WASO as a Stage-Resolved Window for Detectable HRV Differences in Paradoxical Insomnia

Abstract

Paradoxical insomnia (PI) is characterized by a discrepancy between subjective sleep complaints and objectively preserved sleep, yet its autonomic mechanisms remain poorly understood. This study examined stage-specific autonomic characteristics of PI using heart rate variability (HRV) analyses in a large population-based cohort. HRV features were extracted from non-overlapping five-minute windows across non-rapid eye movement (NREM) sleep, rapid eye movement (REM) sleep, and wake after sleep onset (WASO). Group differences were evaluated using FDR-corrected univariate analysis, multivariate embedding, and supervised machine learning. Whole-night, NREM, and REM features showed substantial overlap among groups. In contrast, the most consistent between-group differences emerged during WASO. Multivariate analysis showed the greatest group displacement during WASO, with UMAP centroid distances exceeding those observed during NREM and REM sleep. Supervised models trained on WASO-specific features achieved the highest classification performance, yielding an accuracy of 0.629 and an F1-score of 0.683 for PI versus normal sleep. Taken together, these findings suggest that WASO is the stage in which between-group HRV differences are most consistently detectable across complementary analyses, although several dispersion-based findings were substantially influenced by WASO window count.

1. Introduction

Sleep is a fundamental biological process essential for maintaining physiological homeostasis, cognitive performance, emotional regulation, and long-term health. Disruption of sleep has been associated with impaired cognitive function, mood disturbances, metabolic dysregulation, and increased risk of cardiovascular and neurodegenerative diseases [1,2,3,4,5,6]. Insomnia is one of the most prevalent sleep disorders, affecting approximately 10–25% of adults, and is characterized by difficulties in sleep initiation, maintenance, or early morning awakening despite adequate opportunity for sleep, accompanied by daytime impairment [7,8]. Chronic insomnia is strongly linked to psychological distress, cognitive decline, and elevated cardiometabolic risk [9,10,11,12,13].

Among insomnia phenotypes, paradoxical insomnia (PI), also referred to as sleep state misperception, represents a particularly challenging and poorly understood subtype. PI is defined by a marked discrepancy between severe subjective sleep complaints and objectively preserved sleep as measured by polysomnography (PSG) [14,15,16]. Individuals with PI often report poor sleep quality despite normal total sleep time and sleep efficiency, complicating diagnosis and frequently leading to inappropriate or prolonged pharmacological treatment [17,18]. Although PI has historically been regarded as a primarily perceptual or cognitive phenomenon, accumulating evidence suggests that it is accompanied by measurable neurophysiological alterations. Electroencephalographic studies have reported cortical hyperarousal during sleep, including increased beta, alpha, and sigma activity during NREM sleep [19,20,21,22], while neuroimaging studies have demonstrated structural alterations in subcortical regions involved in arousal regulation and emotional processing [23]. In addition, changes in sleep microarchitecture, such as reduced sleep spindle activity and increased cyclic alternating pattern frequency, have been reported, indicating heightened sleep instability [24,25]. However, despite growing evidence for central nervous system involvement, the peripheral autonomic correlates of PI remain insufficiently characterized.

The autonomic nervous system (ANS) plays a critical role in sleep regulation and is tightly coupled with central arousal systems through hypothalamic–brainstem pathways. Heart rate variability (HRV), derived from beat-to-beat fluctuations in cardiac intervals, provides a non-invasive index of autonomic regulation and has been widely used to investigate autonomic activity in insomnia [26,27,28,29]. Prior studies have reported altered HRV profiles in insomnia, often interpreted within a hyperarousal framework. However, most HRV studies in insomnia have relied on whole-night or averaged metrics, which may obscure transient autonomic alterations associated with dynamic sleep–wake transitions.

Autonomic regulation during sleep is highly stage dependent, with parasympathetic predominance during non-rapid eye movement (NREM) sleep, increased sympathetic instability during rapid eye movement (REM) sleep [30]. Wake after sleep onset (WASO) represents a transitional and physiologically unstable state characterized by frequent arousals and repeated shifts between sleep and wakefulness. If PI is characterized by state-dependent rather than sustained hyperarousal, autonomic alterations may preferentially emerge during such transitional states. However, sleep-stage–resolved HRV studies specifically targeting PI are scarce, and no prior work has systematically examined whether autonomic differentiation in PI clusters during WASO rather than during consolidated sleep stages.

Based on these considerations, the present study aimed to characterize stage-specific autonomic features of PI using sleep-stage–resolved heart rate variability analysis in a large population-based cohort. We hypothesized that (1) whole-night HRV metrics would show substantial overlap among PI, objective insomnia (OI), and normal sleep (N) groups, whereas (2) HRV features derived specifically from WASO would demonstrate greater autonomic differentiation than those derived from NREM or REM sleep. To test these hypotheses, we integrated univariate statistical analysis, multivariate pattern analysis, and supervised machine learning to systematically evaluate whether autonomic alterations in PI are globally expressed or preferentially emerge during specific sleep–wake states.

2. Materials and Methods

2.1. Dataset and Exclusion Criteria

This study utilized data from the Sleep Heart Health Study (SHHS) [31], a large-scale multicenter cohort including overnight in-home PSG, electrocardiography (ECG), and standardized sleep questionnaires. All available recordings from SHHS1 and SHHS2 (n = 10,115) were initially screened.

To ensure reliable autonomic analysis, a multi-step exclusion procedure was applied (Figure 1). Recordings with less than 6 h of scorable PSG data or low-quality ECG signals were excluded, and only recordings belonging to the highest three PSG quality levels were retained to minimize artifacts and segmentation errors affecting HRV estimation. Participants exhibiting extreme physiological values (heart rate >150 bpm or <30 bpm, or SpO₂ < 75% for more than 10% of total sleep time) were excluded to avoid non-physiological and non-stationary RR interval dynamics.

Individuals with moderate-to-severe sleep-disordered breathing (apnea–hypopnea index ≥ 10 events/h) or frequent periodic limb movements were removed to minimize confounding autonomic fluctuations associated with comorbid sleep disorders. In addition, participants taking medications known to affect autonomic regulation, including beta-blockers, sympathomimetics, antidepressants, and benzodiazepines, were excluded. Recordings with missing PSG-derived sleep metrics required for phenotype classification were also discarded. After applying all exclusion criteria, 1970 participants remained eligible for phenotype classification and subsequent analysis.

2.2. Insomnia Phenotypes Classification

Participants were classified into PI, OI, and normal sleep (N) groups using a two-stage procedure integrating subjective insomnia complaints and objective polysomnographic (PSG) criteria (

Table 1. Operational criteria for classification of insomnia phenotypes based on subjective complaints and polysomnographic parameters.

Group	Classification Criteria	n = 1970
OI (Objective Insomnia)	Subjective complaint of insomnia and objective insomnia indicated by PSG PSG criteria: - TST < 6 h or SE < 85% or WASO > 30 min or SL > 20 min [32,33]	625
PI (Paradoxical Insomnia)	① Subjective complaint of insomnia with normal PSG findings PSG criteria: - TST ≥ 6 h and SE ≥ 85% and WASO ≤ 30 min and SL ≤ 20 min ② (oTST-sTST) ≥ 1 h [14]	261
N (Normal Sleep)	No subjective complaint of insomnia and normal PSG findings PSG criteria: - TST ≥ 6 h and SE ≥ 85% and WASO ≤ 30 min and SL ≤ 20 min	201
Unclassified	Participants not meeting the criteria for any of the above groups	883

).

In the first stage, subjective insomnia complaints were assessed using the Sleep Heart Health Study Sleep Habits Questionnaire. Participants reporting insomnia symptoms as “often” or “almost always” for at least one item were classified as having clinically significant insomnia complaints.

In the second stage, objective sleep parameters derived from PSG were applied to define phenotypes. OI was defined as the presence of insomnia complaints in combination with at least one abnormal PSG criterion, including total sleep time (TST) < 6 h, sleep efficiency (SE) < 85%, WASO > 30 min, or sleep latency (SL) > 20 min, consistent with established insomnia definitions [32,33]. PI was defined as the presence of insomnia complaints despite preserved objective sleep (TST ≥ 6 h, SE ≥ 85%, WASO ≤ 30 min, and SL ≤ 20 min) or a marked discrepancy between objective and subjective total sleep time (oTST − sTST ≥ 60 min), reflecting sleep state misperception [14]. N was defined as the absence of insomnia complaints with all PSG parameters within normal ranges.

Participants who did not fully satisfy the operational criteria for PI, OI, or N were categorized as unclassified and excluded from subsequent analyses. This group primarily included individuals with partial or borderline patterns that prevented definitive phenotype assignment. For example, participants without subjective insomnia complaints but with mildly reduced sleep efficiency (SE < 85%) or shortened total sleep time (TST < 6 h) were not classified as OI, as OI requires the presence of subjective complaints. Similarly, individuals reporting insomnia complaints but with preserved PSG parameters and an objective–subjective total sleep time discrepancy below the 60 min threshold were not classified as PI. These conservative rules were applied to minimize phenotype misclassification and ensure clear group separation. According to these criteria, 625 participants were classified as OI, 261 as PI, and 201 as N, while 883 were unclassified.

To ensure balanced group sizes for statistical comparisons and machine learning analyses, a random subsample of 200 participants was selected from each group. This approach was adopted to prevent class imbalance, which can bias non-parametric statistical tests and machine learning models, and to maintain computational feasibility for repeated multivariate analyses and model training procedures. Random selection was performed using a fixed random seed to ensure reproducibility.

To evaluate whether this sampling strategy influenced the study outcomes, additional sensitivity analyses were conducted using the full available dataset without subsampling. These analyses were designed to confirm that the principal stage-dependent HRV patterns identified in the balanced subsample remained consistent when all available participants were included.

2.3. ECG Signal Preprocessing and R-Peak Detection

ECG signals were extracted from overnight polysomnography recordings and processed using a unified preprocessing and R-peak detection pipeline (Figure 2). To ensure consistency across recordings, all ECG signals were resampled to a uniform sampling rate of 250 Hz, as SHHS1 recordings were originally acquired at 125 Hz and SHHS2 recordings at 250 or 256 Hz. Analyses were restricted to the interval between sleep onset and final awakening to focus on sleep-related autonomic activity.

R-peaks were detected using the XQRS algorithm implemented in the Waveform Database (WFDB) toolkit [34], which is based on an enhanced Pan–Tompkins framework [35]. The algorithm employs bandpass filtering, signal differentiation, squaring, and adaptive thresholding to robustly identify QRS complexes under varying signal conditions. Detected beats were subjected to refractory period constraints to reduce false detections.

R–R interval time series were constructed from consecutive R-peaks. Intervals shorter than 0.3 s or longer than 2.5 s were considered physiologically implausible and removed. The proportion of excluded R–R intervals was low across all groups (OI: 0.30%, PI: 0.21%, N: 0.30%), indicating minimal artifact contamination. The resulting cleaned R–R interval series served as the basis for subsequent HRV feature extraction.

2.4. HRV Feature Extraction

HRV features were extracted from the cleaned R–R interval time series to characterize autonomic nervous system dynamics during sleep. HRV analysis was conducted using a short-term, window-based approach to capture transient and stage-dependent autonomic fluctuations that may not be reflected in whole-night averages.

2.4.1. Windowing Strategy and Stage-Wise Segmentation

HRV features were computed using non-overlapping 5 min windows in accordance with established guidelines for short-term HRV analysis. This window length was selected based on the Task Force guidelines of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology, which recommend 5 min segments for reliable short-term HRV estimation [36]. Only windows fully contained within a single sleep stage were included to prevent contamination across stage boundaries.

WASO was defined as all epochs scored as wake occurring between sleep onset and final awakening based on standard polysomnographic scoring criteria. Each valid window was assigned a sleep stage label (NREM, REM, or WASO) according to polysomnographic annotations. For each participant, HRV features were aggregated within each sleep stage using the median value across all valid windows, providing robust stage-specific representations of autonomic activity while minimizing the influence of outliers and transient artifacts.

In addition to the median value, within-subject dispersion of HRV features during each sleep stage was quantified using the interquartile range (IQR), calculated as the difference between the 75th and 25th percentiles across all valid 5 min windows belonging to that stage. These stage-specific IQR values were computed independently for each participant and represent intra-individual variability of autonomic responses within NREM, REM, or WASO periods rather than between-subject variability.

2.4.2. HRV Metrics

Standard HRV metrics were extracted to characterize autonomic regulation from complementary perspectives. Time-domain metrics were computed to quantify overall variability and short-term beat-to-beat fluctuations, reflecting both global autonomic modulation and parasympathetic activity. Instantaneous heart rate (IHR) was also derived from NN intervals as IHR = 60/RR (RR in seconds) and is reported in beats per minute (bpm). Frequency-domain metrics were derived to assess oscillatory components of heart rate dynamics associated with sympathetic and parasympathetic influences. In addition, non-linear HRV metrics were calculated to capture the complexity, irregularity, and fractal properties of heart rate time series that are not adequately described by linear measures alone. A comprehensive list of all HRV features, including their definitions and units, is provided in

Table 2. Summary of heart rate variability features used in this study.

Category	Feature	Description	Unit
Time-domain	AVNN	Mean of normal-to-normal (NN) intervals	ms
	SDNN	Standard deviation of all NN intervals, reflecting overall HRV	ms
	RMSSD	Root mean square of successive NN interval differences, reflecting short-term vagal modulation	ms
	pNN50	Percentage of successive NN intervals differing by more than 50 ms	%
Frequency- domain	Total power	Total spectral power across all frequency bands	ms2
	VLF power	Power in the very-low-frequency band (0.003–0.04 Hz)	ms2
	LF power	Power in the low-frequency band (0.04–0.15 Hz)	ms2
	HF power	Power in the high-frequency band (0.15–0.40 Hz)	ms2
	LF/HF ratio	Ratio of low- to high-frequency power, reflecting sympathovagal balance	-
Non-linear	SD1	Poincaré plot index of short-term variability	ms
	SD2	Poincaré plot index of long-term variability	ms
	SD1/SD2 ratio	Ratio of short- to long-term variability	-
	Sample Entropy	Quantifying signal complexity and irregularity	-
	Approximate Entropy	Measuring regularity of time-series patterns	-
	DFA α1	Short-term fractal scaling exponent from detrended fluctuation analysis	-
	DFA α2	Long-term fractal scaling exponent from detrended fluctuation analysis	-

.

2.5. Statistical Analysis

All statistical analyses were performed to evaluate stage-specific differences in HRV features among insomnia phenotypes. Analyses were conducted separately for NREM sleep, REM sleep, and WASO to avoid dilution of transient autonomic effects in whole-night averages.

The distribution of HRV features was assessed using visual inspection and normality tests. As most features did not satisfy normality assumptions, non-parametric statistical methods were employed. Group-wise comparisons were performed using two-tailed Mann–Whitney U test for PI vs. N and PI vs. OI comparisons within each sleep stage.

Because multiple HRV features were tested within each stage and phenotype contrast, p-values were adjusted to control the false discovery rate using the Benjamini–Hochberg procedure. FDR-adjusted p-values (q-values) are reported, and statistical significance was determined at q < 0.05. Effect sizes were additionally calculated using the common language effect size (CLES) and rank-biserial correlation coefficients.

To assess the robustness of the statistical findings to the sampling strategy, a sensitivity analysis was performed using the full dataset without subsampling, and the resulting patterns were compared with those obtained from the balanced subsample.

Because IQR-based dispersion metrics were derived from valid 5 min windows within each stage, we additionally examined their association with WASO window count. For the WASO stage, Spearman correlation analyses were performed between each IQR-based HRV feature and WASO window count in the pooled sample and within each phenotype group. Covariate-adjusted regression models were then fitted to test whether group differences in WASO dispersion metrics remained after adjustment for WASO window count. The same Benjamini–Hochberg FDR correction procedure was applied to these additional analyses.

HRV features demonstrating consistent stage-specific effects were subsequently used for unsupervised pattern analysis and machine learning–based classification. All statistical analyses were implemented using Python (v3.11)-based scientific computing libraries.

2.6. Unsupervised Pattern Analysis

Unsupervised pattern analysis was performed to explore the intrinsic structure of HRV features and to assess whether insomnia phenotypes exhibited separable autonomic patterns without imposing prior class labels. Analyses were conducted separately for each sleep stage (NREM, REM, and WASO) to preserve stage-specific autonomic characteristics.

Uniform Manifold Approximation and Projection (UMAP) was employed for non-linear dimensionality reduction in high-dimensional HRV feature spaces. UMAP was selected due to its ability to preserve both local neighborhood relationships and global data structure, making it well suited for visualizing complex physiological patterns. Prior to dimensionality reduction, HRV features were standardized to zero mean and unit variance to ensure equal contribution of features.

Two-dimensional UMAP embeddings were generated independently for PI vs. N and PI vs. OI comparisons within each sleep stage. To quantify relative multivariate group differences in the embedded space, Euclidean distances between group centroids were computed for each comparison and stage.

To evaluate whether the observed centroid separation exceeded chance-level expectations, a permutation-based centroid distance test was performed. After computing the UMAP embedding, group labels were randomly shuffled 1000 times while keeping the embedding fixed. For each permutation, centroid distances between groups were recalculated to generate an empirical null distribution. A one-sided empirical p-value was computed as the proportion of permuted centroid distances greater than or equal to the observed centroid distance.

UMAP analyses were primarily used to characterize stage-dependent multivariate patterns in HRV features, and the resulting separation patterns were interpreted in conjunction with univariate statistical analyses and supervised machine learning results.

2.7. Machine Learning Analysis

Supervised machine learning analysis was conducted to evaluate the discriminability of PI from OI and N based on stage-specific HRV features. Random Forest classifiers were trained and evaluated separately for each sleep stage (NREM, REM, and WASO) to assess whether stage-specific HRV feature sets differed in their relative information content for phenotype classification.

To reduce dimensionality and limit potential overfitting, model inputs were selected from HRV features that showed statistically significant between-group differences after FDR correction in the primary univariate analyses. Because some WASO-related significant features were IQR-derived dispersion measures, and these measures were later found to be associated with WASO window count, the machine learning results should be interpreted as reflecting the combined discriminative contribution of HRV features present in the stage-specific analysis rather than the independent predictive value of dispersion metrics alone.

To ensure reproducibility and methodological consistency, all data preprocessing and model training steps were implemented using a unified pipeline constructed with the Scikit-learn framework. First, feature values were standardized using z-score normalization (zero mean and unit variance) via standard scaling to prevent dominance of features with larger numerical ranges. Second, to address class imbalance, a dual strategy was applied at both the data and algorithmic levels. At the data level, the Synthetic Minority Over-sampling Technique (SMOTE) was applied to the training set to synthetically augment minority class samples and achieve balanced class distributions. In parallel, class weights were incorporated into the learning algorithm to impose higher penalties for misclassification of minority classes, encouraging the model to prioritize minority class representation during training.

The dataset was randomly partitioned into training (70%) and testing (30%) subsets using stratified sampling to preserve class proportions. Hyperparameter tuning was performed within the training set using grid search with 10-fold stratified cross-validation. Model performance was evaluated on the held-out test set using accuracy, precision, recall, F1-score, and average precision (AP; area under the precision–recall curve). These analyses were intended to compare the relative information content of stage-specific HRV feature sets rather than to establish a clinically deployable diagnostic classifier.

3. Results

3.1. Overall HRV Differences Across the Entire Sleep Period

Stage-agnostic comparisons of HRV features were first performed across the entire sleep period to examine whether global nocturnal autonomic profiles differed among insomnia phenotypes. HRV features averaged over the whole night were compared between PI and N, as well as between PI and OI.

Across all time-domain, frequency-domain, and non-linear HRV metrics, no statistically significant differences were observed between groups in either comparison. Effect size estimates were consistently small, indicating substantial overlap in whole-night HRV distributions among insomnia phenotypes.

These findings suggest that autonomic characteristics distinguishing PI are not adequately captured by whole-night HRV averages. Consequently, subsequent analyses focused on stage-specific HRV patterns to investigate whether transient autonomic alterations emerge within specific sleep–wake states.

3.2. Stage-Specific HRV Differences (Mann–Whitney U Test)

3.2.1. Stage-Specific HRV Differences Between Paradoxical Insomnia and Normal Sleep

Stage-specific comparisons of HRV features between PI and N were performed for NREM sleep, REM sleep, and WASO using the Mann–Whitney U test. To account for multiple comparisons across HRV features, statistical significance was evaluated based on Benjamini–Hochberg FDR–adjusted p-values (q < 0.05).

During consolidated sleep stages, autonomic profiles of the two groups showed substantial overlap. In both NREM and REM sleep, no HRV feature demonstrated significant between-group differences after FDR correction across time-domain, frequency-domain, or non-linear metrics.

In contrast, statistically significant differences emerged selectively during WASO, although the magnitude of these effects was generally modest and substantial overlap between groups remained. A total of 19 HRV features showed significant differences between PI and N after FDR correction (

Table 3. Stage-specific HRV differences between paradoxical insomnia (PI) and normal sleep (N) during wake after sleep onset (WASO).

Sleep Stage	HRV Features	PI (Median)	N (Median)	q-Value	CLES	Rank-Biserial r
WASO	SampEn IQR	0.276	0.127	<0.0001	0.339	−0.323
	ApEn IQR	0.163	0.087	0.0001	0.352	−0.296
	DFA α2 IQR	0.204	0.120	0.0001	0.357	−0.287
	SDNN IQR	27.385	18.263	0.0002	0.362	−0.276
	pNN50 IQR	4.919	2.185	0.0002	0.362	−0.275
	SD2 IQR	38.119	22.141	0.0005	0.367	−0.265
	SD1 IQR	13.852	8.268	0.0007	0.372	−0.256
	RMSSD IQR	19.590	11.692	0.0007	0.372	−0.256
	AVNN IQR	57.743	32.086	0.0008	0.373	−0.253
	IHR IQR	4.652	2.835	0.0009	0.375	−0.251
	SD1/SD2 IQR	0.840	0.520	0.0034	0.385	−0.229
	HF IQR	445.802	172.727	0.0048	0.389	−0.222
	LF IQR	499.178	278.598	0.0061	0.392	−0.216
	DFA α1 IQR	0.160	0.106	0.0064	0.393	−0.213
	DFA α2 median	1.082	1.165	0.0079	0.604	0.208
	Total Power IQR	4427.589	2599.252	0.0091	0.398	−0.205
	VLF median	4098.166	5603.812	0.0100	0.601	0.203
	LF/HF IQR	0.749	0.382	0.0186	0.407	−0.187
	SD1/SD2 median	3.023	3.533	0.0310	0.588	0.177

Only HRV features showing statistically significant differences between paradoxical insomnia (PI) and normal sleep (N) after Benjamini–Hochberg false discovery rate (FDR) correction (q < 0.05) are presented in this table; features without significant group differences were omitted for clarity. IQR values represent within-subject interquartile ranges (75th–25th percentile) calculated across all valid 5 min windows within the corresponding sleep stage.

). The majority of these differences reflected greater within-stage dispersion of HRV features in PI, quantified by larger within-subject interquartile ranges calculated across valid five-minute windows during the WASO stage.

Among time-domain variability indices, PI exhibited substantially greater dispersion in several short-term variability measures. In particular, SDNN IQR was significantly higher in PI than in N (27.39 vs. 18.26 ms; q = 0.0002), as was RMSSD IQR (19.59 vs. 11.69 ms; q = 0.0007). Similar patterns were observed in related variability measures, including pNN50 IQR, SD1 IQR, SD2 IQR, AVNN IQR, and IHR IQR. These findings indicate increased variability of beat-to-beat autonomic fluctuations during WASO in PI.

Frequency-domain indices showed a comparable pattern. PI demonstrated larger dispersion in HF power IQR (445.80 vs. 172.73 ms²; q = 0.0048) and LF power IQR (499.18 vs. 278.60 ms²; q = 0.0061), as well as greater dispersion in total power and the LF/HF ratio. These results indicate modestly greater variability in spectral components of autonomic modulation during WASO in PI.

Non-linear HRV metrics further reinforced this pattern. Complexity-related indices, including sample entropy (SampEn IQR: 0.276 vs. 0.127; q < 0.0001) and approximate entropy (ApEn IQR: 0.163 vs. 0.087; q = 0.0001), were significantly larger in PI, indicating greater irregularity of heart rate dynamics. Fractal scaling indices derived from detrended fluctuation analysis (DFA α₁ and DFA α₂) also showed greater dispersion in PI.

In addition to dispersion-related measures, several central tendency indices differed between groups. PI exhibited lower DFA α₂ median values and reduced SD1/SD2 ratios compared with N, suggesting subtle alterations in long-term autonomic regulation during WASO.

Taken together, these results indicate that autonomic differentiation between PI and N was not evident during consolidated sleep stages but emerged selectively during WASO. During this stage, PI exhibited a consistent tendency toward greater variability, dispersion, and complexity of HRV dynamics. However, the overall magnitude of separation remained modest, with substantial overlap in HRV distributions between the two groups. These unadjusted WASO findings should be interpreted together with the window-count analysis presented in Section 3.2.4, which showed that the dispersion-based effects were substantially influenced by variation in the number of valid WASO windows.

3.2.2. Stage-Specific HRV Differences Between Paradoxical Insomnia and Objective Insomnia

Stage-specific comparisons between PI and OI revealed minimal differences during consolidated sleep. In both NREM and REM sleep, HRV features showed substantial overlap between the two insomnia subtypes, and no HRV metric demonstrated statistically significant after FDR correction.

In contrast, significant between-group differences emerged during WASO (

Table 4. Stage-specific HRV differences between paradoxical insomnia (PI) and objective insomnia (OI) during wake after sleep onset (WASO).

Sleep Stage	HRV Features	OI (Median)	PI (Median)	q-Value	CLES	Rank-Biserial r
WASO	DFA α1 IQR	0.247	0.160	0.0001	0.357	−0.286
	DFA α2 IQR	0.275	0.204	0.0001	0.359	−0.281
	SD1/SD2 IQR	1.228	0.840	0.0058	0.396	−0.209
	SampEn IQR	0.359	0.276	0.0064	0.397	−0.205
	SD2 IQR	49.357	38.119	0.0071	0.399	−0.202
	VLF median	2628.257	4098.166	0.0120	0.595	0.190
	SDNN IQR	36.906	27.385	0.0142	0.407	−0.186
	ApEn IQR	0.199	0.163	0.0143	0.407	−0.185
	RMSSD IQR	27.836	19.590	0.0143	0.408	−0.184
	SD1 IQR	19.686	13.852	0.0143	0.408	−0.184

Only HRV features showing statistically significant differences between paradoxical insomnia (PI) and objective insomnia (OI) after Benjamini–Hochberg false discovery rate (FDR) correction (q < 0.05) are presented in this table; features without significant group differences were omitted for clarity. IQR values represent within-subject interquartile ranges (75th–25th percentile) calculated across all valid 5 min windows within the corresponding sleep stage.

). After FDR correction, 10 HRV features differed significantly between PI and OI. Most of these were dispersion-based measures, indicating broader within-stage variability of HRV dynamics in OI than in PI.

Several IQR-based indices were higher in OI, including DFA α₁ IQR (0.247 vs. 0.160; q = 0.0001), DFA α₂ IQR (0.275 vs. 0.204; q = 0.0001), SDNN IQR (36.91 vs. 27.39 ms; q = 0.0142), and RMSSD IQR (27.84 vs. 19.59 ms; q = 0.0143). Similar patterns were observed in Poincaré-derived measures (SD1/SD2 IQR, SD1 IQR, and SD2 IQR) and entropy-based measures, including SampEn IQR (0.359 vs. 0.276; q = 0.0064) and ApEn IQR (0.199 vs. 0.163; q = 0.0143).

In contrast to these dispersion-related findings, one central tendency measure was higher in PI: median VLF power during WASO was greater in PI than in OI (4098.17 vs. 2628.26 ms²; q = 0.0120).

Taken together, these results indicate that, before adjustment for WASO window count, autonomic differences between PI and OI were most evident during WASO rather than during consolidated sleep. PI showed relatively higher central tendency values for certain spectral components, whereas OI demonstrated greater dispersion across several variability and complexity-related HRV measures. However, the overall magnitude of separation remained modest, with substantial overlap in HRV distributions between the two groups. These unadjusted findings should be interpreted alongside the covariate-adjusted results in Section 3.2.4, in which the dispersion-based group differences were attenuated after accounting for WASO window count.

3.2.3. Sensitivity Analysis Using the Full Dataset

To evaluate whether the balanced subsampling strategy influenced the primary findings, additional analyses were performed using the full available dataset without random subsampling. The overall pattern of results was consistent with that observed in the balanced subsample. In both phenotype contrasts, FDR-significant HRV differences remained confined to the WASO stage, whereas no significant features were identified during NREM or REM sleep in either analysis.

For the PI vs. N comparison, 18 of the 19 HRV features that were significant in the balanced subsample were also significant in the full dataset, with fully concordant effect directions (Supplementary Table S1). For the PI vs. OI comparison, all features identified in the balanced subsample remained significant in the full dataset, and additional features reached significance with increased statistical power (Supplementary Table S2). Effect size estimates were also highly consistent between the two analyses, with rank-biserial correlation coefficients exceeding 0.95 across comparisons.

These findings indicate that the principal stage-specific HRV patterns observed in this study were robust to the sampling strategy and were not driven by balanced subsample selection.

3.2.4. Influence of WASO Window Count on Dispersion Metrics

Because the significant WASO findings were predominantly based on IQR-derived dispersion measures, additional analyses were performed to examine their relationship with WASO window count. In the pooled sample, all IQR-based WASO HRV features showed significant positive correlations with the number of valid WASO windows after FDR correction. Similar positive associations were also observed within phenotype groups, indicating that larger dispersion values were generally associated with a greater number of analyzable WASO segments.

To determine whether the previously observed between-group differences were independent of this effect, covariate-adjusted regression analyses were performed with WASO window count included as a covariate. After adjustment, none of the previously significant IQR-based differences between PI and N or between PI and OI remained significant after FDR correction.

These findings indicate that the unadjusted WASO dispersion results were substantially influenced by variation in WASO window count and should therefore be interpreted cautiously when considered against the primary univariate group differences reported in Section 3.2.1 and Section 3.2.2.

3.3. Multivariate Pattern Separation Using UMAP

To explore multivariate autonomic patterns across sleep stages, UMAP was applied as an exploratory visualization tool using standardized HRV feature sets. During NREM and REM sleep, UMAP projections revealed substantial overlap between PI, OI, and N groups, indicating weak overall multivariate differentiation during consolidated sleep stages (Figure 3, Figure 4, Figure 5 and Figure 6).

In contrast, during WASO, embedded distributions showed modest shifts in the spatial location of PI samples relative to comparison groups, although group distributions remained largely overlapping (Figure 7 and Figure 8).

To quantify stage-dependent multivariate differences in HRV patterns, centroid distances between phenotype groups were computed in the two-dimensional UMAP-embedded space. To evaluate whether the observed separation exceeded chance-level expectations, a permutation-based centroid distance test (1000 permutations) was performed (

Table 5. Stage-specific centroid distances and permutation-based statistical testing of insomnia phenotype separation in the UMAP-embedded HRV feature space.

Sleep Stage	Group	Centroid Distance	Permutation p	z-Score
NREM	PI vs. N	0.136	0.705	−0.63
	PI vs. OI	0.258	0.275	0.49
REM	PI vs. N	0.310	0.162	0.98
	PI vs. OI	0.112	0.787	−0.85
WASO	PI vs. N	0.876	0.001	6.01
	PI vs. OI	0.596	0.003	3.57

).

Across consolidated sleep stages (NREM and REM), centroid distances between groups were small and did not differ significantly from the permutation-based null distribution (all p > 0.16), indicating substantial overlap of multivariate HRV patterns among insomnia phenotypes during stable sleep.

In contrast, larger centroid separations were observed during WASO. The centroid distance between PI and N reached 0.876 and was significantly greater than expected under random label permutations (p = 0.001, z = 6.01). Similarly, the centroid distance between PI and OI during WASO was 0.596 and also exceeded the permutation-based null distribution (p = 0.003, z = 3.57).

These findings indicate that overall cluster separation in the UMAP space is modest, with substantial overlap remaining between groups. Within this modest overall separation, a statistically significant centroid displacement was detectable during WASO but not during NREM or REM. This stage-dependent multivariate pattern is consistent with the broader result that between-group HRV differences were most detectable during WASO.

3.4. Machine Learning Classification Performance

Supervised machine learning analysis was conducted to evaluate the discriminability of PI from N and OI based on stage-specific HRV features. Random Forest classifiers were trained and evaluated separately for each sleep stage (NREM, REM, and WASO) to assess the stage dependency of autonomic information for phenotype classification. To reduce the dimensionality of the feature space and limit potential overfitting, only HRV features that remained statistically significant after FDR correction in the univariate analysis were used as model inputs.

Across both classification tasks (PI vs. N and PI vs. OI), model performance varied substantially by sleep stage (

Table 6. Random forest classification performance for paradoxical insomnia (PI) vs. normal sleep (N) using stage-specific HRV features selected after FDR correction.

Sleep Stage	Group	Precision	Recall	F1-Score	Accuracy	10-Fold CV AP
NREM	PI	0.509	0.467	0.487	0.508	0.592 ± 0.082
	N	0.508	0.550	0.528
REM	PI	0.500	0.467	0.483	0.500	0.555 ± 0.060
	N	0.500	0.533	0.516
WASO	PI	0.627	0.750	0.683	0.629	0.720 ± 0.074
	N	0.632	0.490	0.552

and

Table 7. Random forest classification performance for paradoxical insomnia (PI) vs. objective insomnia (OI) using stage-specific HRV features selected after FDR correction.

Sleep Stage	Group	Precision	Recall	F1-Score	Accuracy	10-Fold CV AP
NREM	PI	0.542	0.533	0.538	0.542	0.534 ± 0.083
	OI	0.541	0.550	0.545
REM	PI	0.486	0.600	0.537	0.483	0.573 ± 0.086
	OI	0.478	0.367	0.415
WASO	PI	0.619	0.464	0.531	0.60	0.634 ± 0.073
	OI	0.589	0.729	0.652

). Classifiers trained on HRV features derived from NREM and REM sleep demonstrated limited discriminative ability, with accuracy and F1-scores remaining close to chance level, indicating minimal classification value from autonomic features during consolidated sleep stages.

In contrast, models trained on WASO-derived HRV features showed the highest overall performance across both classification tasks. For PI vs. N, the WASO-based Random Forest model achieved an accuracy of 0.629, with F1-scores of 0.683 for PI and 0.552 for N. For PI vs. OI, the WASO-based model also showed relatively better performance than NREM- and REM-based models, although the overall magnitude of discrimination remained modest.

Supplementary baseline comparison further showed that whole-night models provided limited discrimination, whereas performance was generally highest during WASO, while the difference between logistic regression and Random Forest remained relatively small (Supplementary Tables S3 and S4).

Taken together, these findings suggest that the main source of improvement lies in stage-specific localization of autonomic information to WASO rather than in model complexity alone. Accordingly, the present machine learning results should be interpreted as proof-of-concept evidence of stage-dependent physiological information content rather than as a clinically applicable diagnostic classifier.

4. Discussion

4.1. Principal Findings

In this study, we investigated sleep-stage–resolved HRV characteristics of PI in a large population-based cohort. The principal finding is that group differences were not evident as a sustained whole-night pattern but were most detectable during WASO. Whole-night HRV averages and features derived from consolidated sleep stages (NREM and REM) showed substantial overlap among PI, OI, and N groups, indicating limited separation during stable sleep.

In contrast, WASO emerged as the stage in which HRV-related group differentiation was most consistently detectable across the univariate, multivariate, and machine learning analyses. At the same time, the magnitude of these differences remained generally small to moderate, with considerable overlap in HRV feature distributions across groups. Accordingly, the present findings should not be interpreted as evidence of marked physiological separation between insomnia phenotypes, but rather as support for a subtle and state-dependent pattern of autonomic differentiation that becomes relatively more observable during nocturnal awakenings.

4.2. Autonomic Characteristics of Paradoxical Insomnia During Wake After Sleep Onset

WASO is a transitional state marked by repeated shifts between sleep and wakefulness, during which cortical arousal systems and cardiovascular regulation must be repeatedly re-engaged. From this perspective, it is plausible that subtle phenotype-related physiological differences would be more detectable during WASO than during consolidated sleep. The present findings support this stage-specific view by showing that WASO, rather than whole-night or stable sleep periods, was the condition in which HRV-related differentiation was most consistently detectable.

The present findings suggest that the physiological signal associated with PI is not uniformly distributed across the night, but is more detectable during WASO, when sleep continuity is repeatedly interrupted and sleep–wake regulatory systems must be re-engaged. This pattern does not necessarily indicate a large or persistent autonomic abnormality unique to PI; rather, it highlights the importance of transitional sleep–wake states as a context in which subtle phenotype-related differences become more visible. Viewed in this way, the study’s main contribution is to show that stage-resolved analysis, particularly during WASO, may offer greater physiological relevance for phenotyping than whole-night summary measures alone.

At the same time, the present results indicate that dispersion-based WASO metrics are influenced by WASO window count and therefore should not be interpreted in isolation as direct markers of PI-specific autonomic instability. Instead, they are more appropriately interpreted as markers of stage-dependent autonomic differentiation that likely reflect a combination of HRV variability and fragmentation-related sleep structure. Even with this limitation, the consistent concentration of detectable signal in WASO remains meaningful, as it highlights the potential value of stage-aware physiological analysis for characterizing heterogeneity across insomnia phenotypes.

4.3. Implications for Stage-Aware Phenotyping and Digital Biomarker Development

The present findings have important implications for objective phenotyping of insomnia subtypes and the development of physiology-based digital biomarkers. A key contribution of this study is the demonstration that HRV-related differences associated with insomnia phenotypes are more detectable in a stage-resolved framework than in whole-night summary measures alone. In particular, despite only modest overall multivariate separation, WASO emerged as the sleep–wake state in which subtle between-group differences were most consistently identified across complementary analyses. This observation highlights a fundamental limitation of conventional whole-night summary metrics and underscores the need for temporally resolved analytical approaches to capture transient but potentially informative physiological variation.

Within this framework, the machine learning analysis was not intended to produce a clinically deployable diagnostic classifier, but rather to serve as a computational probe for evaluating the relative information content of stage-specific autonomic features. Although models based on WASO-related HRV features showed relatively better performance than those based on NREM or REM features, overall classification performance remained modest. Specifically, an accuracy of approximately 0.60–0.65 and an F1-score below 0.70, while exceeding chance level, are generally considered insufficient for standalone clinical screening or diagnostic applications, which typically require substantially higher sensitivity and specificity to ensure reliable individual-level decision making. Accordingly, the present machine learning results should be interpreted as evidence of relative stage-specific information content rather than as evidence of immediate clinical utility.

Supplementary baseline comparison further supported this interpretation by showing that models based on whole-night HRV summaries provided limited discrimination, whereas classification performance was consistently highest during WASO across both PI vs. N and PI vs. OI comparisons (Supplementary Tables S3 and S4). At the same time, the performance gap between logistic regression and Random Forest remained relatively small, indicating that the principal source of incremental predictive value was not increased classifier complexity but temporal localization of autonomic information to the sleep–wake transition period. This pattern reinforces the view that WASO constitutes a physiologically informative window for phenotype differentiation, even though overall discriminability remains modest. It should not, however, be taken as evidence that the dispersion-based WASO features independently represent PI-specific autonomic dysregulation.

From a technological perspective, these results support the value of stage-aware computational analysis pipelines for physiological signal interpretation. By integrating sleep staging with short-term HRV windowing and multilevel analysis, this study provides a scalable framework for examining where subtle physiological differences are most detectable in large-scale biosignal datasets. At the same time, the present findings indicate that dispersion-based WASO metrics are influenced by WASO structure and therefore should not be interpreted in isolation as standalone biomarkers. Rather, stage-resolved HRV features may be most useful as one component of future multimodal digital phenotyping systems that integrate autonomic, behavioral, and central nervous system measures.

4.4. Limitations and Future Directions

Several limitations of the present study should be acknowledged. First, the analysis was based on cross-sectional polysomnography and electrocardiographic recordings obtained from a single-night assessment, which precludes evaluation of intra-individual variability and the longitudinal stability of the observed autonomic patterns. Although the large sample size enhances statistical robustness, single-night measurements may not fully capture night-to-night fluctuations in sleep–wake dynamics, particularly in individuals with insomnia.

Second, the present findings were derived from the Sleep Heart Health Study, which predominantly consists of middle-aged and older adults. Given that PI is also prevalent in younger populations and may manifest differently across the lifespan, caution is warranted when generalizing these results to younger individuals or clinical samples with different demographic and psychosocial characteristics. Future studies involving age-diverse cohorts and clinically characterized insomnia samples will be necessary to determine the broader generalizability of the stage-specific patterns observed here.

Third, the primary WASO-related differences in the present study were largely based on IQR-derived dispersion metrics. Additional analyses showed that these measures were strongly associated with the number of valid WASO windows and were attenuated after adjustment for WASO window count. Accordingly, these dispersion-based findings should not be interpreted in isolation as direct evidence of PI-specific autonomic instability. Future studies should consider alternative within-stage variability metrics that are less sensitive to sampling density, or explicitly model WASO duration, fragmentation, and window count alongside HRV dynamics.

Fourth, the analysis focused exclusively on heart rate variability as an index of autonomic regulation. While HRV provides a valuable non-invasive window into peripheral autonomic dynamics, it does not directly capture central nervous system activity or subjective sleep experience. HRV alone is unlikely to fully characterize the complex mechanisms underlying sleep state misperception. Future work should adopt multimodal phenotyping approaches that integrate HRV with electroencephalography-derived markers of cortical arousal and instability, as well as subjective and behavioral measures such as ecological momentary assessment of sleep perception and nocturnal awakenings.

Finally, although the present analyses were conducted using in-home polysomnography rather than wearable sensing, extension of this stage-resolved framework to ambulatory and wearable monitoring environments will be important for evaluating feasibility and translational potential in real-world settings. Despite these limitations, the present study highlights the value of stage-resolved analysis for identifying where subtle physiological differences are most detectable and supports further development of multimodal, longitudinal, and stage-aware approaches for characterizing heterogeneity across insomnia phenotypes.

5. Conclusions

In conclusion, this study shows that HRV-related differences among insomnia phenotypes were not evident as a broad whole-night pattern, but were most detectable during WASO. By applying stage-resolved heart rate variability analysis in a large population-based cohort, we identified WASO as the sleep–wake state in which subtle between-group differences were most consistently observed, whereas whole-night, NREM, and REM measures showed substantial overlap.

At the same time, these differences were modest in magnitude, and dispersion-based WASO findings were influenced by WASO window count. Thus, the results should be interpreted cautiously and not as evidence of large physiological separation or autonomous proof of PI-specific autonomic instability.

From a technological perspective, the present findings support the value of stage-aware analysis for physiological phenotyping and suggest that stage-resolved HRV may serve as a useful component of future multimodal digital biomarker frameworks for insomnia heterogeneity.