Predicting neurological recovery following cardiac arrest based on dynamic brain–heart coupling

Abstract

Background

Cardiac arrest (CA) often results in severe neurological injury, with both the central (CNS) and autonomic nervous systems (ANS) playing critical roles in recovery. Brain–heart coupling (BHC) and heartbeat-evoked potentials (HEP) reflect CNS and ANS interactions, yet their temporal evolution after return of spontaneous circulation (ROSC) and prognostic relevance remain unclear.

Objective

This study aims to examine frequency-specific variations in BHC and HEP within the first 96 h after ROSC in cardiac arrest patients and evaluate their prognostic value for neurological recovery.

Methods

We analyzed physiological data from 277 CA patients, focusing on BHC and HEP metrics. Unlike previous studies, hourly analyses were performed for each patient in order to investigate the temporal evolution of BHC and HEP following cardiac arrest. EEG and ECG data were preprocessed, and the Poincaré Sympathetic-Vagal Synthetic Data Generation (PSV-SDG) model was used to compute the strength. Refined Composite Multiscale Entropy (RCMSE) was applied to assess BHC complexity. Additionally, HEP was further examined for spatiotemporal features. Logistic regression (LR) and support vector machine (SVM) models were applied to predict 3-month outcome using BHC and HEP features at multiple post-CA time points to assess their prognostic value over time.

Results

We found that patients with good outcome demonstrated significantly higher BHC strength and complexity, particularly in the delta and theta bands, with notable increases between 38–47 h post-ROSC. HEP analysis showed enhanced positive peaks and widespread cortical activation in this group, suggesting more robust cortical-autonomic integration. In contrast, the poor outcome group exhibited weaker, more localized BHC and unstable HEP responses. Machine learning models combining BHC and HEP features achieved strong prognostic performance, with the highest AUC of 0.98 observed at 70 h post-ROSC, identifying it as the most informative time point for outcome prediction.

Conclusion

This study demonstrates that BHC and HEP metrics offer important insights into post-CA recovery. Notably, measures obtained at 70 h post-ROSC provided the greatest prognostic value, highlighting their potential to inform early clinical decision-making in critical care.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12984-026-01971-2.

Introduction

Cardiac arrest (CA), a catastrophic clinical event characterized by abrupt cessation of cardiac function, leads to immediate loss of consciousness [1]. Numerous studies have demonstrated that 50%–70% of CA patients develop irreversible post-cardiac arrest brain injury (PCABI) secondary to global hypoxic-ischemic insult [2]. A significant proportion of these patients fail to regain consciousness, exhibit severe cognitive impairments, or progress to death.Early prediction of neurological injury is essential for timely intervention and optimizing recovery in CA patients. However, existing approaches for evaluating brain injury in CA patients remain limited by low specificity, delayed applicability, and reduced reliability in unconscious individuals. Consequently, there is a pressing need to develop objective and accurate clinical biomarkers to enhance early detection and prognostic assessment of brain injury following cardiac arrest [3].

Traditional neuroprognostic assessment systems have long focused on cerebral biomarkers such as EEG reactivity [4], serum neuron-specific enolase [5], and neuroimaging features [6]. While these methods provide critical insights into cortical integrity assessment, they overlook the impact of brain–heart interactions on post-CA outcome. Emerging evidence positions the brain–heart axis as a frontier in critical care neuroscience. Research in stroke [7] and traumatic brain injury [8] has revealed bidirectional brain–heart interplay mediated by autonomic dysfunction, where cerebral ischemic cascades trigger autonomic dysregulation characterized by sympathetic overactivation and vagal suppression. This dysregulation creates a vicious cycle: impaired autonomic control exacerbates cerebral hypoperfusion through hemodynamic instability, while progressive neurological injury further disrupts autonomic homeostasis [7]. Notably, although the initial injury in CA originates from cardiac etiology, secondary brain injury-driven autonomic dysfunction emerges as a pivotal mechanism influencing recovery trajectories [9]. Translating these findings into neuroprognostic assessment and prediction in CA populations holds unique clinical value. Heart rate variability (HRV), derived from electrocardiography (ECG), has emerged as a validated proxy for ANS activity and offers valuable complementary insights into brain–heart communication [10]. Integrating HRV with EEG-based monitoring enhances prognostic precision [11, 12].For example, elevated low frequency to high frequency (LF/HF) ratios have been associated with poor outcome (AUC = 0.85) [13], and the combination of EEG entropy features with HRV measures has demonstrated high predictive value (AUC = 0.954) in animal models [14]. Although these approaches improve prediction accuracy by combining EEG and HRV, they primarily treat EEG and HRV as independent data sources and rely on statistical methods for feature aggregation, neglecting their temporal synchrony, phase coupling, or causal relationships. Such simplistic feature fusion overlooks the physiological basis of bidirectional brain–heart communication, resulting in prognostic indicators that lack mechanistic interpretability.

Recent studies have provided compelling evidence that dysfunction of the autonomic nervous system (ANS) was closely linked to impaired BHC, a phenomenon reflecting the bidirectional interaction between cortical neural activity and autonomic cardiovascular regulation. BHC captures the dynamic synchronization between cerebral electrical signals, such as electroencephalography (EEG), and cardiac dynamics quantified by heart rate variability (HRV), thereby offering an integrative window into functional communication between the central and autonomic nervous systems [15]. Hermann et al. used a synthetic data generation model to computed the strength and complexity of bidirectional interactions between EEG frequency bands (delta, theta, and alpha) and ECG heart rate variability frequency bands (low frequency, LF and high frequency, HF).Their findings revealed that reduced brain–heart coupling strength and complexity were associated with poor outcome  [16]. Additionally, Kumral et al. observed attenuated HEP in patients with atrial fibrillation, particularly in the right insular cortex, a central hub of autonomic control [17]. This further supports the relevance of brain–heart coupling as a window into central autonomic integrity. Together, these studies highlight its potential application for integrative monitoring and outcome prediction in critically ill populations.However, despite the promising findings, several limitations persist in current research on BHC in CA: (1). Most existing studies rely on fixed-band spectral analysis to assess cardiac sympathetic activity; however, this approach was limited by overlap with vagal rhythms and fails to capture the nonlinear, time-varying nature of autonomic modulation. (2). Current studies predominantly utilize short-duration data segments from the early post-CA period, overlooking the longer-term evolution of BHC throughout recovery. (3). it remains unclear which post-CA time windows provide the most prognostically informative BHC features.

To address these gaps, the present study systematically investigated the temporal evolution and prognostic value of BHC in a multicenter cohort of 277 comatose patients after cardiac arrest. Continuous EEG and ECG signals were recorded between 13 and 96 h post-ROSC. BHC dynamics were quantified using the PSV-SDG model to capture nonlinear and bidirectional interactions between cortical and autonomic activity. In parallel, HEPs were analyzed to assess event-related cortical responses to cardiac input. To evaluate their prognostic utility, supervised machine learning models (LR and SVM) were applied at multiple time points post-ROSC. This integrative, time-resolved framework aims to identify critical neurophysiological windows and improve the accuracy of early outcome prediction in post-cardiac arrest care.

Material and methods

Dataset

The data used in this study were obtained from the International Cardiac Arrest Registry (I-CARE) database [18],which comprises data collected from seven academic institutions. For the present study, the authors were granted access to data from five participating hospitals, which constituted the study cohort.This cohort includes adult patients who experienced either in-hospital or out-of-hospital cardiac arrest, achieved return of spontaneous circulation (ROSC) and remained in a comatose state, defined as a Glasgow Coma Scale (GCS) score ≤ 8. The dataset includes 606 patients provided by I-CARE, collected from multiple participating hospitals. EEG and ECG monitoring were initiated during hypothermia and continued through the rewarming phase.The dataset comprises recordings from five participating hospitals, each employing different acquisition hardware, EEG electrode layouts, and sampling rates. To address these site-specific variations, a standardized channel naming convention was adopted, with channels categorized into EEG, ECG, reference, and miscellaneous groups.Details on the original acquisition protocols and hospital-specific configurations are provided in Appendix Table 1.

The original data collection were approved by the respective ethics committees, as outlined by PhysioNet. Since the data was fully anonymized, no direct patient consent was required, and the use of the dataset was conducted under the PhysioNet Credentialed Health Data Use Agreement.

Preprocessing

Hourly EEG/ECG recordings 13–96 h post-ROSC, acquired simultaneously via a unified monitoring system, were analyzed. To ensure quality, one 5-min segment per hour was extracted if both signals met predefined quality criteria (no major artifacts, physiologically plausible morphology, stable baselines). hours without a valid segment were excluded. The sections below detail the exclusion rules and modality-specific preprocessing. All preprocessing was performed in MATLAB R2020a.

EEG preprocessing: Signals underwent fourth-order Butterworth bandpass filtering (0.5–45 Hz) and were downsampled to 100 Hz to standardize multi-site data and avoid aliasing. Artifact correction used EEGLAB’s [19] ‘runica’ independent component analysis (ICA) algorithm: two certified technologists identified and removed ocular, muscular, and cardiac components based on scalp topography, waveform morphology, and spectral content; no fixed thresholds were applied. Post-ICA, signals were re-referenced to the common average, and 19 electrodes conforming to the international 10–20 system were retained (Appendix Fig. 2).

EEG spectrograms were generated via STFT with a Hanning window (2-s window, 50% overlap), yielding 1-s time and 0.5-Hz frequency resolution, and were integrated into five bands: δ (0–4 Hz), θ (4–8 Hz), α (8–12 Hz), β (12–30 Hz), γ (30–45 Hz) [20].

ECG preprocessing: ECG signals were bandpass filtered using a fourth-order Butterworth filter (0.5–20 Hz). R-peaks were detected using the Pan-Tompkins algorithm [21]. An automated RR-interval point-process model identified and corrected physiologically implausible intervals, followed by a final manual quality-control step, during which fewer than 5% of beats required adjustment. Two trained annotators independently reviewed the ECG segments, and discrepancies were resolved by consensus. Segments that remained ambiguous after automated correction were excluded [12].

Evaluation of sympathetic and parasympathetic activity

Sympathetic and parasympathetic autonomic activity was quantified using the Poincaré plot method [22], which characterizes HRV by depicting fluctuations between consecutive inter-beat intervals (IBI). The key geometrical features quantified from the Poincaré plot are SD₁ and SD₂, representing short-term and long-term fluctuations of HRV, respectively. These features form an ellipse, and the ellipse ratios, SD₀₁ and SD₀₂, are computed as follows:

1

2

where IBI′ denotes the first-order difference of the inter-beat interval series (IBI′_n = IBI_n+1 − IBI_n), and std(·) represents the standard deviation. Within this framework, BHC are characterized by examining the temporal evolution of Poincaré ellipse descriptors. Accordingly, the time-varying fluctuations of these ellipse ratios are computed using a sliding time window, as defined in Eqs. (3) and (4).

3

4

where Ω_t: t -T ≤ t_i ≤ t, in this study T is fixed in 15 s. Based on these calculated features, the following indices are derived: the CVI, the CSI, and the Sympathovagal Balance Index (SVI), computed as:

5

6

7

where is the demeaned SD_x.

A comprehensive overview of the signal processing pipeline, including the computation of BHC and HEP features and their integration into machine learning models, is illustrated in Fig. 1.

Fig. 1.

Overview of BHC framework and data processing pipeline. A Bidirectional BHC modeled via the PSV-SDG framework. Ascending modulation reflects cardiac-to-cortical influence using cardiovascular indices (CSI, CVI) at time t and EEG at t + 1; descending modulation reflects cortical-to-cardiac regulation using EEG at t and cardiovascular indices at t + 1. Sympathetic and parasympathetic pathways are shown in green and red. B Signal processing workflow. EEG and ECG data undergo preprocessing, R-peak detection, and artifact removal. BHC is computed using PSV-SDG and frequency-domain integration. HEPs are extracted from R-peak-locked EEG epochs. Time-resolved features are fed into machine learning models (SVM, LR) to predict outcome across multiple post-ROSC time points

Brain heart-coupling

Estimation of functional brain–heart interactions

The brain–heart interaction process was estimated using the PSV-SDG model [23]. This model quantifies the brain’s influence on the heart and describes heartbeat generation through a pulse frequency modulation model, parameterized by the features of the Poincaré plot. The detailed steps are as follows:

1) Heartbeat generation model:

The synthetic heartbeats were modeled as Dirac delta functions, denoted as δ(t), which generate a heartbeat at each occurrence time t_k. The heartbeat generation occurs when the integral function of a reference heart rate μ_HR and a modulation function m(t) reaches a threshold value of 1,

8

9

2) Modulation function:

Equation (10) specifies the modulation function m(t) as a combination of two oscillators that represent the sympathetic and vagal autonomic outflows. The oscillators are centered at the frequencies ω_s and ω_V with corresponding amplitudes defined by C_S and C_V, respectively:

10

3) HRV calculation:

The parameters C_S and C_V were calculated based on HRV. The length L and width W of the Poincaré plot are used for this calculation:

11

12

where IBIk is the inter-beat interval.

To link C_S and C_V with HRV, analytical approximations based on the geometric properties of the Poincaré plot and the reference heart rate (μ_HR) were applied. Solving the corresponding system of equations yields the expressions for C_S and C_V as functions of L and W:

13

Here, is a constant term derived from frequency components.

4) Brain–heart interaction coefficients:

Finally, the brain–heart interaction coefficients are quantified as the ratio between C_S and C_V, along with the EEG power in the frequency bands (i.e., δ, θ, α, β, γ) during the previous time window a_F(t-1). The brain-to-sympathetic and brain-to-vagal interaction coefficients are denoted as C _F→CSI and C _F→CVI, respectively.

14

15

Estimation of functional heart-brain interactions

The functional interplay from heart to brain is quantified through a model based on the generation of synthetic EEG series using an adaptive Markov process, as shown in Eq. (16).where is the main frequency in a defined frequency band, and stands for its associated phase. The variation in EEG power is estimated through a coefficient using least squares in an exogenous autoregressive process, as shown in Eq. (17), is a constant, and is a Gaussian white noise term.

16

17

The generation of Markovian neural activity incorporates both its prior neural states and concurrent heartbeat dynamics as input sources for EEG signal modeling. Specifically, the coupling coefficients and are derived by capturing the influence of cardiac dynamics and the external components within the autoregressive framework. As follows:

18

19

Quantification of BHC strength and interaction complexity

The PSV-SDG model yields time-varying coupling coefficients, which were computed for each 5-min EEG and ECG segment using 15-s sliding windows and sampled at 1 Hz. For each frequency band and interaction direction (brain-to-heart and heart-to-brain), coefficients were aggregated across time and EEG channels using the median, yielding a robust and unitless measure of BHC strength. Interaction complexity was further assessed using RCMSE [16] computed on the time-varying coupling coefficient series. RCMSE was calculated with an embedding dimension m = 2, a tolerance r = 0.15 × SD of the coupling coefficient series, and scale factors τ ranging from 1 to 20.

Heartbeat-evoked potentials

Heartbeat-evoked potentials refer to transient neural responses triggered by each heartbeat, which were captured via EEG. To compute HEPs, EEG epochs were extracted based on their alignment with the R-wave peak of the cardiac cycle within a time window of [−200 ms, 400 ms]. Epochs with an amplitude exceeding 300 µV were excluded from further analysis. Additionally, epochs corresponding to IBI shorter than 600 ms were discarded. In cases where more than 20% of EEG epochs were excluded due to insufficient IBI duration, the HEP computation latency was recalibrated to ensure that at least 80% of the epochs were retained [24].

Machine learning

To predict neurological outcome, we trained two supervised classification models: SVM and LR. The CPC score at hospital discharge was used as the outcome label, with CPC 1–2 defined as a Good outcome (label = 1) and CPC 3–5 as a Poor outcome (label = 0). BHC related features extracted from the EEG and ECG signals served as model inputs at each analyzed time point.

All models were evaluated using stratified tenfold cross-validation to preserve the proportion of Good and Poor outcome within each fold. To prevent information leakage, the entire preprocessing pipeline was performed strictly on the training portion of each fold and then applied to the corresponding validation data. Specifically, continuous features were first z-score normalized using statistics derived from the training data only, and the same parameters were applied to the validation set. Because outcome imbalance varied across time points, the Synthetic Minority Oversampling Technique (SMOTE) was applied exclusively to the normalized training data within each fold, ensuring that no synthetic samples influenced model evaluation [25]. Hyperparameter tuning for both LR and SVM was conducted using grid search nested within the training folds. For LR, the regularization strength (C) and penalty type were optimized, whereas for SVM, C and kernel parameters were tuned. The optimal configuration identified within each fold was then used to retrain the model on the full training set (after resampling), which was subsequently evaluated on the held-out validation fold. Model performance was summarized using standard classification metrics, including accuracy, sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC-ROC) [26].

Statistical analysis

To determine whether the observed BHC reflected genuine physiological interactions rather than spurious correlations, we performed a surrogate-data analysis based on isospectral phase randomization. For both EEG signals and RR-interval (HRV) time series, 100 surrogate realizations were generated by randomizing temporal phases while preserving their original power spectra [16]. Surrogate-based p-values were then computed by comparing the empirical coupling coefficients with the surrogate-derived null distribution. Group-level inference was conducted using a nonparametric permutation framework. For each comparison, the null distribution of the test statistic was estimated using 10,000 random permutations of the outcome labels, thereby avoiding assumptions regarding data normality. Multiple comparisons across frequency bands and time points were controlled using the Benjamini–Hochberg false discovery rate procedure (FDR = 0.05).

For HEP, group differences were evaluated using non-parametric cluster-based permutation tests to account for the strong temporal dependency of EEG signals. At each electrode, point-wise comparisons between the good and poor outcome groups were performed using the Wilcoxon rank-sum test, and temporally contiguous samples exceeding an uncorrected threshold of p < 0.05 were grouped into clusters, with a minimum duration of 10 ms. Cluster-level statistics were defined as the sum of absolute z-values within each cluster. Statistical significance was assessed by comparing observed cluster statistics with a null distribution generated from 2,000 random permutations of group labels, thereby controlling the family-wise error rate across time. Clusters with permutation-based p < 0.05 were considered significant.

Results

Population description

In this study, a total of 606 patients were initially identified from the I-CARE dataset. After applying exclusion criteria, including absence of ECG recordings (n = 201), missing targeted temperature management (TTM) documentation (n = 91), poor EEG/ECG signal quality (n = 25), and insufficient data coverage within the 13–96 h post-ROSC window (n = 12), a final cohort of 277 patients was retained for analysis (Appendix Fig. 1). Baseline demographic and clinical characteristics are summarized in Table 1. The mean age of the cohort was approximately 61 years, and 31.4% of patients were female. Among the included patients, 99 (35.7%) had a good outcome (CPC 1–2), whereas 178 (64.3%) had a poor outcome (CPC 3–5). Differences in age, initial cardiac rhythm, and sex distribution between outcome groups are reported in Table 1.

Table 1.

Patient characteristics

	Total N = 277	Good outcome N = 99	Poor outcome N = 178	p-value
Age (years)	61.04 ± 16.52	56.49 ± 14.04	63.15 ± 17.18	0.0007 (t-test)
Female gender (%)	31.4	23.6	35.3	0.08 (chi-square)
Shockable rhythm (VFib/VT,%)	42.5	69.3	31.6	< 0.0001 (chi-square)
TTM (33/36)	277 (72)	99 (24)	178 (48)	–

VFib: ventricular fibrillation; VT: ventricular tachycardia; TTM: targeted temperature management

Frequency specific dynamics of brain–heart coupling

In this study, BHC was evaluated hourly from 13 to 96 h post-resuscitation using the PSV-SDG model on five-minute clean EEG-HRV segments. After multiple comparison correction, significant differences in BHC metrics were observed at distinct time points, notably at 24, 34, 38–41, 47, 51, and 70 h post-ROSC (Fig. 2; full statistics in Appendix Table 11), suggesting critical windows of physiological change. Among frequency bands, the delta band exhibited the most pronounced and sustained alterations, particularly between 38 and 47 h, with a peak at hour 47-potentially reflecting key recovery mechanisms such as neural plasticity and autonomic stabilization. In contrast, alpha-band BHC showed more localized changes around discrete time points, possibly indicating the reactivation of cortical-autonomic regulatory circuits. Theta-band variations were less consistent and more transient, suggesting limited involvement in sustained recovery dynamics.

Fig. 2.

Filtered Significant Heatmap of Brain–Heart Coupling Metrics (-log10 (FDR-adjusted P-Value))

Dynamics of brain–heart coupling strength and complexity

In this study, we systematically evaluated the dynamic strength and complexity of bidirectional BHC, encompassing both cortical-to-cardiac and cardiac-to-cortical interactions.From the perspective of coupling strength, patients with good outcome exhibited a clear upward trend in both brain-to-heart (Fig. 3) and heart-to-brain (Fig. 4) directions, especially in low-frequency bands (delta and theta). In the delta band (Brain-to-heart), coupling strength increased from 0.4 to 0.8 by the 70th hour (mean: 0.6 ± 0.15), while the poor outcome group remained relatively stable (0.5 ± 0.05; p_fdr = 0.043 < 0.05,Cohen’s d = 0.677). Although transient increases were observed in the poor outcome group between 30–50 h, these did not reach statistical significance. The good outcome group also showed a more consistent upward trend in higher-frequency bands (e.g., alpha and gamma).

Fig. 3 .

Median coupling strength from EEG to CSI and CVI across frequency bands. A EEG-to-CSI and B EEG-to-CVI coupling strength over time in delta, theta, alpha, beta, and gamma bands. Data were presented as median with interquartile range (IQR) over time. *Pfdr < 0.05

Fig. 4 .

Median coupling strength from CVI/CSI to EEG across frequency bands. A CSI-to-EEG and B CVI-to-EEG coupling strength over time in delta, theta, alpha, beta, and gamma bands. Data were presented as median with interquartile range (IQR) over time. *Pfdr < 0.05

Regarding complexity, the good outcome group showed a sustained increase across both brain-to heart (Fig. 5) and heart-to-brain (Fig. 6) directions. In the delta band (heart-to-brain), complexity rose from 0.22 to 0.46 by the 70th hour (mean: 0.34 ± 0.10), while remaining below 0.18 in the poor outcome group. A similar trend was observed in the gamma band (brain-to-heart), where complexity increased from 0.004 to 0.007 in the good outcome group, but stayed below 0.004 in the poor outcome group (p_fdr = 0.0034 < 0.01,Cohen’s d = 0.784). Notably, although the poor outcome group occasionally exhibited comparable coupling strength in low-frequency bands, its consistently lower complexity suggests transient and less stable BHC dynamics.

Fig. 5 .

Temporal evolution of brain-to-heart coupling complexity across EEG frequency bands. A EEG-to-CSI and B EEG-to-CVI coupling complexity, quantified by RCMSE, across delta, theta, alpha, beta, and gamma bands. Data were presented as median with interquartile range (IQR). *Pfdr < 0.05

Fig. 6 .

Temporal evolution of heart-to-brain coupling complexity across EEG frequency bands. A CSI-to-EEG and B CVI-to-EEG coupling complexity, quantified by RCMSE, across delta, theta, alpha, beta, and gamma bands. Data were presented as median with interquartile range (IQR). *Pfdr < 0.05

Spatiotemporal dynamics of heartbeat-evoked potentials

In this study, HEPs were examined at nine post-ROSC time points selected based on statistically significant differences in BHC metrics. Group-level comparisons revealed distinct HEPs dynamics between patients with good and poor outcome, with differences predominantly emerging during the early post-R-wave period, corresponding to cortical processing of interoceptive cardiac signals.

As shown in Fig. 7, cluster-based permutation testing identified significant amplitude differences between outcome groups across multiple scalp regions. Robust group differences were most consistently observed at fronto-central electrodes (Fz, Cz) and sensorimotor sites (C3, C4), primarily within the early post-R-wave interval (approximately within the first 100 ms). Compared with the poor outcome group, patients with good outcome exhibited significantly larger positive HEPs deflections at Fz and Cz, indicating enhanced cortical responsiveness to cardiac afferent input. In sensorimotor regions, particularly at C3, the good outcome group demonstrated markedly greater positive amplitudes, while C4 showed similar but less pronounced trends. Although this pattern may suggest lateralized sensorimotor involvement, no definitive conclusions regarding hemispheric dominance were drawn due to the limited spatial resolution of the 19-channel EEG montage and the absence of source-level analyses.

Fig. 7.

Grand-averaged HEPs at representative EEG channels for patients with Good (blue) and Poor (red) outcome, with black traces showing the difference waveforms (Good-Poor). Signals are time-locked to the R-peak (0 s) and displayed for frontal (FP1, FP2, FZ), central (CZ), and motor-related (C3, C4) electrodes. The x-axis denotes time (s) and the y-axis denotes EEG amplitude (µV)

The scalp topographies of HEPs activity (Fig. 8) provide a complementary, descriptive view of the spatiotemporal evolution of cortical cardiac processing across recovery stages. During the early post-ROSC phase (24–34 h), HEPs activity was largely confined to midline and frontal regions. In the intermediate phase (38–41 h), activity extended toward bilateral sensorimotor cortices, potentially reflecting re-engagement of motor-related cortical networks. By the later phase (47–70 h), HEPs responses became more spatially distributed, while central regions continued to exhibit dominant activity, suggesting progressive restoration of large-scale cortical-autonomic integration.

Fig. 8.

Scalp topographies illustrating group-level differences in HEPs at representative time points after ROSC. Each map displays the spatial distribution of HEP amplitude differences between patients with good and poor outcome. Warmer colors indicate greater HEP amplitudes in the Good Outcome group relative to the Poor Outcome group

Prognostic analysis of patients with cardiac arrest

We selected nine representative time points to further evaluate the prognostic value of BHC, HEP, and their combination (BHC + HEP). For each time point, we applied machine learning models SVM and LR to assess the predictive performance of each feature set.

Figure 9 illustrates the trends in the AUC for the SVM and LR models across different hours, using the BHC, HEP, and combined BHC + HEP datasets. Specifically, we focus on the predictive performance at hour 70. To illustrate peak performance, we highlight the results at the 70th hour (Table 2, Fig. 9). At this time point, the SVM model using BHC features achieved the highest AUC of 0.962 [95% CI 0.946–0.976], along with high accuracy (0.885), sensitivity (0.883), specificity (0.883), and F1 score (0.879). The LR model also performed well (AUC: 0.944). For the HEP features, both models showed robust AUCs (SVM: 0.923, LR: 0.901), and perfect specificity (1.000). When combining BHC and HEP features, the LR model achieved the highest AUC (0.980 [0.962–0.989]) and specificity (0.953). The complete model performance results for all nine time points were reported in Tables 2–9 (Appendix). To further quantify the added prognostic value of feature integration, baseline comparisons with EEG-only and HRV-only models at the best-performing time point (70 h post-ROSC) were provided in Table 10 (Appendix).

Fig. 9.

Temporal evolution of AUC for outcome prediction at different post-ROSC time points using BHC, HEP, and combined features with SVM and LR classifiers. Bars represent mean AUC values, and error bars indicate 95% confidence intervals obtained by bootstrap resampling

Table 2.

Performance metrics at 70th hour for HEP, BHC, and HEP + BHC models

	ACC	AUC[95% CI]	Sen	SPE	F1
BHC
SVM	0.885	0.962[0.946–0.976]	0.883	0.883	0.879
LR	0.803	0.944[0.924–0.964]	0.852	0.783	0.805
HEP
SVM	0.795	0.923[0.899–0.947]	0.616	1.000	0.666
LR	0.855	0.901[0.881–0.925]	0.716	1.000	0.746
BHC + HEP
SVM	0.888	0.951[0.930–0.970]	0.802	0.962	0.864
LR	0.866	0.980[0.962–0.989]	0.832	0.953	0.853

Discussion

To the best of our knowledge, this study was one of the first to examine the dynamic evolution of BHC across 96 h post-ROSC using hourly EEG-ECG data. The results show that BHC exhibits pronounced time- and frequency-specific dynamics, with discrete post-ROSC windows particularly between 38 and 47 h. showing the most prominent differences between neurological outcome groups. Across frequency bands, delta-band coupling emerged as the most consistent marker of recovery, while patients with Good outcome demonstrated both stronger and more complex bidirectional BHC compared with those with poor outcome. HEP showed complementary spatiotemporal differences, supporting progressive re-establishment of cortical-autonomic integration. Prognostic analyses further identified approximately 70 h post-ROSC as the most discriminative time point, at which combining BHC and HEP features yielded the highest predictive performance (AUC = 0.980).

Temporal dynamics and complexity of BHC in post-CA

The temporal and frequency-specific dynamics of BHC revealed distinct trajectories associated with neurological recovery following CA. Among all frequency bands, the delta band exhibited the most sustained and robust modulation. Between 38 and 47 h post-ROSC, patients with good outcome demonstrated progressive increases in both BHC strength and complexity, a pattern that aligns with previous research linking delta oscillations to neuroplasticity, metabolic restoration, and autonomic stabilization [27].These delta-driven changes may reflect enhanced cortical-subcortical communication and cerebral perfusion, which are consistent with processes involved in post-ischemic neural repair [28].

The theta band also displayed meaningful, albeit more transient, dynamics. Notably, fluctuations occurred between 30 and 50 h, potentially representing phases of autonomic reorganization and stress adaptation. Theta oscillations have been associated with sympathetic-parasympathetic balance and regulatory responses to physiological stress [16, 29]. However, the lack of sustained theta-BHC complexity suggests that autonomic control mechanisms remained unstable during this window in some patients. In contrast, the alpha band showed localized increases in BHC, especially between 40 and 51 h post-ROSC, potentially reflecting the reactivation of cortical inhibitory control and higher-order regulatory processes. This aligns with studies suggesting that enhanced alpha activity may indicate parasympathetic re-engagement and the recalibration of central autonomic networks [30, 31].

Beyond frequency-specific changes, a critical distinction emerged when analyzing the relationship between BHC strength and complexity across outcome groups. Patients with good outcome showed a coordinated and parallel increase in both metrics, particularly in heart-to-brain coupling, suggesting a more refined integration of peripheral cardiac signals and effective top-down modulation of autonomic activity [32]. In contrast, the poor outcome group often exhibited a "high-strength but low-complexity" BHC pattern, most notably between 30 and 50 h post-ROSC. This dissociation implies that elevated coupling strength alone may not be beneficial and could instead reflect maladaptive or non-specific hyper-synchronization, a compensatory mechanism that lacks informational richness [33].Such interpretations are supported by theoretical and empirical studies demonstrating that higher physiological complexity is a hallmark of healthy autonomic regulation and cardiovascular adaptability [34, 35]. Therefore, the progressive emergence of complexity in BHC, especially within delta and theta bands, may serve as a dynamic biomarker of functional recovery, integrating both neurocardiac coordination and regulatory flexibility.

Spatiotemporal dynamics of heartbeat-evoked potentials

HEP reflect cortical responses to cardiac activity, providing a direct measure of BHC[36]. In this study, our analysis of HEP dynamics revealed significant spatiotemporal differences during recovery, particularly in central (Cz), frontal (Fz), and motor-related regions (C3, C4).

The scalp topography analysis of HEP reveals that, during the early phase, activity was predominantly localized in the central (Cz) and frontal (Fz, FP1, FP2) regions. This pattern of localized and relatively low-intensity activity was consistent with prior research indicating that early post-cardiac arrest brain responses were mainly limited to supporting essential survival mechanisms. Guo et al. demonstrated that subcortical structures, such as the brainstem and hypothalamus, play a leading role in the hyperacute recovery phase, contributing significantly to autonomic function stabilization in the absence of extensive cortical engagement [37]. Similarly, Allen et al. observed that early autonomic regulatory activity was concentrated in subcortical networks, while cortical involvement remains minimal during initial recovery stages [38]. These findings were in line with our results, reinforcing the notion that early neural reactivation prioritizes autonomic homeostasis over higher-order cognitive processes.As the recovery process progresses into the mid-phase (38–41 h post-resuscitation), HEP activity extends toward motor-related regions, including C3 and C4. This spatial redistribution may suggest a re-emergence of cortical-autonomic integration, particularly involving sensorimotor networks associated with somatic-autonomic coordination.This change was consistent with evidence from studies such as that by Satterthwaite et al. [39], which demonstrated increased cortical involvement-particularly in the motor cortex-during subacute post-CA phases, facilitating motor-autonomic coupling. These findings underscore that the mid-phase represents a transitional period, where both neural responsiveness and autonomic stability begin to reestablish in tandem.In the late recovery phase (47–70 h), HEP activity becomes more widespread and symmetrical, with a clear centralization at Cz. This spatial balance suggests the consolidation of distributed brain networks involved in interoceptive and autonomic regulation. Craig et al. emphasized that the anterior insula plays a critical role in integrating autonomic information and conscious awareness, suggesting that enhanced activity in central cortical regions reflects the maturation of integrative neural mechanisms necessary for maintaining homeostasis [40]. These late-stage changes indicate a shift toward globally coordinated brain-body interactions, representing a hallmark of functional recovery.

Based on HEP topographical analysis, waveform analysis further revealed significant differences between the Good and Poor Outcome groups, particularly within the 0–200 ms window, indicative of early cortical processing of cardiac afferent signals [41]. The Good Outcome group exhibited stronger positive peaks in CZ, FZ, and C3, reflecting enhanced cortical-autonomic integration, consistent with prior studies linking robust HEP responses to good recovery outcome [42]. Notably, the increased amplitude in CZ suggests a greater top-down neural regulatory role, while the C3 response supports the left hemisphere's dominant role in autonomic recovery [43]. In contrast, the Poor Outcome group demonstrated lower amplitude and unstable responses, pointing to impaired cortical processing and more severe neural injury, indicative of disrupted neuroautonomic regulation [44].

Comparison with classification performance in previous studies

Compared to previous studies, our results demonstrate a more reliable and effective approach for classifying outcome in CA patients. As summarized in Table 3, earlier works primarily relied on EEG features for classification, which may be limited by the lack of temporal dynamics and the inability to capture the full complexity of brain–heart. For example,Zheng et al. employed recurrent neural networks (RNNs) to analyze longitudinal EEG features, achieving 88% accuracy [45]. Kim et al. demonstrated that EEG biomarkers, including power spectral density and spectral entropy, can predict outcome in cardiac arrest survivors, achieving a classification AUC of 90.3% using a linear mixed model [46]. Similarly, Tjepkema et al. used EEG-based Cerebral Recovery Index (CRI) to assess recovery in comatose patients, reporting a classification AUC of 92% with RF classifiers based on integrative EEG metrics [47]. However, their approach did not incorporate autonomic features, which are essential for understanding BHC, nor did it explore the temporal dynamics of recovery or the evolving interactions between brain and heart functions. Recent studies have sought to address these gaps by incorporating both EEG and ECG features. For example, Krones et al. used multimodal deep learning to combine EEG functional connectivity with HRV, achieving 93% classification accuracy [48]. Similarly, Hermann et al. integrated heart-rate variability with EEG complexity, utilizing logistic regression to predict outcome with 90% accuracy [16]. While these studies enhance our understanding of brain–heart interactions, they still rely on fixed time intervals, limiting their ability to fully capture the evolving dynamics of recovery.

Table 3.

Comparative summary of predictive models for neurological prognostication

Research	Number of subjects	Feature set	Classifier	The most discriminative features	Best AUC (%)
[45]Zheng et al., IEEE Transactions (2021)	1038 cardiac arrest patients	Longitudinal EEG features, including power spectral density (PSD), coherence, and dynamic changes in EEG patterns over time	Deep learning (RNN)	Longitudinal EEG patterns during recovery phase	88.0
[46]MJ Kim et al., Scientific Reports (2022)	54 cardiac arrest survivors	Quantitative EEG biomarkers: power spectral density (PSD), event-related spectral perturbation (ERSP), and spectral entropy (SE) across multiple frequency bands	Linear mixed model	Alpha-power in EEG	90.3
[47]MC Tjepkema-Cloostermans et al., Critical Care Medicine (2017)	283 comatose patients	EEG-based Cerebral Recovery Index (CRI): phase-amplitude coupling, coherence, alpha band power, and other integrative metrics	Random forest classifier	Integrative quantitative EEG measures	92.0
[48]FH Krones et al., arXiv Preprint (2024)	300 comatose patients	Multimodal features combining EEG (functional connectivity, phase synchrony) and autonomic nervous system (HRV, baroreflex sensitivity) metrics	Multimodal deep learning	Combined EEG and autonomic nervous system features	93.0
[16]B Hermann et al., Wiley Online Library (2024)	181 ICU patients post-cardiac arrest	Brain–heart coupling metrics: heart-rate variability (time-domain and frequency-domain analysis) and EEG complexity measures	Logistic regression	Heart-rate variability combined with EEG complexity	90
Present Study	277 Cardiac arrest patients	BHC and HEP	SVM and LR	Brain–heart interplay and Heartbeat-evoked potentials	98.0

Reported performance metrics are provided for contextual reference only and are not directly comparable due to differences in patient cohorts, feature extractionmethods, and validation strategies

In summary, our findings show that incorporating dynamic brain–heart coupling into prognostic models enhances existing EEG and HRV-based approaches by capturing the temporal evolution of cortical-autonomic interactions, offering more precise insight into post-cardiac arrest recovery.

Limitations and future directions

Although this study provides new insights into the temporal evolution of BHC after cardiac arrest, several limitations should be acknowledged. First, the predictive models were developed and evaluated within the same cohort. While stratified cross-validation was applied to mitigate overfitting, this approach may still limit the generalizability of the findings.Second, BHC and HEP metrics were derived from discrete 5-min EEG/ECG segments with minimal artifacts. While this improves signal quality, it may fail to capture transient or continuous fluctuations in brain–autonomic interactions.Additionally, respiratory activity and cerebral hemodynamic parameters were not available in the dataset; given their known influence on autonomic and cortical dynamics, the absence of these physiological signals may restrict the interpretability of the coupling metrics.

Our findings suggest that BHC and HEP can operationalize the brain–heart axis as a quantitative, time‑resolved biomarker of post‑cardiac arrest recovery, rather than treating EEG and ECG as independent signals. In this cohort, models incorporating BHC (with or without HEP) outperformed those using only EEG or only HRV, indicating that indices of brain-autonomic integration provide additional outcome‑relevant information beyond traditional single‑signal measures. Notably, patients with good outcome exhibited a clearer and more sustained recovery of delta band BHC between approximately 60 and 75 h after ROSC, suggesting that this epoch may represent a critical window for incorporating BHC‑based metrics into multimodal neuroprognostication. In a real‑time ICU workflow, these metrics could be computed automatically from routinely acquired EEG and ECG in rolling 5‑minute windows and visualized as trajectories of frequency‑specific coupling strength and complexity. The preliminary thresholds identified here are exploratory and not suitable for stand‑alone decision‑making, but they illustrate how BHC and HEP based features could serve as mechanistically interpretable, time‑sensitive adjuncts to established neuroprognostic assessments, especially when clinical examination is limited.

Conclusion

This study shows that dynamic BHC and HEPs have potential as time-sensitive biomarkers associated with neurological outcome in comatose patients after cardiac arrest. Delta- and theta-band coupling changes were most prominent in patients with Good outcome. Integrating BHC and HEP features, particularly around 70 h post-ROSC, yielded high classification performance in this cohort (AUC = 0.98). These findings suggest that BHC and HEP may serve as complementary, exploratory tools for neuroprognostication, rather than standalone clinical biomarkers, pending validation in independent cohorts.

Author contributions

Yanxiang Niu and Ruxin Tan contributed equally to this work, including study conception, data analysis, and manuscript drafting. Xin Chen performed computational modeling and statistical analysis. Jianqi Fan and Ziquan Liu contributed to data preprocessing and machine learning implementation. Xiangyan Meng and Yanqing Liu assisted with literature review and manuscript editing. Lu Lu and Zongya Zhao provided supervision, methodology design, and manuscript revision. Haojun Fan led the overall study design, coordinated the research efforts, and critically revised the manuscript.

Funding

This study was supported by the National Key R&D Program of China, Project No. 2024YFC3016604.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

This study was approved by the Institutional Review Boards (IRBs) of all participating institutions, including Partners Healthcare in the United States (#2013P001024). Given the retrospective nature of the study and the use of anonymized data, informed consent was waived.

Consent for publication

Not applicable. The manuscript does not include individual person’s data requiring consent for publication.

Competing interests

The authors declare no competing interests.