonomic regulation and HRV metrics have been increasingly recognized, necessitating careful consideration of biological variability [11,12,13]. Based on these theoretical developments, HRV has been applied across diverse domains, including early screening of cardiovascular and neurological diseases, postoperative prognosis, and the estimation of fatigue, drowsiness, stress, and mental states. One of the fundamental strengths of HRV analysis lies in its derivation from electrocardiographic (ECG) signals, which provide high temporal precision and robustness against noise. Moreover, accumulating evidence supporting the concept of the brain–heart axis indicates that beat-to-beat heart rate dynamics reflect not only cardiac autonomic regulation but also central nervous system activity. This neurocardiac coupling has further reinforced the role of HRV as a noninvasive window into human physiological and psychological states.

1.2. Conventional and Emerging HRV Metrics

Traditionally, HRV has been quantified using time-domain, frequency-domain, and nonlinear metrics [14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36]. Among these, a major line of research focuses on the analysis of complexity and fractal properties of heart rate dynamics, particularly through detrended fluctuation analysis (DFA). Multiscale extensions of DFA have been proposed to improve computational efficiency and to characterize scale-dependent physiological complexity under various conditions, including circadian variations [14,15,16,20,21], while complementary studies have examined the theoretical validity of DFA and its relationship with conventional spectral analysis, especially in short data segments [17,18,19]. More advanced approaches, such as asymmetric multiscale analysis, further capture differences in fractal behavior between heart rate acceleration and deceleration phases [35].

In parallel, another important group of studies has focused on deceleration capacity (DC) and acceleration capacity (AC) as physiologically interpretable markers of vagal and sympathetic modulation with strong clinical relevance. Notably, DC has been demonstrated as a powerful predictor of mortality following myocardial infarction [29], and subsequent studies have extended its application to various pathological and environmental conditions, including hypertension, vasovagal syncope, cardiomyopathy, and hypoxic exposure [22,23,24,27,28]. In addition, DC/AC metrics have been investigated as prognostic indicators in patients with heart failure and cancer [25,26].

More recent work has introduced the concept of heart rate fragmentation, which captures the breakdown of continuous heart rate dynamics into rapid, disorganized fluctuations associated with aging and cardiovascular dysfunction [31,33]. This framework suggests that such fragmentation may reflect not only autonomic dysregulation but also intrinsic sinoatrial node impairment. Furthermore, its impact on conventional HRV metrics and potential to confound their interpretation have been critically evaluated [32].

Finally, efforts to improve analytical robustness and extend HRV assessment to real-world conditions have led to the development of new mathematical frameworks and methodological considerations. These include critical discussions on the limitations and potential pitfalls of HRV as a surrogate for autonomic function [30], validation of HRV indices derived from ultra-short-term recordings relevant to wearable sensing [36], and the introduction of topological data analysis approaches, such as persistent homology, to capture higher-order structural features of heart rate dynamics beyond conventional time-series methods [34].

Collectively, these studies move beyond simple statistical characterization of variability and instead aim to more accurately interpret the complex dynamical interactions between cardiac function and autonomic regulation using advanced mathematical frameworks.

Building upon these advances in mathematical modeling and physiological interpretation, recent years have witnessed a rapid expansion of novel HRV-related indices that further exploit developments in signal processing and computational methodologies. In parallel, sensor and hardware technologies have undergone remarkable progress over the past decade, driven by improvements in semiconductor performance, miniaturization, and power efficiency. These technological innovations have enabled long-term and ambulatory acquisition of RR intervals (RRIs) using compact and wearable devices, including Holter ECG recorders, thereby greatly enhancing the feasibility of continuous HRV monitoring in real-world and non-clinical environments. Together, these methodological and technological developments have facilitated a shift from controlled, short-term assessments toward continuous, high-resolution monitoring of autonomic dynamics in daily life.

1.3. Challenges in Wearable-Based HRV Measurement

Despite these advances, fundamental questions remain unresolved: How reliable are the growing number of HRV metrics, and what can realistically be achieved using these indicators with current wearable technologies? These questions have become increasingly important with the widespread adoption of wearable sensors, many of which estimate cardiac dynamics not from ECG but from peripheral pulse signals.

1.4. Distinction Between HRV and PRV

The recent proliferation of wearable devices has enabled routine measurement of pulse-to-pulse intervals (PPI) derived from photoplethysmography (PPG) [37,38,39,40,41,42,43,44,45]. These studies represent a translational phase in which the previously established mathematical and physiological frameworks of HRV are extended toward practical implementation using wearable and contactless sensing technologies. A substantial body of work has focused on validating the accuracy and reliability of consumer-grade devices, including ring-, wrist-, and chest-worn systems, by comparison with clinical-grade measurements, demonstrating their feasibility for long-term monitoring in both general and elderly populations [37,38,39,40]. In addition to PPG-based approaches, novel sensing modalities have been explored, such as the use of inertial measurement units (IMUs) embedded in smartwatches to detect physiological conditions like sleep apnea through subtle motion patterns, either independently of or in combination with cardiac signals [41,42]. Furthermore, non-contact measurement techniques, including radar-based sensing, have been introduced to assess physiological and psychological states, enabling low-burden monitoring of stress and mental health [43]. From a clinical perspective, these approaches have also contributed to establishing HRV and pulse rate variability (PRV) as meaningful biomarkers, with evidence linking vagal activity to disease prognosis, such as in cancer outcomes [44], and proposing PRV as a distinct physiological marker rather than merely a surrogate of HRV [45]. Collectively, these studies shift the focus from theoretical characterization toward real-world applicability, aiming to integrate advanced physiological insights into scalable, unobtrusive monitoring systems for healthcare and daily life applications.

Although PPI-based indices are often used as substitutes for RRI-based HRV, pulse rate variability (PRV) and HRV are inherently distinct phenomena and should not be used interchangeably without careful consideration. This issue was first clearly articulated by Constant et al. in 1999 [46]. By examining short-term variability of finger pulse waves and ECG signals in children with fixed ventricular pacing (80 beats/min) and comparing them with healthy controls in sinus rhythm, they demonstrated that respiratory-related fluctuations in pulse signals are mechanically influenced by changes in cardiac output and arterial wall pressure through pulse wave propagation velocity. Their findings indicated that, particularly in upright healthy individuals and patients with low HRV, respiratory PRV does not accurately reflect respiratory HR variability. Consequently, they concluded that heart rate modulation should be studied using ECG signals rather than distal pulse waveforms.

1.5. Supporting Evidence from Subsequent Studies

Subsequent studies have reinforced this distinction. Yuda et al. (2020) [45] provided direct evidence that HRV and PRV are not equivalent by simultaneously measuring ECG and pulse wave signals in elderly patients with fixed ventricular pacing (VOO mode). Similarly, Hejjel and Bére (2021) [47] reported that the agreement and interchangeability between HRV and PRV metrics are not clearly established. More recently, Kantrowitz et al. (2025) [48], using large-scale and diverse clinical datasets, demonstrated through big data analysis that pulse rate variability is not the same as heart rate variability, further underscoring the physiological and methodological differences between these measures.

1.6. Practical Use Cases of PRV in Wearables

Nevertheless, important nuances remain. Constant et al. (1999) [46] acknowledged that distal pulse signals may serve as an acceptable alternative when ECG recording is not feasible. Yoshida et al. (2017) [49] further suggested that under conditions of minimal body motion and involuntary respiration—such as during sleep—PRV indices can approximate HRV metrics with reasonable validity. Given that sleep monitoring is one of the most prominent applications of wearable devices, PRV-based analysis retains practical value, and continued investigation into the conditions under which PRV can reliably substitute for HRV remains scientifically and clinically meaningful.

1.7. Toward Multimodal and Advanced HRV Analysis

In addition to these measurement considerations, the landscape of HRV analysis itself has evolved substantially. Novel indices such as heart rate fragmentation (HRF), introduced in 2017 [31,32,33], and markers such as Hsi—reflecting power concentration of respiratory–cardiac coupling during non-rapid eye movement (NREM) sleep—proposed in 2020 [50], exemplify a shift toward more physiologically nuanced interpretations of cardiac dynamics. These developments are closely linked to the rise of multimodal analysis. Earlier studies often focused on extracting information from one-dimensional RRI time series alone; in contrast, contemporary research increasingly integrates HRV with accelerometry, skin temperature, respiration, speech activity, and other biosignals. Modern wearable sensors routinely capture multiple synchronized physiological and behavioral signals, making multimodal state estimation not only feasible but essential for robust interpretation in real-world settings.

1.8. Aim of This Review

Against this backdrop, the present review surveys the expanding body of literature on HRV metrics and their application in wearable sensor technologies. We examine both established and emerging HRV indices, critically discuss the distinction and relationship between HRV and PRV, and highlight methodological considerations inherent to wearable-based measurements. By synthesizing findings across diverse studies, we aim to clarify the current capabilities and limitations of HRV analysis and to outline future directions toward reliable, multimodal, and personalized physiological monitoring.

2. Methodology and Concepts of HRV Analysis in Wearable Technology

The assessment of autonomic nervous system (ANS) activity through wearable devices has evolved from simple heart rate monitoring to complex signal processing. This section outlines the conventional metrics, their applications in health and medicine, and the emergence of fragmentation-based indices.

2.1. Conventional HRV Metrics: Time, Frequency, and Non-Linear Domains

Conventional HRV analysis is generally categorized into three domains, each providing different insights into cardiac autonomic regulation (Figure 1 and Figure 2,

Table 1. Heart Rate Variability (HRV) Metrics.

Category/Metric	Definition
1. Time-domain measures
a. Moment statistics
Mean NN	Mean of all normal-to-normal (NN) intervals over 24 h (ms)
SDNN	Standard deviation of all NN intervals over 24 h (ms)
SDANN	Standard deviation of the average NN intervals calculated over 5-min segments across 24 h (ms)
RMSSD	Square root of the mean squared differences in successive NN intervals (ms)
SDNN index	Mean of the standard deviations of NN intervals for all 5-min segments over 24 h (ms)
SDSD	Standard deviation of successive differences between adjacent NN intervals (ms)
NN50 count	Number of pairs of successive NN intervals differing by more than 50 ms
b. Geometric measures
Triangular index	Total number of NN intervals divided by the height of the histogram (maximum frequency) over 24 h
TINN	Baseline width of the triangle approximating the NN interval histogram over 24 h (ms)
2. Frequency-domain measures
a. Short-term (5-min) HRV measures
Total Power (TP)	Variance of NN intervals over 5 min (ms2)
VLF	Power in the very-low-frequency band (≤0.04 Hz) (ms2)
LF	Power in the low-frequency band (0.04–0.15 Hz) (ms2)
LF amp	LF oscillation expressed as average amplitude: √(2 × LF) (ms)
LF norm	LF power in normalized units: LF/(TP − VLF) × 100 (%)
LFccv	LF power expressed as component coefficient of variance: 100 × √LF/mean NN (%)
HF	Power in the high-frequency band (0.15–0.40 Hz) (ms2)
HF amp	HF oscillation expressed as average amplitude: √(2 × HF) (ms)
HF norm	HF power in normalized units: HF/(TP − VLF) × 100 (%)
HFccv	HF power expressed as component coefficient of variance: 100 × √HF/mean NN (%)
LF/HF	Ratio of LF to HF power: LF (ms2)/HF (ms2)
b. Long-term (24-h) HRV measures
Total Power	Variance of NN intervals over 24 h (ms2)
ULF	Power in the ultra-low-frequency band (≤0.003 Hz) (ms2)
VLF	Power in the very-low-frequency band (0.003–0.04 Hz) (ms2)
LF	Power in the low-frequency band (0.04–0.15 Hz) (ms2)
HF	Power in the high-frequency band (0.15–0.40 Hz) (ms2)
Power-law β	Slope of the regression line of the log–log power spectrum below 0.04 Hz (spectral exponent)
3. Nonlinear dynamics measures
DFA α1, α2	Short-term and long-term scaling exponents derived from detrended fluctuation analysis
Non-Gaussianity λ	A measure of the deviation from a Gaussian distribution across multiple scales (Kiyono et al.).
MF-DFA	Multifractal DFA, describing the singularity spectrum and heterogeneity of the signal.
ApEn	Approximate entropy, a measure of signal complexity
SampEn	Sample Entropy; an improved version of ApEn with reduced bias and better consistency.
MSE	Multiscale Entropy; evaluates complexity by calculating entropy over multiple time scales.
Poincaré plot	Morphological classification of the scatter plot of consecutive NN intervals
Symbolic Dynamics	Analysis of bit-sequences (0V, 1V, 2LV, 2UV) representing sympathetic and parasympathetic modulation.
Correlation dimension	Fractal dimension estimated using correlation dimension analysis
Capacity dimension	Fractal dimension estimated using box-counting method
DC (Deceleration Capacity)	Quantifies the capacity of heart rate deceleration using PRSA; a powerful predictor of mortality.
AC (Acceleration Capacity)	Quantifies the capacity of heart rate acceleration, reflecting sympathetic activation.

The notation NN intervals stands for normal-to-normal intervals and is an international standard. The definition of normalized units using TP—VLF follows the Task Force (1996). This table was created with reference to Hayano J. (http://www.med.nagoya-cu.ac.jp/igak.dir/nmj-pdf/50-2/P093-100hayano.pdf, accessed 28 March 2026). Deceleration Capacity (DC) is categorized as a nonlinear measure derived from Phase-Rectified Signal Averaging (PRSA). Unlike traditional linear frequency domain analyses, DC selectively quantifies the heart rate’s deceleration dynamics, reflecting asymmetric autonomic modulation (primarily vagal activity) and providing superior prognostic value for mortality risk.

) [1,2,3,4,5]. Time-Domain Indices: These are the simplest to compute and include metrics such as SDNN (Standard Deviation of NN intervals), reflecting overall variability, and RMSSD (Root Mean Square of Successive Differences), which is widely used in wearable applications to estimate parasympathetic (vagal) activity. Frequency-Domain Indices: By applying Power Spectral Density (PSD) analysis, HRV is divided into frequency bands: High Frequency (HF: 0.15–0.40 Hz): Reflects parasympathetic tone. Low Frequency (LF: 0.04–0.15 Hz): Reflects a mix of sympathetic and parasympathetic activity. LF/HF Ratio: Historically used as an index of sympathovagal balance, though its interpretation remains a subject of debate in recent literature. Non-linear Indices: These metrics quantify the complexity and unpredictability of the heartbeat, such as Poincaré plots (SD1, SD2), Sample Entropy (SampEn), and Detrended Fluctuation Analysis (DFA). They are particularly useful for capturing the “fractal” nature of heart rate dynamics that linear methods often overlook.

In frequency-domain HRV analysis, the parameters F and P represent the location and magnitude of each spectral component, respectively. F (Frequency), expressed in hertz (Hz), denotes how rapidly a given oscillatory component fluctuates. For example, a peak around 0.12 Hz corresponds to a slow rhythm occurring approximately once every 10 s, typically associated with baroreflex-related Mayer waves within the LF band. In contrast, a peak near 0.27 Hz reflects faster oscillations driven by respiratory sinus arrhythmia, representing parasympathetic (vagal) modulation within the HF band. P (Power), expressed in ms², quantifies the strength or amplitude of each frequency component and corresponds to the “energy” or area under the curve of the R–R interval fluctuations. Larger power values indicate stronger autonomic modulation, whereas stress, fatigue, or aging generally reduce overall spectral power.

Using an autoregressive (AR) model, these F and p values allow complex heart-rate fluctuations to be decomposed into sympathetic-related (LF) and parasympathetic-related (HF) components with high resolution. In Figure 2A, the spectrum shows an HF peak of approximately 443 ms² centered near 0.27 Hz, indicating stable extraction of vagal respiratory modulation. Meanwhile, the LF component exhibits a dominant peak of approximately 708 ms² around 0.12 Hz, suggesting a moderate level of sympathetic engagement or physiological arousal.

Figure 2A: HRV is a simple, non-invasive method that quantifies beat-to-beat fluctuations in heart rate and serves as an indicator of cardiac autonomic nervous system activity. HRV decreases with aging and is further reduced when autonomic dysfunction occurs, shifting the autonomic balance toward sympathetic dominance. Reduced HRV is associated with increased cardiovascular risk, including heart failure, coronary artery disease, and mortality after acute myocardial infarction.

HRV analysis consists of evaluating overall variability and decomposing the periodic fluctuations of the R–R intervals (RRIs) into frequency components using power spectral analysis. The HRV spectrum is typically divided into three frequency bands:

VLF (Very Low Frequency, 0–0.05 Hz): reflects vascular tone regulation, renin–angiotensin system activity, and thermoregulation.

LF (Low Frequency, 0.05–0.20 Hz): associated with baroreflex modulation and influenced by both sympathetic and parasympathetic activity.

HF (High Frequency, 0.20–0.35 Hz): corresponds to respiratory sinus arrhythmia and reflects parasympathetic (vagal) activity.

Because parasympathetic responses are rapid and sympathetic responses are slower, these two branches of the autonomic nervous system occupy different frequency ranges. HF power represents vagal modulation, whereas LF power reflects a mixture of sympathetic and parasympathetic influences. The LF/HF ratio is commonly used as an index of sympathovagal balance.

The figure illustrates the analysis procedure. In (a), the intervals between successive heartbeats (T1–T6) vary slightly. Plotting these intervals over time yields the RRI time series shown in (b), where the beat-to-beat fluctuations appear as a jagged waveform. Applying Fast Fourier Transform (FFT) to this signal decomposes the variability into its constituent frequencies, producing a spectrum with a peak around ~0.1 Hz (LF) and another around ~0.25 Hz (HF). The power within these frequency bands is used as an estimate of autonomic nervous system function: HF power as an index of parasympathetic activity, and LF/HF as an index of sympathetic modulation.

Figure 2B: Autonomic Modulation Reflected in HF and LF Components of HRV; Schematic illustration showing how posture, tilt testing, and mental stress influence the frequency components of heart rate variability (HRV). The HF and LF components of HRV serve as useful indicators of autonomic nervous system activity because they change predictably with physiological conditions. In the supine position, parasympathetic (vagal) activity predominates, resulting in a larger HF component. When standing, sympathetic activation increases and vagal tone decreases, producing a relative rise in the LF component. This shift is also visible in the tachogram: the fine, rapid oscillations associated with respiratory sinus arrhythmia diminish upon standing, while slower fluctuations become more prominent. A mechanical tilt test, in which the subject is moved from supine to upright and then returned to supine, clearly demonstrates these autonomic transitions. The shift to upright posture suppresses HF power and enhances LF power, whereas returning to the supine position restores vagal dominance. Mental stress produces a similar autonomic pattern: HF power decreases and LF power increases, reflecting sympathetic activation and vagal withdrawal. HRV analysis therefore provides an objective method for assessing the current state of autonomic function, its dynamic changes, and the effects of therapeutic interventions. In addition, autonomic activity exhibits circadian variation, and disturbances in this daily rhythm—caused by chronic stress or disease—can be detected more sensitively through HRV than by single-time-point measurements. (This figure was created by the author with reference to the webpage, http://www.take-clinic.com/psm/hrv/hrv_autonomic2.htm (accessed 28 March 2026))

Supplementary note for Figure 2B; the units used in the power spectrum depend on whether the analysis is based on R–R intervals or instantaneous heart rate. When HRV is computed from R–R intervals, the spectral power is expressed in ms² (or ms²/Hz for power spectral density), which is the standard convention in physiological and clinical research. In contrast, when the analysis is performed on instantaneous heart rate (beats/min), the corresponding units become (beats/min)² or (beats/min)²/Hz. These represent the variance or spectral density of heart-rate fluctuations rather than interval-based variability. Because the numerical scale differs substantially between ms-based and bpm-based representations, care must be taken to ensure unit consistency when interpreting or comparing results. In most HRV studies—including the present example—the spectral components in Figure 1b are expressed in ms², reflecting variability in the R–R interval signal.

2.2. Applications in Clinical Medicine and Health Science

HRV indices derived from wearable devices have become powerful tools for both clinical screening and daily health management (

Table 2. Comparison Between HRV and PRV.

Aspect	HRV (Heart Rate Variability)	PRV (Pulse Rate Variability)
Primary signal	ECG-derived NN (RR) intervals	PPG-derived pulse-to-pulse intervals (PPI)
Physiological origin	Electrical depolarization of the heart	Mechanical pulse wave propagation
Relation to autonomic regulation	Direct reflection of cardiac autonomic modulation	Indirect; influenced by vascular and hemodynamic factors
Sensitivity to respiration	Reflects respiratory sinus arrhythmia (RSA) directly	Modulated by respiration via pulse wave velocity and stroke volume
Influence of vascular properties	Minimal	Significant (arterial stiffness, blood pressure, vasomotion)
Effect of posture and preload	Relatively small	Pronounced, especially in upright posture
Motion sensitivity	Moderate (depends on electrode quality)	High (motion artifacts strongly affect PPG)
Equivalence to HRV metrics	Reference standard	Not equivalent; metric-dependent and condition-dependent
Valid substitution for HRV	Yes (gold standard)	Limited; acceptable only under specific conditions
Evidence for non-equivalence	—	Demonstrated by Constant et al. (1999) [46], Yuda et al. (2020) [45], Hejjel et al., Kantrowitz et al. (2025) [48]
Recommended use	Clinical research, diagnosis, mechanistic studies	Wearable monitoring, screening, and trend analysis under constrained conditions

HRV and PRV represent related but distinct physiological phenomena and should not be used interchangeably without explicit validation of measurement conditions and target outcomes.

and

Table 3. Applicability of HRV/PRV Metrics in Wearable Sensing Contexts.

Measurement Context	ECG-Based HRV	PPG-Based PRV	Notes
Resting, supine	〇	△	PRV approximates HRV under minimal motion
Controlled laboratory tasks	〇	△	PRV validity depends on task design and posture
Sleep monitoring	〇	〇	PRV acceptable due to low motion and involuntary respiration
Ambulatory daily life	△	×	Motion artifacts strongly affect PRV
Exercise	△	×	Rapid hemodynamic changes degrade PRV accuracy
Long-term (24 h) analysis	〇	△	PRV unsuitable for ULF/VLF interpretation
Frequency-domain metrics	〇	△	LF and HF may be distorted by vascular effects
Nonlinear HRV metrics	〇	×	Fragmentation and entropy metrics are not interchangeable
Stress/fatigue estimation	〇	△	Multimodal support recommended for PRV
Clinical screening	〇	×	PRV lacks diagnostic equivalence
Consumer wearable use	△	〇	Usability prioritized over physiological precision

Legend: 〇 = Appropriate/reliable △ = Conditionally acceptable × = Not recommended.

). In clinical and prognostic applications, reduced HRV is a well-established predictor of mortality following myocardial infarction and has been widely used for early screening of diabetic neuropathy and cardiovascular diseases [51,52,53,54,55,56,57,58,59,60,61,62,63]. These findings highlight the strong association between autonomic dysfunction and disease progression, underscoring the clinical relevance of HRV as a noninvasive biomarker.

In addition to traditional clinical use, HRV-based approaches have been increasingly applied in sleep analysis. A substantial body of literature has demonstrated the utility of HRV for sleep stage estimation as well as for screening of sleep disorders such as sleep apnea syndrome (SAS) [64,65,66,67,68,69,70,71,72,73,74]. These methods enable continuous and unobtrusive monitoring of sleep quality in real-world environments, particularly when integrated with wearable sensor technologies.

Furthermore, in the fields of health science and psychophysiology, wearable devices allow real-time monitoring of mental and physiological states. HRV has been extensively studied as an indicator of stress, emotion, and mental health status [75,76,77,78,79,80,81,82,83,84,85]. Specifically, lower HRV indices—particularly RMSSD and HF components—are consistently associated with increased perceived stress and impaired emotional regulation. In addition, changes in frequency-domain metrics such as the LF/HF ratio and total power have been used to detect fatigue and drowsiness in drivers and workers, supporting the development of early warning systems for safety-critical environments.

Overall, these applications demonstrate that HRV serves as a versatile biomarker bridging clinical medicine, sleep science, and psychophysiology, with wearable technologies further expanding its practical utility in continuous and real-world monitoring.

2.3. New Frontiers: Heart Rate Fragmentation (HRF), Hayano Sleep Index (His) and Heart Rate Recovery (HRR)

• Heart Rate Fragmentation (HRF)

Recent studies have identified a phenomenon called Heart Rate Fragmentation (HRF), which refers to a breakdown in the smooth, neuroautonomic control of the sinoatrial node, leading to frequent, short-term fluctuations in heart rate acceleration [31,32,33]. PIP (Percentage of Inversion Points): As a core metric of HRF, PIP quantifies the frequency of “directional changes” in RR intervals. High PIP values indicate a fragmented rhythm, often seen in aging or diseased hearts, which can confound traditional HF power measurements.

PIP, a key index of HRF, calculates the percentage of inversion points where the direction of heart rate fluctuations switches from “acceleration to deceleration” or “deceleration to acceleration.” First, when the difference between consecutive RR intervals is defined as

$$∆{RR}_i={\mathrm{R}\mathrm{R}}_i-{\mathrm{R}\mathrm{R}}_{i-1}$$

the inversion point (IP) is defined as follows:

$${IP}_i=\left\{\begin{array}{c}1if\left(∆{RR}_i·∆{RR}_{i+1}\right)<0\\ 0\mathrm{otherwise}\end{array}\right.$$

$$\mathrm{P}\mathrm{I}\mathrm{P}={\displaystyle \frac{1}{N-2}}\sum _{I=2}^{n-1}{IP}_i\times 100$$

HRF (especially PIP) is attracting attention because it calls into question the conventional paradigm of “high frequency = vagal activity.” The conventional interpretation is

Smooth fluctuations due to respiratory arrhythmia (RSA) = High HF power

HRF interpretation: Jagged fluctuations due to cardiac “control breakdown” caused by aging or disease = High HF power (Fragmented).

When the HF component is high but DFA α₁ is low and PIP is high, this suggests that autonomic nervous system activity is not due to activity, but rather that control at the SA node (sinus node) is fragmented.

• Hsi (Hayano Sleep Index)

Developed as a robust marker for Non-REM (NREM) sleep detection, Hsi quantifies the degree of high-frequency (HF) power concentration into a narrow frequency range [51]. This index is based on the physiological characteristic that respiratory frequency becomes highly regular during NREM sleep. Unlike conventional HRV indices such as LF/HF or scaling exponent α, Hsi demonstrates superior accuracy (AUC = 0.86) in discriminating NREM from wakefulness and REM sleep, making it an ideal biomarker for sleep stage monitoring via wearable sensors.

The Utility and Clinical Significance of the Hayano Sleep Index (Hsi)

(1)

High Discriminant Performance as a Univariate Marker

The Hayano Sleep Index (Hsi) functions as a powerful univariate indicator for detecting non-rapid eye movement (NREM) sleep. Previous studies typically required complex multivariate models combining 10–80 features to achieve approximately 80% classification accuracy. In contrast, Hsi alone demonstrates comparable or superior performance, with an area under the curve (AUC) of 0.86. Using a cutoff value of >70%, Hsi achieves 77% sensitivity and 80% specificity in discriminating NREM sleep from wakefulness and REM sleep. Furthermore, when combined with the amplitude of body movement (ABM), Hsi enables effective three-stage classification (wake, NREM, and REM), highlighting its practical utility in simplified sleep staging.

(2)

Robustness and Physiological Basis

A key advantage of Hsi lies in its strong physiological foundation and robustness. Unlike conventional HRV indices such as LF, HF, LF/HF ratio, or scaling exponents (α), which depend on relative or indirect measures of autonomic activity, Hsi quantifies the spectral concentration of the high-frequency (HF) component. This design makes it largely independent of absolute power and less affected by confounding factors such as age, respiratory frequency, and health status.

Importantly, Hsi reflects the increased regularity of central respiratory drive during NREM sleep, providing a direct physiological interpretation. Notably, Hsi derived from R–R intervals has been shown to be more robust to noise than that derived from direct respiratory signals, likely because respiratory measurements are more susceptible to motion artifacts and environmental disturbances.

(3)

Applications in Wearable and Clinical Monitoring

Hsi is particularly well suited for applications in wearable technology and large-scale physiological data analysis. As a marker of sinoatrial rhythm stability, it serves as a key feature in automated algorithms for estimating sleep stages from long-term ECG or pulse-based recordings.

In clinical settings, Hsi enhances Holter ECG analysis by enabling precise identification of sleep periods, thereby improving the interpretation of transient cardiac events such as arrhythmias. In addition, Hsi contributes to the screening of sleep-disordered breathing (SDB), including sleep apnea. By accurately identifying true sleep periods, it reduces the risk of false-negative results that may occur when the absence of cyclic variation is due to wakefulness rather than the absence of pathological events (

Table 4. Summary of Hsi Characteristics.

Feature	Description
Metric Type	Univariate, Spectral-shape-based (Nonlinear/Complexity Domain)
Physiological Trigger	Concentration of HF power due to regularized respiration in NREM
Key Advantage	High AUC (0.86) independent of age and Autonomic Tone (Power)
Primary Use Case	Automated NREM detection for Wearables and Holter Monitoring

).

• Heart Rate Recovery (HRR)

Heart Rate Recovery (HRR) is a simple yet clinically meaningful index that quantifies the rate at which heart rate declines immediately after the cessation of exercise [60,61]. HRR primarily reflects the rapid reactivation of parasympathetic (vagal) activity and the withdrawal of sympathetic drive following physical exertion [62,63,86,87,88,89,90].

HRR is typically calculated as the difference between heart rate at the end of exercise and heart rate measured at a fixed time point during recovery, most commonly 1 min post-exercise:

$${\mathrm{HRR}}_{1\min }={\mathrm{H}\mathrm{R}}_{\mathrm{e}\mathrm{n}\mathrm{d}}-{\mathrm{H}\mathrm{R}}_{1\mathrm{m}\mathrm{i}\mathrm{n}}$$

A larger HRR value indicates faster autonomic recovery, whereas a smaller value reflects delayed recovery. Physiologically, the initial phase of HRR (within the first 30–60 s) is dominated by vagal reactivation, making HRR a surrogate marker of parasympathetic function. Consequently, impaired HRR has been consistently associated with autonomic dysfunction, reduced cardiorespiratory fitness, increased cardiovascular morbidity, and elevated all-cause mortality risk.

From a clinical perspective, commonly used reference thresholds for HRR at 1 min are as follows:

• ≥20 bpm: normal autonomic recovery;
• 12–19 bpm: borderline or mildly impaired recovery;
• <12 bpm: markedly impaired recovery, associated with increased cardiovascular risk.

It should be noted that these thresholds depend on exercise modality, protocol, age, fitness level, and medication status, and therefore should be interpreted in context.

Unlike conventional heart rate variability (HRV) indices, which require stationary resting conditions, HRR captures dynamic autonomic responsiveness during a physiological transition. In contrast to Heart Rate Fragmentation (HRF), which reflects instability and non-autonomic irregularity of sinoatrial node control under resting conditions, HRR evaluates the efficiency of autonomic reorganization following sympathetic activation.

Thus, HRR provides complementary information to HRV- and HRF-based metrics: while HRV and Hsi characterize autonomic modulation and rhythm integrity at rest, HRR reflects the adaptability and resilience of the autonomic nervous system under stress and recovery conditions. Integrating HRR- with HRF-related indices may therefore offer a more comprehensive assessment of cardiovascular autonomic regulation across both static and dynamic states (

Table 5. Summary of Heart Rate Recovery (HRR) Research by Key Themes.

Research Theme	Key Reference	Core Findings/Significance
Mortality Prediction	Cole et al., NEJM (1999) [61]	Demonstrated that delayed HRR is strongly and independently associated with increased mortality.
Reference Values	Jou et al., EJPC (2025) [60]	Established large-scale normative values for HRR and confirmed its association with survival in cycle exercise testing.
Parasympathetic Link	Fonseca et al., Sci Rep (2024) [89]	Found that the speed of HRR is closely linked to resting cardiovagal modulation (e.g., SD1 from Poincaré plots).
Autonomic Remodeling	Facioli et al., Sci Rep (2021) [90]	Investigated the relationship between HRR, HRV, and Baroreflex Sensitivity (BRS) during the post-exercise recovery phase.

).

3. Future Perspectives: From Big Data to Edge Intelligence and Non-Ergodic Dynamics

The field of wearable HRV analysis is transitioning from descriptive monitoring to predictive intervention. This section explores the convergence of machine learning, edge computing, and fundamental reassessments of biological time series.

3.1. Big Data, Artificial Intelligence, and Future Perspectives

In recent years, the widespread adoption of wearable sensors has led to a rapid expansion of HRV data from individual-level measurements to large-scale population datasets. Large databases, such as ALLSTAR, have enabled the identification of long-term variability patterns and inter-individual differences that were difficult to capture in small-scale studies [91,92]. As a result, HRV research is shifting from individual assessment toward the understanding of population-level dynamics, accelerating the transition from descriptive analysis to predictive and interventional medicine.

Within this big data context, machine learning and deep learning have become essential tools. However, the use of high-dimensional and nonlinear models introduces a fundamental trade-off: improved predictive performance often comes at the cost of reduced interpretability. To address this issue, Explainable Artificial Intelligence (XAI) has been proposed as a framework to provide human-understandable explanations of model decisions [93]. Nevertheless, the concept of “explainability” itself warrants careful reconsideration. From a theoretical perspective, large-scale neural networks are not inherently “unexplainable”; rather, they can be viewed as high-dimensional languages in their own right. A neural network with billions of parameters may be interpreted as encoding its entire structure as a single “word” within an extremely large symbolic space. In this sense, if the full parameter set can be transmitted and reconstructed, the system can be considered “understood” in a machine-readable sense. However, such representations far exceed human cognitive limits and cannot be reduced to conventional, human-readable explanations. Therefore, the term “Explainable AI” may be somewhat misleading. A more appropriate interpretation may be “Human-Readable AI,” emphasizing that explanations are not complete representations of the model itself, but rather approximations, abstractions, and compressions tailored to human understanding. This perspective aligns with concepts in fuzzy logic and approximate reasoning, where explanations are not exact reproductions of underlying mechanisms but practical mappings into interpretable forms. Importantly, humans themselves do not fully understand their own decision-making processes. In this regard, the opacity of AI systems is not an anomaly but a characteristic shared by complex systems in general. Although artificial neural networks can, in principle, be expressed within extended logical frameworks, making them theoretically describable, the resulting representations are often too large and complex to be practically interpretable. Thus, “explainability” should be regarded as a relative concept, dependent on the observer’s cognitive constraints.

Future HRV research must therefore balance machine readability and human readability. In clinical applications, predictive accuracy alone is insufficient; the rationale behind model outputs must also be interpretable and clinically meaningful. This necessitates the integration of XAI approaches with physiologically grounded feature design and model constraints. In parallel, advances in biofeedback research [94,95] are transforming HRV from a passive monitoring signal into an actionable physiological target. The integration of real-time HRV analysis with AI enables adaptive interventions, such as stress regulation and autonomic training, paving the way for personalized healthcare systems.

Overall, HRV research is evolving toward the integration of big data analytics, artificial intelligence, and the study of non-ergodic physiological dynamics. Future developments will require not only statistical modeling but also the incorporation of physiological interpretability and dynamic system perspectives, ultimately supporting more reliable personalized medicine and real-world health monitoring.

3.2. Long-Term Monitoring and Circadian Integration

The paradigm shift toward long-term monitoring is increasingly supported by the evolution of wearable sensors, transitioning from simple fitness trackers to clinically validated diagnostic tools (Convergence of wearable technology and Big Data Analytics).

Validation of Consumer Wearables: Recent studies have rigorously assessed the accuracy of commercial devices. For instance, the Oura Ring [38] and Polar H10 [39] have demonstrated high correlation with gold-standard ECG for nocturnal HRV and resting states. However, research by Yuda et al. (2020) [45] cautions that pulse rate variability (PRV) obtained via photoplethysmography (PPG) should be treated as a distinct biomarker rather than a direct surrogate for ECG-derived HRV, especially during physical activity or autonomic stress.

Automated Detection of Sleep Pathologies: Wearable integration has revolutionized the screening for Sleep-Disordered Breathing (SDB). By utilizing the Cyclic Variation in Heart Rate (CVHR), researchers can now quantitatively detect sleep apnea using only watch-based sensors [71]. More recently, the inclusion of Inertial Measurement Units (IMU) within wrist-worn devices [41,42] and smartphones [75] has further refined these algorithms, allowing for the decoupling of respiratory effort from autonomic responses.

Strategic Integration for Population Health: The synergy between high-fidelity biomarkers like Hsi [50] and long-term big data (e.g., ALLSTAR database) enables a “preventive” rather than “reactive” healthcare model. By identifying the “random component” of HRV that increases with both development and degeneration [10], and monitoring the suppression of vagal control in chronic conditions like chronic pain [64], clinicians can move toward personalized autonomic profiling. This integration is essential for managing the rising demand for automated health evaluation in the era of digital big data.

3.3. From LLMs to SLMs: The Rise of Edge Intelligence and Real-Time Prediction

While Large Language Models (LLMs) currently dominate the AI landscape, their high computational costs and significant latency make them unsuitable for power-constrained wearable applications. The future of personalized medicine lies in Small Language Models (SLMs) and Edge Computing, where specialized, efficient algorithms process biological signals locally.

Real-time Feedback at the Edge: Processing data directly on the wearable device (Edge AI) eliminates the latency inherent in cloud computing. This is critical for life-saving interventions. For instance, the integration of indices like Hsi [50] or DC [29] into edge-compatible architectures allows for the immediate identification of high-risk physiological states, such as obstructive sleep apnea episodes [71] or acute autonomic shifts prior to cardiovascular events.

The Challenge of Sudden Death and Seizures: Predicting catastrophic events like sudden cardiac death (SCD) or epileptic seizures remains a formidable challenge. This difficulty is linked to the physical limits of proving randomness in complex biological systems. However, modern approaches such as Heart Rate Fragmentation (HRF) [31,32,33] and ultra-short-term fluctuation analysis [58] provide new windows into the breakdown of neuroautonomic control. These “fragmented” rhythms serve as precursors to system instability that traditional linear models fail to capture.

Convergence with Molecular and Genetic Insights: The limitation of pure HRV analysis is mitigated when data-driven science converges with molecular biology and genetic research. While phenotypic HRV is a strong predictor of all-cause mortality [57], its combination with genetic markers for myocardial infarction allows for a more stratified risk assessment.

The Role of Big Data in Preventive Notification: By leveraging large-scale databases like ALLSTAR [91,92], edge algorithms can be “pre-trained” on millions of hours of human data to recognize the “peak windows of vulnerability” [7]. This enables wearables to move beyond simple monitoring to providing “Early Notifications.” Even if a catastrophic event cannot be predicted with 100% certainty due to biological stochasticity, the ability to detect a significant deviation from a patient’s circadian baseline [63] will become a cornerstone of preventive medicine.

3.4. Reassessing Ergodicity: Analysis Within the 100-Year Human Lifespan

A fundamental theoretical reconsideration is required regarding the nature of biological time series. Conventional heart rate variability (HRV) indices are predominantly calculated under the assumption of ergodicity—the premise that the statistical properties of a single individual over a long period (time average) are identical to the average properties observed across a large population at a single point in time (ensemble average). However, this assumption is increasingly scrutinized in biological systems for the following reasons: The Non-convergence of Averages: In a strictly ergodic system, the observation time must approach infinity (t → ∞) for the time average to converge to the ensemble average. Given the finite nature of the human lifespan—limited to approximately 100 years—biological processes are inherently non-ergodic. When the observation window is constrained by these natural limits, the time average of an individual’s autonomic state may never converge to the population mean.

Autonomic nervous system data are non-ergodic, meaning that the individual time average $\overline{{\mathrm{x}}_{\mathrm{i}}}$ does not necessarily match the ensemble average E[x(t)]. In particular, when the observation window T is constrained, the discrepancy indicator Var( $\overline{{\mathrm{x}}_{\mathrm{i}}}$) > 0 holds, making individual order optimization essential. The concept of being non-ergodic, that is, the time average of an individual does not match the ensemble average (group average), can be expressed mathematically as follows. First, let x(t) be an individual’s physiological index such as heart rate variability. The time average, for an individual i, over an observation window T, is defined as follows (Definitions of time average and ensemble average).

$$\overline{{\mathrm{x}}_{\mathrm{i}}}=\lim \left(\mathrm{T}\to \mathrm{\infty }\right)\left({\displaystyle \frac{1}{\mathrm{T}}}\right){\int }_0^T{\mathrm{x}}_{\mathrm{i}}\left(\mathrm{t}\right)\mathrm{d}\mathrm{t}$$

(1)

The expected value of the entire population at a given time t is defined as follows (ensemble mean):

$$\mathrm{E}\left[\mathrm{x}\left(\mathrm{t}\right)\right]=\int \mathrm{x}\mathrm{P}\left(\mathrm{x},\mathrm{t}\right)\mathrm{d}\mathrm{x}$$

(2)

If the system is non-ergodic, the time average of the individuals will not converge to the ensemble average (the condition for non-ergodicity).

$$\overline{{\mathrm{x}}_{\mathrm{i}}}\ne \mathrm{E}\left[\mathrm{x}\left(\mathrm{t}\right)\right]$$

(3)

More strictly, if the observation window T is constrained to a finite T_max by natural limits (physiological constraints or limits on the measurement duration), the mean discrepancy will persist stochastically.

If the time-averaged variance of each individual in the ensemble is not zero, the system is non-ergodic (a measure of inconsistency).

$$\mathrm{V}\mathrm{a}\mathrm{r}\left(\overline{{\mathrm{x}}_{\mathrm{i}}}\right)=\mathrm{E}\left[{\left(\overline{{\mathrm{x}}_{\mathrm{i}}}-\mathrm{E}\left[\mathrm{x}\right]\right)}^2\right]>0$$

(4)

If the system is ergodic, over a long enough period of time, one person will experience all of the possible states of the population, so the individual average will match the population average and the variance in the above equation will converge to 0. However, if individual differences are fixed, as in the autonomic nervous system, or if there is a limit to the observation time, the above equation will continue to maintain a positive value, and the characteristics of an individual will not be submerged in the population average. The autonomic nervous system indices targeted in this study have non-ergodic properties due to the finite observation window T and stationary heterogeneity. Therefore, rather than setting a uniform threshold based on the population average, it is important to ensure robustness in the order selection p_opt, which is adapted to the individual’s time series characteristics.

The Individualized Aging Trajectory: Biological time is “arrow-like” and irreversible. As demonstrated by Hayano et al. (2018) [10], the random component of HRV increases with both developmental and degenerative stages of life. These shifts represent a non-stationary process where the “state space” of the individual is constantly shrinking or shifting, making the cross-sectional population data (ensemble) a poor proxy for an individual’s future (time). New Frontiers in HRV Metrics Recognizing this non-ergodicity opens the door to the development of novel HRV metrics that explicitly account for the bounded and individualized nature of physiological time series: Life-stage Sensitive Metrics: Instead of comparing an individual to a static “normal range,” new metrics should incorporate the known temporal boundaries of human life. This allows for the assessment of “physiological age” versus “chronological age” by analyzing how an individual’s autonomic complexity deviates from their own longitudinal baseline [63]. Accounting for Pathological Discontinuity: In non-ergodic systems, rare but catastrophic events (like the onset of SDB or acute cardiac events) fundamentally alter the system’s dynamics. Metrics like Hsi [50] or Heart Rate Fragmentation [31,32,33] provide a way to detect these discrete shifts in the “biological state,” offering more accurate representations of an individual’s specific aging trajectory and physiological decline. In conclusion, by moving away from the ergodic assumption, we can transition from “population-based medicine” to a truly personalized chronobiology, where the 100-year lifespan is treated as a finite, unique, and non-repeating dynamical system.

3.5. Nonlinear Association Analysis: Beyond Linear Correlations Using Chatterjee’s Xi

In medical and physiological research, many biomarkers exhibit J-shaped or U-shaped relationships with health outcomes rather than simple linear trends. These patterns indicate the presence of an “optimal range” in which risk is minimized, while values that are either too low or too high are associated with increased morbidity or mortality.

Well-known examples of J-shaped or U-shaped associations are widely observed across medical research. For instance, body mass index (BMI) demonstrates increased mortality risk at both low and high extremes, with the lowest risk typically occurring within the high-normal or mildly overweight range. Similarly, diastolic blood pressure (DBP) follows a J-shaped pattern, as values below approximately 70–80 mmHg may impair coronary perfusion, while elevated levels increase the likelihood of cardiovascular events. Alcohol consumption presents a classic J-curve relationship in which moderate intake (approximately 10–20 g/day) has been associated with lower mortality compared with both complete abstinence and heavy drinking. Sleep duration likewise exhibits a U-shaped association, with 6–7 h generally considered optimal, whereas both shorter and longer sleep durations are linked to increased mortality risk.

Comparable nonlinear characteristics have also been reported in cardiac parameters, including basal heart rate (BHR). However, conventional statistical methods such as Pearson’s and Spearman’s correlation coefficients are primarily designed to detect linear or monotonic relationships. Consequently, they may underestimate—or entirely fail to detect—clinically meaningful nonlinear dependencies (Figure 2).

To overcome this limitation, Chatterjee’s correlation (Xi coefficient), introduced in 2021, provides a model-free and rank-based measure capable of detecting general functional relationships, including highly nonlinear and non-monotonic patterns such as U-shaped associations [96]. Unlike traditional correlation measures, Xi does not assume linearity or monotonicity, making it particularly suitable for complex physiological data.

Chatterjee’s Xi coefficient is a relatively new, rank-based correlation measure designed to overcome the limitations of classical indices like Pearson’s r or Spearman’s rho. While Pearson’s correlation only detects linear relationships, the xi coefficient is capable of capturing any functional dependency, including highly non-linear and non-monotonic relationships (e.g., parabolic or sinusoidal patterns). A key feature of the xi coefficient is that it converges to 1 if and only if one variable is a measurable function of the other, and converges to 0 if the variables are independent.

To calculate the coefficient ξ_n(X, Y) for a sample of size n. Tie-free Case (No duplicate values in Y) If all observations of Y are distinct, the formula is simplified as:

$${\mathsf{\xi }}_{\mathrm{n}}\left(\mathrm{X},\mathrm{Y}\right)=1-{\displaystyle \frac{3{\sum }_{i=1}^{n-1}\left|r_{i+1}-{\mathrm{r}}_{\mathrm{i}}\right|}{{\mathrm{n}}^2-1}}$$

General Case (With ties/duplicates in Y), if there are duplicate values in Y, the formula is defined as:

$${\mathsf{\xi }}_{\mathrm{n}}\left(\mathrm{X},\mathrm{Y}\right)=1-{\displaystyle \frac{\mathrm{n}{\sum }_{i=1}^{n-1}\left|r_{i+1}-{\mathrm{r}}_{\mathrm{i}}\right|}{2{\sum }_{i=1}^n\mathrm{}{\mathrm{l}}_{\mathrm{i}}(\mathrm{n}-l_i)}}$$

Variable Definitions n, total number of pairs (X, Y). r_i is the rank of Y assigned after sorting the data by the values of X. l_i is the number of indices j such that Y_j ≥ Y_i, the sum of absolute differences between successive ranks of Y (after sorting by X) is

$${\sum }_{i=1}^{n-1}\left|r_{i+1}-{\mathrm{r}}_{\mathrm{i}}\right|$$

It can identify a relationship even if the correlation is zero in the classical sense. And it is based on ranks, it is robust against outliers and does not require the assumption of normality. Application in HRV Research. In the context of Heart Rate Variability (HRV), the xi coefficient is particularly useful for quantifying the deterministic structure of the autonomic nervous system. While aging often leads to a loss of complexity, traditional linear measures may only show a decrease in amplitude” The xi coefficient allows us to quantify whether the beat-to-beat dynamics are losing their functional coupling and transitioning toward a more stochastic (random) state.

In the context of wearable HRV monitoring, incorporating Chatterjee’s Xi may substantially improve the detection of optimal physiological ranges and enhance risk stratification, especially in scenarios where traditional linear models fail to capture clinically relevant nonlinear dynamics.

The nonlinear relationship between basal heart rate (BHR) and age was examined with reference to Hayano et al. (2017) [97], which characterized age- and sex-dependent variations in daily-life heart rate and proposed the concept of an “optimal range” for heart rate metrics. In the present analysis, BHR was preliminarily estimated using the ALLSTAR database, with approximately 10 samples per age group to provide an illustrative comparison across the lifespan. Figure 3 presents both linear and nonlinear representations of the association between age and resting heart rate. Mean BHR values for each age group are shown with error bars reflecting physiologically reported ranges. The orange line indicates a linear regression corresponding to the assumptions of the Pearson correlation coefficient, which models a simple linear relationship between age and heart rate. In contrast, the green curve represents a nonparametric LOWESS regression, capturing the underlying nonlinear structure of the data. The red curve illustrates a monotonic but nonlinear inverse relationship, consistent with rank-based association measures such as the Chatterjee correlation coefficient. The observed pattern suggests a rapid decline in BHR during younger ages followed by a more gradual change in adulthood and aging. This demonstrates that physiological data often exhibit inherently nonlinear dynamics, which cannot be adequately described by linear correlation alone. While Pearson correlation captures only linear trends, rank-based or nonparametric approaches are more suitable for detecting general functional relationships. Overall, this figure highlights the importance of adopting nonlinear analytical frameworks when interpreting age-related changes in cardiovascular physiology.

4. Conclusions

This review has summarized the current landscape of heart rate variability (HRV) metrics and their applications within wearable sensor technologies. The integration of HRV analysis into wearable devices has enabled a paradigm shift from episodic clinical assessment to continuous, real-time physiological monitoring. Conventional time- and frequency-domain indices remain foundational; however, increasing attention has been directed toward nonlinear measures—including Heart Rate Fragmentation (HRF) and symbolic indices such as Hsi—which offer additional insight into autonomic and non-autonomic regulatory dynamics associated with aging, cardiovascular disease, and systemic physiological stress.

Advances in data science have further expanded the potential of wearable HRV analytics. Large-scale physiological repositories and machine learning frameworks have facilitated improved pattern recognition and risk stratification. At the same time, practical deployment in wearable systems requires computationally efficient implementations. Emerging approaches such as edge AI and lightweight models provide promising solutions for real-time processing, reduced energy consumption, and on-device inference, thereby enhancing the feasibility of continuous cardiovascular risk monitoring in daily life.

Taken together, these considerations highlight the necessity of maintaining a clear distinction between measurement and inference, and underscore the importance of physiological interpretability as a central criterion in the development and evaluation of HRV-related methodologies.

Importantly, HRV indices are calculated from directly measured beat-to-beat fluctuations in cardiac intervals, and therefore inherently reflect underlying autonomic and cardiovascular dynamics. In contrast, some recent approaches attempt to estimate HRV from indirect signals, such as physical activity or behavioral data, without actually measuring these physiological fluctuations. This is conceptually similar to estimating body temperature without using a thermometer: although values that appear reasonable can be generated, it remains unclear whether they truly reflect physiological states or clinical conditions. For this reason, simply demonstrating that estimated HRV values resemble or correlate with measured HRV is not sufficient. The key question is whether these estimates preserve the physiological meaning of HRV—namely, its relationship to autonomic regulation and disease states. Without clear evidence of this, such estimated metrics have limited interpretability and should be treated with caution.

It is also important to distinguish these approaches from pulse rate variability (PRV)-based methods. Unlike indirect estimation, PRV is derived from pulse wave signals that directly capture beat-to-beat cardiovascular fluctuations, similar to HRV itself. Because both are based on the same underlying physiological processes, the use of PRV as a surrogate measure is physiologically justified.

Furthermore, caution is needed when interpreting studies that use machine learning to model relationships between behavioral signals and autonomic activity. Although such models may achieve high predictive accuracy by reproducing known correlations (for example, between activity level and HRV), prediction alone does not imply physiological understanding. Accurately predicting a variable is not the same as explaining the mechanisms that generate it. Therefore, claims of new physiological insights based solely on predictive performance should be carefully evaluated, particularly when mechanistic evidence is lacking.

This review also highlights the importance of accounting for inter-individual variability and the time-dependent nature of physiological signals. Human biological systems exhibit substantial heterogeneity, and future HRV research should increasingly emphasize personalized modeling strategies rather than relying solely on population-averaged assumptions. Incorporating nonlinear dynamics and individualized baselines may improve the clinical interpretability and predictive utility of wearable-derived HRV metrics.

In conclusion, wearable sensor technologies combined with advanced HRV analytics represent a rapidly evolving field with significant implications for preventive medicine and remote health monitoring. Continued methodological refinement, validation in diverse populations, and integration with intelligent edge-based systems will be essential for translating HRV-based insights into reliable, scalable healthcare applications.