HRV features as potential biomarkers for auxiliary diagnosis in epilepsy

Abstract

Epilepsy affects around 70 million people worldwide, and diagnosis is often difficult and delayed, exposing patients to avoidable morbidity and psychosocial burden. Heart rate variability (HRV) is a non-invasive marker of autonomic nervous system function that may be altered in epilepsy and may support clinical decision-making. In this single-center case–control study, we recorded short-term HRV during a standardized cardiovascular autonomic reflex test including supine resting, deep-breathing and three challenges (active standing, Valsalva manoeuvre and sustained handgrip) in 200 adults with epilepsy and 200 age- and sex-matched healthy controls. Patients with epilepsy showed consistently lower HRV than controls. Using HRV and demographic features, we developed logistic regression models to distinguish epilepsy from health in an independent test set. A model integrating rest and sustained handgrip achieved the highest performance, although still only moderate (area under the curve 0.68; sensitivity 0.821; specificity 0.484). Standardized multi-paradigm HRV assessment may therefore provide a feasible, low-cost adjunct to support, but not replace, conventional diagnostic evaluation. However, the single-center design, relatively short recordings and inclusion of only healthy controls limit generalizability, and larger multicenter studies including patients with paroxysmal conditions that mimic epilepsy are needed to determine clinical utility.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-34682-0.

Introduction

Epilepsy is a chronic neurological disorder caused by abnormal electrical discharges of brain neurons, affecting approximately 70 million people worldwide^1,2. Despite significant medical advances, diagnosing epilepsy remains challenging due to its unpredictable nature. While advanced tools like electroencephalography (EEG), neuroimaging, and genetic testing can improve diagnostic accuracy, they are often limited by high costs and low detection rates, especially in underdeveloped regions. These limitations frequently cause diagnostic delays, increasing the burden on both patients and society^3,4. Indeed, the challenge of rapid and accurate diagnosis is common across neurological disorders, particularly in primary care settings. For instance, in ischemic stroke, reliance on physical examinations by non-specialists can lead to lower diagnostic accuracy⁵, creating a critical need for objective decision-support tools.

A critical aspect of epilepsy is the presence of underlying autonomic nervous system (ANS) dysfunction, a key mechanism linked to severe outcomes, including an elevated risk of cardiovascular complications and sudden unexpected death in epilepsy (SUDEP). This dysfunction highlights the need for tools that can assess autonomic function objectively, enabling more accurate diagnosis and timely intervention. While current research often focuses on EEG-based seizure diagnosis^6,7, recent advancements in IoT and wearable devices have spurred interest in multimodal approaches. Among these, heart rate variability (HRV) monitoring is emerging as a promising non-invasive tool for the diagnostic assessment of this crucial autonomic function in epilepsy.

HRV analysis is a reliable, non-invasive method for assessing autonomic nervous system (ANS) function, which is often impaired in epilepsy patients^8,9. While higher HRV in healthy populations indicates good autonomic function, lower HRV is a known marker for various disease states, including cardiovascular disease and depression^10,11. In epilepsy, HRV alterations reflect the complex brain-heart interaction, where abnormal neuronal discharges affect cardiac activity via the ANS^12,13. However, a validated diagnostic model for epilepsy based on these non-invasive autonomic markers is currently lacking, highlighting a crucial gap in developing objective assessment tools for the disorder.

Previous research has consistently shown that epilepsy patients exhibit altered HRV patterns, such as reduced parasympathetic activity and altered sympathovagal balance^14,15, with changes detectable during both ictal and interictal periods^16,17. These findings highlight HRV’s potential for both early diagnosis and continuous monitoring. Despite these findings, clinical translation has been limited by inconsistent methodologies, such as non-standardized or single-paradigm assessments like steady-state recordings¹⁸, and the lack of large-scale validation and comprehensive control groups, which limits the generalizability and specificity of results^19–21. These challenges underscore the need for a rigorous, multidimensional approach to HRV assessment in epilepsy.

To address these limitations, this study was designed to develop and internally validate a standardized, multi-paradigm HRV assessment for the auxiliary diagnosis of epilepsy. Our primary objective was to integrate HRV parameters across multiple standardized autonomic challenges into an internally validated, non-invasive diagnostic model, thereby exploring a potential objective biomarker candidate that could modestly improve diagnostic assessment.

Methods

Participants: inclusion and exclusion criteria

Between November 2023 and August 2024, we initially recruited 493 consecutive adults with epilepsy from the Department of Neurology at West China Hospital of Sichuan University and 213 community-recruited healthy controls. The inclusion criterion for all participants was an age range of 18–65 years. For patients, an additional inclusion criterion was a diagnosis of epilepsy confirmed by two or more specialists according to the International League Against Epilepsy (ILAE) diagnostic criteria (2014)²² and classification standards (2017)²³ The following exclusion criteria were applied to all participants: (1) cognitive impairment preventing study cooperation; (2) hypertension, arrhythmia, diabetes, or other cardiac diseases; and (3) progressive brain diseases, psychiatric disorders, or other major illnesses. All subjects provided written informed consent, and the study was approved by the Biomedical Ethics Committee of West China Hospital of Sichuan University (Approval No.: NO.2023–2394).

For HRV analysis, the same quality-control criteria were applied to both groups. Of the 493 PWE, 17 did not complete all autonomic testing tasks and 8 withdrew consent before HRV recording, leaving 468 patients who underwent HRV testing. Recordings were excluded if they were uninterpretable because of excessive background noise or signal artefacts. Under these criteria, 32 of the 468 PWE recordings and 13 of the 213 control recordings were excluded, resulting in 436 patients with epilepsy and 200 healthy controls with complete and analyzable HRV data. Next, we constructed an age- and sex-matched sample by randomly selecting 200 PWE from the 436 eligible patients and matching them 1:1 with the 200 healthy controls, yielding a final analysis sample of 400 participants. Within this matched sample, analyzable 5-minute HRV segments were available for all four autonomic testing paradigms (supine rest, active standing, Valsalva and sustained handgrip).

Demographic and clinical data collection

For all participants, we recorded key demographic and anthropometric data, including sex, age, height, weight, Body Mass Index (BMI), and waist circumference. For the epilepsy group, detailed clinical information was collected from patient medical records and structured interviews. This included seizure frequency over the past year, clinically confirmed seizure type, current antiseizure medication regimen, and relevant neuroimaging findings.

In patients with epilepsy, we additionally recorded potential autonomic confounders that might influence HRV. These included antiseizure medication (ASM) status (no medication, monotherapy, or polytherapy), smoking history (yes/no), alcohol consumption (yes/no), habitual coffee intake (yes/no), the presence of sleep-related seizures (awake-only, sleep-only, or both awake and sleep), and comorbid anxiety and/or depression. These variables were collected systematically in the epilepsy group, whereas corresponding information was not available for healthy controls.

HRV assessment

HRV recording

This study employed a standardized protocol to acquire ECG data (250 Hz). After a 30-minute quiet rest in a supine position for heart rate stabilization, a 5-minute baseline ECG was recorded (Supine Resting Test). Immediately following this, participants underwent four standardized Cardiovascular Autonomic Reflex Test (CART) manoeuvres in a fixed sequence: (1) Active standing test: after supine rest, participants stood up actively and maintained standing for ≥ 5 min; ECG was recorded continuously. (2) Deep breathing test: participants performed six breaths per minute (0.1 Hz) for 1 min under metronome guidance, followed by 2 min of recovery; ECG and respiratory signals were recorded. (3) Valsalva manoeuvre: participants performed three 15-second expiratory strains at 40 mmHg using a mouthpiece with manometer (with a small leak to avoid glottic closure), each separated by 1-minute rest, followed by 3 min of recovery after the final strain; ECG was recorded continuously. (4) Sustained handgrip test: each participant’s maximal voluntary contraction (MVC) was determined using a digital dynamometer. Participants then performed a continuous isometric handgrip at 30% MVC for 5 min, with real-time visual feedback to maintain target force; ECG was recorded continuously²⁴.

Each active manoeuvre was designed to last a minimum of 5 min (including rest/recovery periods where applicable) to ensure sufficient data for reliable analysis per HRV Task Force guidelines²⁵. All recordings were performed using a cardiovascular autonomic nervous function multi-parameter evaluation system (Model R6000, REEM (Shenzhen) Healthcare Co., Ltd.), which integrated a wireless ECG patch (Fig. 1), a respiratory flow sensor, and a digital dynamometer.

Fig. 1.

Schematic diagram of ECG patch.

The deep-breathing manoeuvre in our protocol lasted 1 min, whereas the other autonomic paradigms and baseline rest were analysed using 5-minute segments. Because reliable frequency-domain HRV estimation, particularly in the low-frequency range, requires several minutes of data, the deep-breathing epoch was not analysed as a separate condition. However, this 1-minute segment was retained as part of the full autonomic test period so that its contribution to autonomic activation was reflected in the whole-period HRV indices. Accordingly, single-condition analyses considered four paradigms—supine rest, active standing, Valsalva and sustained handgrip—while the “entire test period” summary included all epochs, including deep breathing.

All HRV recordings were obtained during clinically interictal assessments. For patients with epilepsy, the interval between the last clinically evident seizure and the HRV testing was extracted from patient reports and medical records and categorized as ≤ 48 h, 48 h–7 days, or > 7 days^26–28.

Signal processing and HRV indices

The R-wave detection was performed using the Pan-Tompkins method²⁹(26), with an adaptive artifact rejection algorithm³⁰(27). The resultant RR interval series was resampled at 4 Hz and detrended. A comprehensive panel of 17 standard HRV indices was calculated from the time-domain, frequency-domain, and non-linear domains for each test condition. To comprehensively assess autonomic modulation, a panel of standard HRV indices was calculated for each of the four test conditions and for the entire recording period. These indices were selected based on their established physiological interpretations^25,31,32. Overall HRV: SDNN, Triangular Index, avgRR, Total Power. Parasympathetic Modulation: RMSSD, SD1, HF power, NN50, pNN50, DC. Markers Influenced by Sympathetic Activity: LF power (under stress conditions), AC, SD2. Index of Sympathovagal Balance: LF/HF ratio. Non-linear/Complexity Measures: α1, α2. Other: VLF power.

Rationale for multidimensional and integrated analysis

The inclusion of parameters from the entire CARTs battery, both individually and as a single integrated sample, enables a robust, multidimensional assessment of autonomic regulatory capacity. This event-related approach aims to enhance statistical power by aggregating responses to sequential autonomic challenges, thereby increasing sensitivity to detect subtle but consistent pathological patterns that may be non-significant in isolated tests.

Statistical analysis

Data analysis and visualization were performed using R. All continuous variables were assessed for normality using Shapiro-Wilk tests. The majority of variables violated the normality assumption (Shapiro-Wilk p < 0.05 for > 80% of distributions), justifying the use of nonparametric Mann-Whitney U tests, and parameters were expressed as median (interquartile range). To account for multiple comparisons, raw p-values were adjusted using the Benjamini-Hochberg false discovery rate (FDR) procedure, with FDR-adjusted p-values < 0.05 considered statistically significant. For exploratory subgroup analyses, only unadjusted P values were reported and no correction for multiple testing was applied.

Autonomic confounders not included in Table 1 (lifestyle factors, comorbid anxiety/depression) and seizure-to-testing intervals in the epilepsy group were summarized descriptively as counts and percentages in Supplementary Table S1.

Table 1.

Demographic and clinical characteristics of the study Population.

	PWE group (N = 200)	HC group (N = 200)	p-value
Demographic characteristics
Sex			1
Female	121 (60.5)	121 (60.5)
Male	79 (39.5)	79 (39.5)
Age (year), Median (IQR)	33.2 (25.2, 41.2)	33.2 (24.6, 41.8)	0.737
Weight (Kg), Median (IQR)	61.5 (53.0–70.0)	60.9 (53.3–68.5)	0.737
BMI, Median (IQR)	22.8 (20.3, 25.3)	22.6 (20.3, 24.9)	0.549
Height (cm), Median (IQR)	163.8 (158.1, 169.5)	164.0 (158.4, 169.6)	0.818
Body surface area (m2), Median (IQR)	1.6 (1.5, 1.7)	1.6 (1.5, 1.7)	0.739
Waist circumference (cm), Median (IQR)	76.9 (69.4, 84.4)	74.2 (67.1, 81.3)	0.767
Epilepsy characteristics
Drug-resistant epilepsy
Yes	73 (36.5)		N.A.
No	127 (63.5)		N.A.
Epilepsy type, n (%)
Generalized	47 (23.5)		N.A.
Focal	125 (62.5)		N.A.
Unknown	28 (14.0)		N.A.
Duration (year), Median (IQR)	11.1 (4.4, 17.8)		N.A.
Severity (NHS3)	6.0 (3.0, 9.0)		N.A.
Seizure frequency category, n(%)
<1/Month	71 (35.5)		N.A.
≥ 1/Month	129 (64.5)		N.A.
Seizure timing
Awake	105(52.5%)		N.A.
Sleep	19(9.5%)		N.A.
Both	76(38.0%)		N.A.
Antiseizure medications use
Monotherapy, n (%)	75 (37.5)		N.A.
Polytherapy, n (%)	108 (54.0)		N.A.
ASM-free, n (%)	17 (8.5)		N.A.

Diagnostic model development

For diagnostic modelling, the full dataset was randomly split into a training set (70%) and a held-out test set (30%) with stratification by group (epilepsy vs. controls). All model specification, feature selection and hyperparameter tuning were performed exclusively in the training set; the test set was used once for final performance evaluation.

We used logistic regression models as the primary modelling framework. As sensitivity analyses, we evaluated L1 (LASSO) and L2 (ridge) penalized regression and random forest classifiers to assess whether alternative approaches might improve discrimination. For penalized models, the regularization parameter (λ) was selected by K-fold cross-validation (K = 5) in the training set, using cross-validated AUC as the model selection criterion. Random forest used default hyperparameters. The final penalized model was refitted on the full training set, and all models were evaluated in the held-out test set, with the operating point chosen by the Youden index. Nested cross-validation results and model comparisons are reported in Supplementary Tables S2 and S3. Because all model specifications yielded similar performance, we report results from the standard logistic regression model in the main text.

To obtain split-independent estimates of model performance, we additionally carried out nested cross-validation on the full sample. In each of K outer folds (K = 5), a penalized logistic regression model was tuned in the corresponding training portion using an inner K-fold cross-validation loop to select the optimal λ, and evaluated in the held-out outer fold. The mean AUC and its standard deviation across outer folds were used as internally validated estimates of discrimination for each autonomic paradigm and for the primary combined model (supine rest + sustained handgrip). These results are summarized in Supplementary Table S2.

From the 17 initially calculated HRV parameters, a final subset of 13 was selected for modelling based on physiological interpretability and reliability in short-term recordings. VLF power and the LF/HF ratio were excluded because their physiological meaning is debated and they are strongly influenced by very-low-frequency components that are not well captured in 5-minute segments^25,33. The non-linear parameters DFA α1 and α2 were excluded because their stability and test–retest reliability are reduced in short-term recordings^34,35. These four indices are therefore reported descriptively but were not entered into the diagnostic models, to focus on HRV features with more established and interpretable properties under our short-term recording conditions.

The final models included demographic and anthropometric characteristics (age, sex, waist circumference, BMI) and the 13 selected HRV parameters as independent variables. Separate models were constructed for each of the five paradigms (supine rest, active standing, valsalva, sustained handgrip, and the entire test period), as well as combined models that integrated resting-state data with each stimulation paradigm. Model performance was primarily evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC), along with accuracy, sensitivity, and specificity. To assess the contribution of each feature, feature importance was ranked based on the standardized coefficients of the primary (unpenalized) logistic regression models.

Subgroup analyses were conducted to explore whether HRV abnormalities and model performance differed across clinically relevant strata within the epilepsy group. Patients with epilepsy were classified along three binary dimensions: (1) seizure type (focal vs. generalized epilepsy; patients with unknown seizure type were excluded from this comparison), (2) treatment response (drug-resistant vs. controlled epilepsy) and (3) sleep involvement (wake-only vs. sleep-related seizures. For each dimension, HRV indices were compared between subgroups using Mann–Whitney U-tests, and results are reported as median (interquartile range) with unadjusted P values (Supplementary Table S4). In addition, the LASSO-regularized version of the primary combined HRV model (supine rest + sustained handgrip) was applied without re-fitting to each subgroup in the held-out test set, and subgroup-specific AUCs, sensitivities and specificities were calculated (numerical results shown in Supplementary Table S5, with corresponding ROC curves illustrated in Supplementary Figure S1). These subgroup analyses were exploratory in nature and were not adjusted for multiple comparisons.

Results

Table 1 summarizes the demographic and clinical characteristics of the study participants. A total of 200 people with epilepsy (PWE) and 200 healthy controls (HC) were enrolled. The groups were successfully matched, with identical sex distributions (60.5% female) and no significant differences in any baseline demographic or anthropometric measurements, including age, weight, height, BMI, body surface area, and waist circumference (all p > 0.05).

Among PWE, focal epilepsy was the predominant type (62.5%), followed by generalized (23.5%) and unknown types (14.0%). The median epilepsy duration was 11.1 years (IQR: 4.4–17.8), with a median severity score (NHS3) of 6.0 (IQR: 3.0–9.0). Drug-resistant epilepsy was present in 36.5% of patients. Most patients (64.5%) experienced at least one seizure per month, with seizures occurring exclusively during wakefulness in 52.5%, exclusively during sleep in 9.5%, and in both states in 38.0% of patients. Regarding treatment, 37.5% received monotherapy, 54.0% received polytherapy, and 8.5% were not on antiseizure medications.

Distributions of additional autonomic confounders and seizure-to-testing intervals in the epilepsy group are summarized in Supplementary Table S1. There was broad heterogeneity in lifestyle factors (smoking, alcohol and coffee consumption), comorbid anxiety/depression and the timing of HRV assessment relative to the last clinical seizure; in particular, 8% of patients were assessed within 48 h of their last seizure, 17.5% between 48 h and 7 days, and 74.5% more than 7 days after the last event, complementing the ASM use and sleep-related seizure patterns reported in Table 1.

Comparison of HRV parameters under different testing paradigms

Overall, PWE exhibited significantly reduced heart rate variability compared to HC across all analyzed paradigms. The detailed statistical comparisons for all 17 calculated HRV parameters are presented in Table 2. The key findings concerning the 13 primary parameters selected for modeling are summarized below.

Table 2.

Comparison of HRV parameters under different testing paradigms between EP group and HC group.

	Supine resting test			Active standing test			Valsalva test			Sustained handgrip test			Whole experiment test
	PWE	HC	p-value	PWE	HC	p-value	PWE	HC	p-value	PWE	HC	p-value	PWE	HC	p-value
SDNN (ms)	27.55(21.02, 40.62)	35.3(24.95, 46.02)	1.82e−05	29.5(22.98, 38.2)	35.2(27.38, 46.08)	7.42e−06	44.85(34.83, 56.72)	51.35(39.35, 64.53)	2.11e−04	24.85(19.0, 31.45)	29.8(22.78, 37.05)	1.03e−05	37.85(29.98, 45.9)	43.15(34.75, 54.1)	2.06e−05
RMSSD (ms)	29.55(21.02, 43.48)	35.85(25.5, 54.12)	9.88e−05	24.45(19.3, 35.1)	29.95(21.08, 42.05)	4.97e−04	31.35(23.7, 42.9)	36.4(28.6, 47.58)	4.53e−04	19.6(14.4, 27.42)	24.5(18.23, 30.85)	7.22e−05	27.15(21.12, 39.82)	32.9(25.18, 43.2)	1.39e−04
NN50	34(4, 94)	55(15, 144)	1.16e−04	28(8, 77)	52(16, 106)	8.03e−04	30(16, 62)	45(24, 78)	4.34e−04	9(2, 27)	17(5, 50)	6.44e−05	118(55, 297)	214(89, 374)	1.06e−04
pNN50 (%)	7.7(1.08, 23.7)	13.45(3.78, 37.42)	5.78e−05	4.4(1.3, 13.02)	8.75(2.6, 19.62)	4.39e−04	8.0(3.8, 16.97)	13.15(6.52, 21.7)	1.56e−04	1.4(0.3, 6.8)	3.75(1.0, 10.22)	2.67e−04	5.8(2.48, 15.12)	10.35(4.57, 19.23)	9.82e−05
Triangular index	7.55(5.88, 10.38)	9.13(6.97, 11.71)	2.47e−06	7.43(6.12, 9.74)	9.06(7.04, 11.55)	1.99e−05	10.23(7.9, 12.42)	11.68(9.05, 14.17)	6.33e−04	7.01(5.36, 8.49)	8.03(6.38, 9.83)	3.24e−05	8.77(6.82, 10.73)	10.58(8.21, 13.18)	1.22e−05
avgRR (ms)	815.20 (738.88, 881.75)	866.70 (798.05, 935.02)	2.47e−06	744.50 (676.05, 814.45)	781.50 (727.42, 838.00)	6.36e−04	744.90 (676.77, 813.75)	784.60 (725.22, 840.92)	1.31e−04	716.75 (651.58, 781.80)	746.20 (689.47, 809.70)	0.0011	740.65 (678.00, 810.28)	783.10 (722.72, 838.27)	2.94e−04
SD1	20.8(14.75, 30.65)	25.05(17.88, 38.15)	9.63e−05	17.15(13.5, 24.42)	20.95(14.8, 29.52)	5.55e−04	21.75(16.45, 29.65)	25.4(19.9, 33.12)	4.30e−04	13.8(10.07, 19.2)	17.2(12.75, 21.65)	7.89e−05	19.0(14.7, 27.92)	23.0(17.5, 30.28)	1.38e−04
SD2	33.05(24.78, 47.05)	40.65(30.18, 54.02)	8.53e−06	37.0(28.75, 48.45)	44.3(36.33, 58.92)	9.52e−07	59.3(46.1, 73.7)	69.55(51.25, 84.43)	1.84e−04	31.8(24.95, 40.4)	39.0(29.38, 47.5)	3.86e−06	49.15(38.98, 58.7)	57.2(45.85, 70.47)	1.04e−05
α1	0.9(0.74, 1.04)	0.85(0.69, 1.06)	0.208	1.15(1.0, 1.27)	1.14(1.0, 1.31)	0.904	1.34(1.16, 1.47)	1.31(1.17, 1.45)	0.508	1.23(1.06, 1.39)	1.23(1.05, 1.38)	0.973	1.33(1.21, 1.42)	1.29(1.16, 1.41)	0.0999
α2	0.38(0.32, 0.48)	0.35(0.28, 0.44)	0.00245	0.58(0.49, 0.67)	0.57(0.49, 0.64)	0.273	0.64(0.53, 0.77)	0.58(0.48, 0.71)	0.0103	0.5(0.42, 0.58)	0.46(0.39, 0.55)	0.0106	0.5(0.42, 0.57)	0.47(0.42, 0.53)	0.00996
DC (ms)	11.19(8.03, 15.82)	13.99(9.85, 18.97)	4.65e−06	10.02(7.59, 13.1)	12.08(8.94, 16.42)	9.36e−06	14.39(11.06, 18.75)	16.39(12.07, 21.39)	0.00172	8.19(6.18, 10.39)	9.85(7.63, 12.71)	8.03e−06	11.65(9.32, 14.84)	13.81(10.69, 17.92)	1.45e−05
AC (ms)	11.74(7.96, 16.51)	13.99(9.94, 20.14)	7.70e−05	10.75(8.33, 13.57)	12.72(9.5, 16.89)	5.22e−05	12.96(10.12, 16.4)	15.76(11.75, 19.7)	2.10e−05	8.97(6.62, 11.58)	10.71(8.17, 14.22)	6.97e−06	11.85(9.17, 14.62)	14.11(11.09, 17.83)	5.50e−06
VLF (ms²)	34.5(18.75, 74.25)	48.0(26.0, 91.0)	7.08e−04	144.0(80.0, 269.0)	222.5(115.75, 415.0)	5.33e−05	386.0(197.75, 673.25)	469.0(211.75, 861.5)	0.0631	40.0(23.75, 79.5)	64.0(34.0, 107.25)	3.14e−05	151.5(98.75, 219.0)	177.0(127.0, 265.5)	0.00174
LF (ms²)	245.5(123.0, 437.75)	364.5(188.25, 680.0)	1.55e−04	486.5(271.25, 805.5)	705.5(397.75, 1061.75)	1.06e−04	1093.0(690.0, 1786.75)	1558.5(851.0, 2497.0)	2.25e−04	319.5(176.75, 542.75)	477.0(247.0, 731.5)	3.19e−05	954.5(557.0, 1343.5)	1215.5(782.0, 1824.5)	5.18e−05
HF (ms²)	302.0(150.5, 660.75)	498.5(227.0, 1082.0)	8.70e−05	245.0(133.5, 470.0)	374.0(167.5, 703.25)	7.38e−04	366.0(207.75, 641.75)	529.5(274.75, 943.0)	5.46e−04	137.5(66.5, 263.25)	196.0(99.0, 361.0)	1.99e−04	249.0(159.25, 520.25)	420.0(216.5, 718.25)	6.91e−05
LF/HF	0.76(0.44, 1.55)	0.71(0.38, 1.33)	0.295	1.97(1.31, 3.14)	1.83(1.06, 3.33)	0.416	3.16(2.07, 4.68)	2.75(1.85, 4.27)	0.288	2.36(1.35, 3.91)	2.15(1.26, 3.82)	0.596	3.38(2.23, 4.69)	2.89(1.95, 4.34)	0.0298
TP (ms²)	624.5(359.5, 1422.75)	1085.0(510.75, 1887.75)	2.12e−05	948.5(536.0, 1654.25)	1401.0(856.75, 2311.0)	1.40e−05	2017.5(1195.5, 3181.5)	2703.5(1591.75, 4218.0)	2.85e−04	542.0(307.5, 897.25)	794.5(472.5, 1223.0)	6.20e−06	1470.0(871.0, 2105.25)	1922.0(1173.0, 2912.75)	3.05e−05

PWE: people with epilepsy, HC: healthy control, BMI: body mass index, NHS3: National Hospital Seizure Severity Scale

Supine resting test: At baseline, the PWE group demonstrated significantly lower values across all 13 primary HRV parameters compared to the HC group. This included markers of overall HRV (e.g., SDNN: 27.55 vs. 35.30 ms), parasympathetic modulation (e.g., RMSSD: 29.55 vs. 35.85 ms), and markers influenced by sympathetic activity (e.g., SD2: 33.05 vs. 40.65) (all p < 0.001). This pattern is consistent with impaired vagal modulation and an altered sympathovagal balance even at rest. Active standing test: In response to orthostatic stress, the PWE group continued to show significantly blunted autonomic responses. All 13 primary HRV metrics remained significantly lower than in the HC group (all p < 0.001), suggesting an impaired ability to adapt to postural changes.

Valsalva test: During this baroreflex challenge, all 13 primary HRV parameters again showed significantly lower values in the PWE group (all p < 0.01), confirming a diminished autonomic response to pressure stimuli. Sustained handgrip test: This test, which primarily elicits a sympathetic response, also revealed significantly lower values for all 13 primary HRV parameters in the PWE group compared to controls (all p < 0.01). Whole experiment test: When analyzing the entire recording period as a single epoch, the PWE group consistently showed significantly lower values across all 13 primary HRV parameters, confirming a robust and persistent pattern of autonomic nervous system dysfunction.

In summary, the results consistently demonstrate that individuals with epilepsy have widespread cardiac autonomic dysregulation, characterized by reduced overall heart rate variability and blunted responses to various physiological stressors.

Performance of the HRV-based diagnostic models

A total of 400 participants were randomly split into a training set (n = 280) and a test set (n = 120) at a 7:3 ratio. Logistic regression models were developed for five distinct paradigms using HRV and demographic/anthropometric features.

In the training set, all single-paradigm models demonstrated moderate-to-good discriminative ability, with Area Under the Receiver Operating Characteristic Curve (AUC) values ranging from 0.71 to 0.74 (Fig. 2A, C, E, G, I). The model based on the active standing test achieved the highest AUC of 0.74, while the sustained handgrip model showed a robust combination of sensitivity (0.722) and specificity (0.566) with an AUC of 0.73.

In the independent testset, a performance drop was observed, though all models maintained an AUC between 0.60 and 0.71 (Fig. 2B, D, F, H, J). The sustained handgrip model proved to be the most robust, achieving the highest AUC of 0.71 with a sensitivity of 0.750 and a specificity of 0.375. In contrast, the other models showed a more noticeable decrease in performance, with AUCs in the range of approximately 0.60–0.63. This performance gap indicates that the apparent discrimination of the initial, unpenalized models in the training set was somewhat optimistic and justified our use of penalization and internal cross-validation to obtain more realistic estimates of generalizability.

Taken together, these results suggest that single-paradigm HRV models provide only modest but genuine diagnostic information and motivate the evaluation of combined-paradigm models that may capture complementary autonomic responses.

Fig. 2.

Performance of single-paradigm logistic regression models for epilepsy diagnosis. Receiver Operating Characteristic (ROC) curves for five logistic regression models, each based on a single autonomic testing paradigm. Panels (A, C, E, G, I) show model performance on the training set (n = 280), while panels (B, D, F, H, J) show performance on the independent test set (n = 120). The paradigms are: (A, B) Supine Resting, (C, D) Active Standing, (E, F) Valsalva, (G, H) Sustained Handgrip, and (I, J) Whole Experiment. The Area Under the Curve (AUC), accuracy (ACC), sensitivity, and specificity, calculated at the optimal cutoff point determined by the Youden index, are displayed with their 95% confidence intervals.

Feature importance in diagnostic models

To identify the most influential features, we evaluated the feature importance in the logistic regression models based on their standardized coefficients (Fig. 3).

Fig. 3.

Importance ranking of features in the epilepsy diagnostic models across five autonomic paradigms. This figure illustrates the contribution of each feature to the logistic regression models for five different paradigms. Feature importance was quantified by the magnitude of the standardized coefficients and ranked accordingly. The paradigms shown are: (A) Supine Resting Test, (B) Active Standing Test, (C) Valsalva Test, (D) Sustained Handgrip Test, and (E) Whole Experiment Test.

Supine resting test (Fig. 3A): In the resting state, parameters reflecting parasympathetic modulation were most important. RMSSD showed the highest importance, followed by TP and SD1. Notably, the anthropometric variable Waist Circumference ranked higher than several other HRV metrics. Active standing test (Fig. 3B): During the orthostatic challenge, the importance of parasympathetic markers was amplified. RMSSD remained the most dominant feature with a markedly elevated importance score, followed by SD1 and SD2, underscoring the sensitivity of these metrics to vagal withdrawal during postural changes.

Valsalva test (Fig. 3C): In response to this baroreflex challenge, SD1 emerged as the most influential feature, closely followed by RMSSD. This highlights that altered short-term cardiovagal variability is a key indicator during respiratory manoeuvres. Sustained handgrip test (Fig. 3D): During isometric effort, parasympathetic withdrawal markers remained highly discriminative. RMSSD and SD1 retained the leading positions, while pNN50 and NN50 also demonstrated substantial importance, ranking higher than in other paradigms. Whole experiment test: (Figure 3E): When integrating all paradigms, RMSSD and SD1 were overwhelmingly the most dominant features. The importance of SDNN was enhanced in the combined model, suggesting its value in capturing overall autonomic dysregulation across multiple physiological states. In contrast, demographic variables (age, sex) consistently showed minimal contributions.

In summary, across all paradigms, HRV parameters were the predominant features. Specifically, metrics sensitive to rapid, vagally mediated heart rate changes, such as RMSSD and SD1, consistently emerged as the most important features for distinguishing between individuals with epilepsy and healthy controls. The mathematical relationship between these two metrics (SD1 ≈ RMSSD / √2) explains their co-occurrence and confirms that interictal autonomic dysfunction in epilepsy is prominently characterized by altered vagal tone.

Performance of combined-paradigm models

To investigate whether combining paradigms could enhance diagnostic performance, we developed three models integrating the supine resting data with each of the three stimulation tasks.

In the training set, all three combined models showed moderate-to-good discrimination, with AUCs between 0.74 and 0.76 (Fig. 4A,C,E). The “Supine Resting + Active Standing” model achieved an AUC of 0.76, sensitivity of 0.750 and specificity of 0.596. The “Supine Resting + Valsalva” model yielded an AUC of 0.74, sensitivity of 0.736 and specificity of 0.588. The “Supine Resting + Sustained Handgrip” model also reached an AUC of 0.75, sensitivity of 0.750 and specificity of 0.588.

Fig. 4.

Performance of combined-paradigm models for epilepsy diagnosis. Receiver Operating Characteristic (ROC) curves evaluating the diagnostic performance of three combined-paradigm models. Panels (A, C, E) show performance on the training set (n = 280), while panels (B, D, F) show performance on the independent test set (n = 120). The combined paradigms are: (A, B) Supine Resting + Active Standing, (C, D) Supine Resting + Valsalva, and (E, F) Supine Resting + Sustained Handgrip. Performance metrics, including Area Under the Curve (AUC), accuracy (ACC), sensitivity, and specificity, were calculated at the optimal cutoff determined by the Youden index.

In the independent test set, performance declined and the AUCs were 0.61 and 0.62 for the “Supine Resting + Active Standing” and “Supine Resting + Valsalva” models, respectively (Fig. 4B,D). By contrast, the “Supine Resting + Sustained Handgrip” (Fig. 4F) model achieved the highest AUC in the test set, with an AUC of 0.68(95% CI 0.59–0.77), sensitivity of 0.821 and low specificity of 0.484. Given the low specificity, this model is not suitable as a stand-alone screening tool, particularly in low-prevalence populations. However, its high sensitivity indicates that it may serve as a useful auxiliary diagnostic tool when combined with clinical history, EEG findings, and other diagnostic tests. Overall performance remained moderate, with specificity limited, but integrating the sustained handgrip challenge with baseline resting state resulted in somewhat better discrimination than other tested combinations, primarily by enhancing sensitivity.

In internal K-fold cross-validation within the training set, penalized logistic regression models showed very similar discrimination to the single-split analyses, with mean AUCs remaining consistent (approximately 0.67) across the combined-paradigm models. For the primary combined model (supine rest + sustained handgrip), nested cross-validation on the full cohort yielded a mean AUC of approximately 0.67 (Supplementary Table S2), closely matching the performance observed in the held-out test set. These findings suggest that, while the HRV-based models do not achieve high discrimination, their moderate performance is robust to sampling variation and is not driven by severe overfitting.

Exploratory subgroup analyses

Within the epilepsy group, HRV indices were broadly similar across the predefined clinical subgroups (Supplementary Table S4). Median values of time- and frequency-domain measures did not differ significantly between focal and generalized epilepsy, between drug-resistant and controlled epilepsy, or between patients with wake-only versus sleep-related seizures, with all P values remaining above 0.05 in these exploratory comparisons. These findings suggest that the chronic reduction in vagally mediated HRV and altered sympathovagal balance observed in epilepsy is a relatively general feature rather than being confined to a specific seizure type, treatment response pattern or sleep-related phenotype.

For subgroup analysis, we applied an L1-regularized (LASSO) HRV model (supine rest + sustained handgrip) to test-set subgroups without re-fitting. Subgroup AUCs, sensitivities and specificities are in Supplementary Table S5 and Supplementary Figure S1. In the held-out test set, subgroup-specific AUCs ranged from 0.55 to 0.69 For seizure type, the AUC was 0.61 (95% CI 0.50–0.71) in focal epilepsy and 0.66 (0.48–0.82) in generalized epilepsy. For treatment response, AUCs were 0.62 (0.51–0.72) in controlled and 0.63 (0.49–0.77) in drug-resistant epilepsy. For sleep involvement, AUCs were 0.55 (0.43–0.67) in the wake-only group and 0.69 (0.57–0.80) in the sleep-related group. All subgroups were evaluated using the same decision threshold, resulting in identical specificity (0.67) and subgroup sensitivities ranging from 0.310 to 0.615. Overall, these values indicate consistently moderate discrimination across strata, with slightly higher AUCs in generalized epilepsy and sleep-related seizures and lower performance in wake-only seizures. However, subgroup sample sizes—especially for generalized (n = 13) and drug-resistant (n = 17) epilepsy—were modest and confidence intervals were wide. The subgroup findings should therefore be regarded as exploratory and hypothesis-generating rather than definitive.

Model calibration and clinical utility

To further evaluate the performance of the combined-paradigm models, we assessed their calibration and clinical utility using calibration plots and decision curve analysis (Figs. 5 and 6).

Fig. 5.

Calibration curves for the combined-paradigm epilepsy diagnostic models. These plots assess the agreement between predicted and observed probabilities for three combined-paradigm models. The diagonal dashed line represents perfect calibration; curves closer to this line indicate better diagnostic accuracy. Calibration curves for both the training set (Trainset) and test set (Testset) are shown. The three models combine supine resting data with: (A) Active Standing, (B) Valsalva, and (C) Sustained Handgrip.

Fig. 6.

Decision curve analysis evaluating the clinical utility of the combined diagnostic models. Decision curve analysis of the three combined models on the training set (A, C, E) and test set (B, D, F). The dashed line (pre) represents the diagnostic model, which provides a greater net benefit than either the “treat all” (solid gray line, All) or “treat none” (horizontal black line, None) strategies when it lies above them. The models combine supine resting data with: (A, B) Active Standing, (C, D) Valsalva, and (E, F) Sustained Handgrip.

First, calibration plots for the training and test sets were generated to compare the predicted probabilities against the observed outcomes (Fig. 5). In the test set, the “Supine Resting + Sustained Handgrip” model demonstrated the best calibration, with its bias-corrected curve closely tracking the ideal 45° line, indicating reasonably good calibration in this dataset (Fig. 5C). In contrast, the other two models showed signs of slight over-prediction, with their test set curves falling slightly below the ideal line, particularly in the mid-probability range (Fig. 5A, B).

Next, decision curve analysis was performed to evaluate the clinical net benefit of the models across a range of threshold probabilities (Fig. 6). In the crucial test set analysis (Fig. 6B,D,F), the "Supine Resting + Sustained Handgrip" model showed the highest net benefit among the three models across a range of threshold probabilities (approximately 0.20–0.75) in this dataset. Comparable patterns were observed in the training set (Figure 6A,C,E). These findings suggest that, under the assumptions of decision curve analysis, using this model to guide further diagnostic evaluation could offer some net benefit over “refer all” or “refer none” strategies, although the absolute gains are modest and require external validation.

Collectively, these analyses indicate that, within this sample, the “Supine Resting + Sustained Handgrip” model is the most discriminative of the three evaluated combinations and shows acceptable calibration, with some suggestive evidence of potential clinical utility. However, all models achieved only moderate discrimination, and these findings should be interpreted cautiously until replicated in independent cohorts.

Discussion

This study suggests that a multidimensional analysis of heart rate variability (HRV) can provide moderate discrimination between patients with epilepsy and healthy controls. Our principal finding is that epilepsy is characterized by a consistent pattern of cardiac autonomic dysregulation, observable both at rest and during physiological stress. We developed a diagnostic model based on these non-invasive biomarkers, in which the combination of a supine resting state and a sustained handgrip test showed comparatively better performance, achieving an AUC of 0.68 in an independent test set, corresponding to moderate overall discrimination with high sensitivity but limited specificity.

The underlying mechanism for these findings appears to be a widespread impairment of autonomic regulation in epilepsy. Our results consistently showed reduced HRV across time-domain, frequency-domain, and non-linear metrics (e.g., SDNN, RMSSD, HF), indicating diminished parasympathetic tone and altered sympathovagal balance. This dysfunction likely stems from the interference of epileptic activity with central autonomic networks, particularly the insular cortex and amygdala, thereby disrupting the brain-heart axis^36–39. The blunted responses observed during the active standing and Valsalva tests further suggest impaired baroreflex sensitivity and vagal reactivity in the interictal period⁴⁰. The significant reduction in non-linear indices such as Deceleration Capacity (DC) and Acceleration Capacity (AC) further reveals how underlying neuronal hyperexcitability may propagate to autonomic outflows, resulting in disordered cardiac rhythm dynamics^41,42. The comparatively better performance of the “supine rest combined with sustained handgrip” model may reflect its ability to capture both baseline autonomic tone and the dynamic response to a sustained sympathetic challenge, providing a more comprehensive autonomic profile than any single paradigm alone^16,43,44.

To achieve this comprehensive autonomic assessment, we adopted a standardized methodological approach. Specifically, we chose a short-term HRV protocol with 5-minute segments for each autonomic paradigm. This approach is consistent with the 1996 Task Force guidelines for HRV measurement and with many subsequent HRV studies²⁵, in which short-term recordings are considered appropriate for most time-domain indices and for high-frequency power as a marker of vagal modulation. In contrast, reliable estimation of very-low-frequency power and long-range correlation or complexity measures generally requires substantially longer recordings (e.g. ≥30-minute or 24-hour Holter monitoring), which capture slower oscillations and circadian influences that cannot be fully assessed in 5-minute segments. Within this framework, our models should therefore be interpreted as reflecting short-term autonomic modulation under standardized CARTs conditions, whereas complementary long-term or wearable-based monitoring may be needed to characterize multi-timescale autonomic dynamics more comprehensively in epilepsy.

Within this methodological framework, our multidimensional and standardized approach extends prior research by applying a comprehensive CARTs protocol in a relatively large, well-matched cohort. Although previous studies have established a link between reduced HRV and epilepsy, often in the context of SUDEP risk or peri-ictal prediction^19,45,46, many were limited by non-standardized, single-paradigm recordings or small sample sizes^47,48. In contrast, our prospective design with internal cross-validation provides preliminary, internally validated evidence that interictal HRV features can contribute diagnostic information. By employing a reproducible CARTs protocol, this study addresses some of the methodological inconsistencies that have previously limited the clinical translation of HRV research in epilepsy and offers a structured framework that can be further tested and refined in future work^49,50.

For comparison, routine scalp EEG—the primary electrophysiological biomarker in epilepsy—similarly exhibits only moderate diagnostic sensitivity. Meta-analytic data indicate that after a first unprovoked seizure, a single routine EEG detects epileptiform discharges with a sensitivity of roughly 15–20% and a specificity around 90–95% in adults⁵¹, while a recent prospective study reported sensitivities of 11% and 22% for the first and second routine EEG, respectively, with specificities of 98–100%⁵². In comparison, our combined HRV model achieved an AUC of 0.68 in the independent test set, with a sensitivity of approximately 0.82 but a lower specificity of about 0.48. These data suggest that HRV-based models may offer greater sensitivity than a single routine EEG but substantially inferior specificity, reinforcing the view that HRV should be considered as a complementary, low-cost auxiliary marker that may add incremental information alongside clinical assessment and EEG, rather than as a replacement for established electroencephalographic diagnostics.

These findings may have clinical implications, particularly for settings where access to specialist assessment and EEG is limited. Our HRV-based assessment provides an objective, low-cost and non-invasive measure of cardiac autonomic function that could support the auxiliary diagnosis of epilepsy rather than replace existing clinical standards⁵³. In particular, short, standardized recordings combined with simple physiological challenges may be practical for outpatient settings or for future integration with wearable or Internet of Things (IoT) technologies. However, given the only moderate discrimination (with a test-set AUC of 0.68) and the limited specificity of the primary combined model (sensitivity 0.821 at specificity 0.484), HRV-based models are not suitable as stand-alone screening tools, especially in populations with low pre-test probability. Instead, they should be viewed as complementary markers that might add incremental information when interpreted together with clinical history, EEG findings and other investigations.

Despite these strengths, several limitations must be acknowledged. First, this was a single-center study, so the generalizability of our findings and of the diagnostic models to other settings and populations remains uncertain and requires multicenter validation. Second, specificity was evaluated only against healthy controls rather than paroxysmal conditions that mimic epilepsy (e.g. psychogenic nonepileptic seizures, vasovagal syncope, sleep disorders), which may overestimate performance and does not address differential diagnosis. In addition, the standardized deep-breathing epoch in our protocol lasted only 1 min and was therefore not analyzed as a separate HRV condition, which may have limited the completeness of our parasympathetic assessment.

Third, although we excluded major cardiometabolic disease and systematically recorded antiseizure medication use, lifestyle factors, mood and seizure-to-testing intervals in the epilepsy group, these variables were not available in controls and were not included as covariates; residual confounding and differential timing effects between groups therefore cannot be ruled out. Finally, our HRV protocol relied on short-term (5-minute) recordings and the models focused on indices validated for this time scale, so the present results primarily characterize short-term autonomic modulation, while longer-term recordings will be needed to determine whether incorporating additional temporal scales and very-low-frequency or long-range complexity measures can further improve diagnostic performance.

In conclusion, this study indicates that a standardized, multi-paradigm HRV assessment may offer a feasible approach to identifying autonomic dysfunction in epilepsy and providing auxiliary diagnostic information. The comparatively better performance of the combined rest and handgrip model highlights the potential value of integrating diverse physiological assessments. This work should be viewed as an initial step toward the development of biomarker-informed, non-invasive tools for the auxiliary diagnosis of epilepsy, rather than a definitive solution, and will require rigorous external validation.  （It seems that the section on Abbreviations of HRV parameters is missing, and we have added it.） Abbreviations of HRV parameters AvgRR = Average RR interval (ms)SDNN = Standard Deviation of Normal-to-Normal intervalsRMSSD = Root Mean Square of Successive DifferencesNN50 = Number of pairs of successive NN intervals differing by more than 50mpNN50 = Percentage of NN50SD1 = Standard Deviation 1SD2 = Standard Deviation 2DC = Deceleration CapacityAC = Acceleration CapacityVLF = Very Low-Frequency powerLF = Low-Frequency powerHF = High-Frequency powerLF/HF = Low Frequency to High Frequency ratio

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Y.L. drafted the manuscript, Y.L. and W.L. conducted the study. Y.L., Q.C., X.G. and M.T. carried out the statistical analysis. R.Y., Q.C. and X.X. performed HRV assessment. R.Y. and H.Q. recruited participants, L.C. designed the study and provided financial support. All authors have read and agreed to the published version of the manuscript.

Funding

This research study was supported by Sichuan Science and Technology Program (2024ZDZX0018).

Data availability

The data of the present study are available upon reasonable requests from the corresponding author.

Declarations

Competing interests

The authors declare no competing interests.

Ethical approval

This study was conducted in accordance with the Declaration of Helsinki, and approved by the Ethics Committee of West China Hospital of Sichuan University (No. 2023–2394).