Cross-modal synchronization of EEG and ECG reveals hidden signatures of recovery in traumatic brain injury

Abstract

Accurate assessment of traumatic brain injury (TBI) is critical for customization of neurorehabilitation treatments and clinical decision-making. Existing monitoring approaches either rely on subjective evaluation or isolated physiological signals, limiting real-time responsiveness and multimodal insight. This study introduces a novel framework integrating electroencephalography (EEG) and electrocardiography (ECG) to explore heart-brain synchronization as a biomarker for neurological state in TBI patients. We first define a synchronization metric using EEG delta power and heart rate variability (HRV), capturing both the degree and direction of synchronization. A two-stage contrastive learning approach is then proposed: Clinically Consistent Contrastive Learning (CCCL) leverages clinical metrics to guide positive sample selection, while Multimodal Heart-brain Contrastive Learning (MHCL) aligns synchronization features with clinical outcomes. Applied to long-term ICU recordings, the proposed approach identifies distinct synchronization patterns associated with recovery trajectories. Although the sample size (N=11) is expected to be extended, this work offers an exploratory, proof-of-concept demonstration of heart-brain synchronization as a potential real-time biomarker for neurophysiological recovery in severe TBI.

Introduction

The rising incidence of traumatic brain injury(TBI) [1–3] has created a need for accurate and timely diagnostic approaches. For patients with severe injuries, accurate assessment is essential to guide clinical interventions that may support recovery and improve long-term outcomes. However, monitoring changes in patients’ states remains a significant challenge in routine clinical settings. There is a clear need for assessment methods that can reliably track patient condition, motivating research into more objective and continuous monitoring strategies.

Current assessment methods largely depend on subjective assessments [4–6] or isolated physiological signals [7–13]. While both approaches have made substantial contributions, each of them has some limitations. For subjective assessment methods, they depend heavily on the clinician’s judgment and may fail to detect subtle or evolving changes in consciousness. Furthermore, such evaluations are typically performed at fixed intervals, limiting their ability to capture real-time changes. On the other hand, isolated physiological assessment methods often focus primarily on neural signals in the head, which neglects the potential contributions of other systemic physiological changes. Recently, more studies [14–16] have proposed the treatment of both the brain and the heart. It is because an increasing number of studies have found that the role of heart in patients with TBI cannot be ignored over the same time period.

Meanwhile, there are also some studies [17–23] that discuss the interactions between the brain and the heart. It is worth noting that a study [21] on the interoceptive prediction coding model has received a lot of attention. This model posits that interoception, the brain’s perception of internal bodily states, is not purely driven by bottom-up sensory input, but also by top-down predictions originating from the autonomic motor cortex. These predictions are constantly updated through the correction of interoceptive prediction errors. This study has inspired more investigation [17, 22, 23] to heart-brain interactions in terms of dynamic predictive coding. Such theoretical insights motivate further investigation into heart-brain interactions in TBI.

Despite these advances, few studies have directly examined the temporal synchronization between brain and heart signals or its relevance to patient status. Understanding this heart-brain synchrony may offer a novel perspective for patient monitoring, potentially serving as an objective and continuous indicator of neurological state. However, the clinical significance of such synchronization must be rigorously validated. Different patient may exhibit distinct synchronization patterns, emphasizing the need to explore how this coordination varies across clinical profiles. This relies on feature extraction from clinical multimodal data.

There have been many studies [24–28] demonstrating the effectiveness of deep learning methods for feature extraction in clinical multimodal data. In particular, contrastive learning [29–31] has emerged as a powerful approach for discovering potential relationships in data by constructing positive pairs and negative pairs. It is especially well-suited to clinical practice, where high-quality labeled data are limited but large number of unlabeled data are available. A contrastive learning framework that integrates clinical metrics with heart-brain synchronization features in brain-injured patients is required. This approach enables us to explore the potential relationship between physiological synchrony and clinical status, even in the lack of real-time label.

Figure 1 illustrates the limitations mentioned above and summarizes the methods and results of this paper. The contributions of this paper are summarized as follows:

• We collected data on EEG,ECG and clinical metrics from patients with different TBI in the ICU and selected the data to build a high-quality TBI multimodal dataset. The patients were under continuous monitoring ranging from 12 to 90 h to capture sufficent physiological changes.

• We propose the heart-brain synchronization metric that are superior to other synchronization metrics in clinical outcome mapping and is used to optimize the model.

• We propose a pre-trained model optimized with clinical metrics(GCS, clinical outcome) that selects positive samples carrying similar clinical information.

• We propose a contrastive learning model optimized with heart-brain synchronization metrics and clinically consistent positive samples, which exploratorily explains the relationship between heart-brain synchronization and clinical outcome in brain-injured patients.

Fig. 1.

Graphical summary of this study. ’Motivation’ outlines the critical need for accurate assessment of patients with TBI and highlights the limitations of current evaluation methods. Given the growing recognition of heart-brain interactions in TBI, we propose investigating heart-brain synchronization as a potential clinical marker. ’Method’ details the approach used to quantify heart-brain synchronization and how these metrics were mapped to clinical outcomes. It also introduces the contrastive learning framework used to extract and integrate relevant features. ’Results’ presents how the proposed MHCL model successfully identified distinct patient clusters and demonstrated its potential to infer recovery based on physiological data

Related works

The growing number of patients with TBI has increased the need for accurate and timely diagnosis. In particular, for patients with severe injuries, a correct diagnosis can guide clinical interventions that may promote recovery and improve long-term outcomes. Understanding changes in brain function requires continuous monitoring, an approach that is challenging in clinical settings. Most current assessment methods rely on subjective evaluations or physiological signals.

Subjective assessment methods

Subjective methods typically use clinical scales to rate a patient’s neurological status. For instance, Shukla et al. [4] reviewed various outcome scales and their application across different patient populations. Castaño Monsalve et al. [5] examined widely used neurobehavioral scales for assessing non-cognitive symptoms in TBI, while von Steinbüchel et al. [6] introduced the QOLIBRI, a tool to evaluate health-related quality of life after TBI and to monitor rehabilitation progress.

Despite their utility, these scales have limitations. They rely heavily on the clinician’s judgment and may overlook subtle changes in consciousness. Furthermore, such evaluations are typically performed at fixed intervals, limiting their ability to capture real-time changes. These constraints highlight the need for objective, continuous, and sensitive monitoring methods–leading to increased interest in physiological signal-based assessments.

Physiological signal assessment methods

Many current approaches assess TBI using physiological signals, typically focusing on a single domain such as brain activity or oxygen levels. EEG-based methods are common: Nuwer et al. [7] reviewed the use of EEG and quantitative EEG in mild TBI, while Perera et al. [8] analyzed EEG patterns to predict outcomes in comatose patients after hypoxic injury. Tewarie et al. [9] used early EEG markers to forecast recovery in moderate to severe TBI cases.

Other study explores additional physiological indicators. Hansen et al. [10] investigated oxygen supplementation and its link to in-hospital mortality in TBI patients, and Roldán et al. [11] examined near-infrared spectroscopy (NIRS) as a non-invasive technique for brain monitoring. Imaging has also been studied: Benjamini et al. [12] highlighted the sensitivity of advanced MRI techniques in detecting diffuse axonal injury, and Currie et al. [13] discussed the prognostic value of imaging in TBI.

While these methods offer valuable insights, they often isolate the brain from other internal organ systems. Multimodal approaches combining different brain signals are emerging, but they largely remain focused on the head. Recently, more studies have proposed the treatment of both the brain and the heart. It is because an increasing number of studies have found that the heart changes in patients with TBI cannot be ignored. Ofer Havakuk et al. [14] reviewed the pathophysiology of TBI in heart failure and described its impact on the prognosis of patients. Hsueh Chen Lu et al. [15] believe that heart rate variability biofeedback training can enhance the integration of the central nervous system and the autonomic nervous system measured by neuropsychology in patients with mild TBI. Giacomo Coppalini et al. [16] described the clinical consequences and management of cardiac injury after TBI.

The proposal of the combined treatment plan for heart and brain is not only since the changes in the heart of patients with TBI cannot be ignored, but also because of the interaction between the heart and the brain. Denise Battaglini et al. [18] the pathophysiologic mechanisms involved in brain-heart interactions and their clinical outcomes for patients with AIS. Giovanna Maria DimitriIt et al. [19] evaluated mortality using intracranial pressure and heart rate in TBI patients. El-Menyar et al. [20] described brain-heart interactions in TBI patients by explaining the effects of beta-adrenergic blockers. It is worth mentioning that a study on interoception has received widespread attention. Lisa Feldman Barrett et al. [21] proposed Embodied Predictive Interoception Coding. They argued that interoception is not only bottom-up signal processing, but is driven by the autonomic motor cortex in the ungranular layer through prediction, limited by prediction error correction of the interoceptive signal. For example, abnormal interoceptive predictions (e.g., chronic high metabolic demand predictions) occur in depression leading to dysregulation of the HPA axis, an inflammatory state, and eventually depressive behavioral symptoms. This research has inspired more researchers to discuss the interaction between the heart and the brain and has provided explanations for the interoception of both. Tahnée Engelen et al. [17] discussed how the brain integrates and coordinates information in distributed regions, by means of oscillatory synchrony, predictive coding or multisensory integration. Sahib S. Khalsa et al. [22] provided molecular and cellular insights into interoceptive mechanisms between the heart and brain. Andrea Zaccaro et al. [23] thought that heart-brain interactions are regulated by respiration-related attentional modulation that exists between interoception and exteroception.

The predictive interoceptive coding model requires dynamic coordination between organs. Due to the existence of human self-awareness and external stimuli [17], each organ changes synchronously in a certain time window. Although some studies have considered both brain and heart metrics in TBI, few have explored their temporal synchronization or its relationship with patient state. Understanding this heart-brain synchrony could provide new insights into patient monitoring and recovery, potentially offering a continuous, objective indicator of clinical status.

Importantly, even if heart-brain synchronization is observed, its clinical relevance must be validated. Specifically, we intend to explore what role heart-brain synchronization can play in clinical assessment. Different patient conditions may exhibit resemble synchronization patterns, emphasizing the need to explore how this coordination varies across clinical profiles. Investigating these variations could pave the way for novel, real-time monitoring tools tailored to the complex physiology of brain-injured patients.

Deep learning for assessment methods

As mentioned above, identifying potential relationships in multimodal data is essential. Deep learning has become widely used for clinical multimodal data feature extraction [24–28].

However, clinical settings often suffer from a lack of labeled data. As noted, data from a single day cannot be categorized under a single label. Contrastive learning addresses this issue by focusing on clustering similar data points through positive and negative sample pairs. This approach is particularly effective in uncovering relationships within multimodal data without requiring labeled information. For instance, Pan et al. [29] applied a self-supervised contrastive learning model to EEG and ECG data, tackling the challenge of limited high-quality labeled data in clinical practice. Zhang et al. [30] utilized contrastive learning to predict clinical events with a limited number of positive training examples. Similarly, Kiran Kokilepersaud et al. [31] introduced a novel strategy for selecting positive and negative sets for contrastive learning of medical images, demonstrating significant improvements in biomarker detection.

Numerous studies have shown the effectiveness of contrastive learning in the absence of labeled data in clinical environments. This method holds great potential for further exploring the relationship between heart-brain synchronization and patient status assessment in brain-injured patients. We believe that contrastive learning and corresponding results will reveal hidden signatures of recovery in traumatic brain injury.

Methods and materials

In this section, we describe data collection and method foundation. We first used correlation to select the EEG channel that carried the most information and calculated its delta power. Then synchronization metrics were proposed by using delta power and HRV. In terms of this, heart-brain synchronization labels were constructed for subsequent contrastive learning. Furthermore, two contrastive learning modules are proposed, Clinically Consistent Contrastive Learning(CCCL) and Multimodal Heart-Brain Contrastive Learning(MHCL). Cross-modal Contrastive Learning uses clinical metrics (GCS scores and clinical outcomes) as an optimization direction to fuse clinical assessment information into the heart-brain feature space. As a pre-trained model, it helps multimodal mind-brain contrastive learning to select positive sample pairs that carry information about clinical metrics. Finally, MHCL uses heart-brain synchronization information as an optimization direction, describes the potential relationship between clinical metrics and heart-brain synchronization. The framework is shown in Fig. 2.

Fig. 2.

The heart-brain synchronization study framework. The proposed framework contains two stages with different supervision levels. (1) Clinically Consistent Contrastive Learning (CCCL) is a weakly supervised pretraining module. Although the contrastive objective itself is self-supervised, CCCL additionally incorporates coarse-grained clinical labels (GCS category and clinical outcome) to guide the organization of the latent feature space. (2) Multimodal Heart-Brain Contrastive Learning (MHCL) is subsequently trained without any clinical labels and is therefore self-supervised with respect to clinical annotations. MHCL relies only on synchronization-derived labels and positive pairs identified by CCCL

Dataset

All EEG and ECG recordings were obtained from continuous bedside monitoring of patients with traumatic brain injury (TBI) in the Department of Neurosurgery, West China Hospital, Sichuan University. To ensure data quality and clinical validity, we established the following inclusion criteria:

1. Age > 18 years, with TBI confirmed by CT or MRI;

2. Severe or extremely severe TBI (Glasgow Coma Scale, GCS 8 at admission);

3. Continuous and synchronized EEG-ECG recordings lasting 24 h, with no major interruptions;

4. Stable clinical condition during monitoring, excluding patients who underwent major interventions (e.g., electrical stimulation, acupuncture, or secondary craniotomy);

5. Complete clinical outcome data, including dynamic GCS changes and follow-up outcomes;

6. Good signal quality, free of significant artifacts, electrode detachment, or synchronization errors;

7. Informed consent obtained from legal guardians, with approval from the Ethics Committee of West China Hospital.

Exclusion criteria included:

1. Severe motion artifacts or electrode detachment causing signal discontinuity;

2. Cardiac arrhythmias, pacemaker implantation, or unstable baseline ECG activity;

3. Severe systemic diseases (e.g., acute respiratory failure, sepsis, or multiple organ failure) affecting autonomic rhythms;

4. EEG or ECG signal loss 20%;

5. Incomplete clinical or follow-up data.

After screening, a total of 11 patients with severe or extremely severe TBI were included. All cases were consecutively enrolled, with high-quality recordings of sufficient duration, covering various TBI subtypes (including cerebral contusion, subarachnoid hemorrhage, and both closed and open head injuries) and coma states (GCS 3-8).

It includes gender, age, diagnosis, GCS score, surgery or not, clinical outcome, EEG, ECG, etc. It is shown in Table 1. EEG is collected through eight channels, while ECG is a single channel. The patients were all in coma and under continuous monitoring ranging from 12 to 90 h.(Original EEG recordings: S1:24 h, S2:24 h, S3:24 h, S4:18 h, S5:12 h, S6:12 h, S7:90 h, S8:18 h, S9:18 h, S10:12 h, S11:24 h). Sampling rate is 500Hz. To ensure perfect alignment between the two modalities, we collected EEG data over a longer period of time. Within this period, ECG was recorded for a specific continuous duration (Original ECG recordings: S1:18 h, S2:18 h, S3:18 h, S4:12 h, S5:12 h, S6:6 h, S7:12 h, S8:12 h, S9:12 h, S10:6 h, S11:6 h). During preprocessing, we precisely extracted the corresponding EEG segment that temporally matched the ECG recording window. As the result, we finally chose 268 h of data to ensure a high level of temporal consistency and to guarantee the quality of the data. 134 h EEG data and time-consistent 134 h ECG data are effectively used by the contrastive learning framework.

Table 1.

Patient Information

	Gender	Age	Diagnosis	GCS	Surgery or not	Clinical Outcome
S1	Male	49	Severe Brain Injury	1T1	Y	Daily Care
S2	Male	43	Closed Brain Injury	1T4	Y	Death
S3	Male	84	Bilateral Frontal Parietal Temporal Subdural Effusions	4T4	Y	Daily Care
S4	Male	24	Severe Closed Brain Injury	1T1	N	Good Prognosis
S5	Male	60	Severe Brain Injury	2T3	Y	Persist Coma States
S6	Male	53	Severe Closed Brain Injury	1T1	N	Death
S7	Male	71	Brain Contusion	1T1	N	Death
S8	Female	63	Subarachnoid Hemorrhage	3T4	Y	Daily Care
S9	Female	45	Cerebral Hernia	1T1	Y	Good Prognosis
S10	Male	48	Injury From a Fall	3T1	N	Daily Care
S11	Male	36	Severe Brain Injury	1T1	Y	Persist Coma States

We eventually hope to construct a large multimodal TBI dataset. Furthermore, the data collection will continue in West China Hospital.

Multimodal synchronization metrics

Data preprocessing and selection

Preprocessing of EEG and ECG signals is necessary due to disturbances in the clinical environment as well as in the patient’s condition. The raw EEG data was filtered through a band-pass filter at 0.5–30 Hz and the ECG data at 0.04–0.15Hz to remove DC signals and high frequency noise. The filter consists of a low-pass filter and a high-pass filter. The bands contain most of the physiological electrical signal activity required for analysis. The data were downsampled to reduce size and visually inspected to remove artifacts and data segments with obvious manipulation errors. The artifacts should then be further removed by independent component analysis(ICA). Subsequent stages involved the removal of inappropriate time segments based on voltage extremes. The z-score standardization is performed on delta power and HRV in the following calculation of metrics.

Since patients are injured in different brain regions, analyzing EEG signals from different channels will lead to different results. Choosing a channel with more information is necessary. Brain network analysis as a topological method describes channels as nodes and relationships between channels as edges [32–34]. Inspired by these studies, we presume that a channel that is more related to other channels or has more connected edges in the brain network will carry more information.

Channel selection was not the focus of this study, therefore only correlations were used to construct the correlation matrix. We used the EEG signal to calculate the correlation coefficient between the data channels as

1

where and represent the EEG data of channel i and channel j. If =1, it means that the two signals are highly correlated with each other. r takes the value . Channels that are more highly correlated with other ones carry more information.

Synchronization metrics

After identifying the channels to be analyzed, we need to determine the metrics to be analyzed. qEEG was found to be effective in the analysis of TBI patients [7, 35, 36].We performed spectral analysis of the EEG after channel selection and analyzed the EEG activity in different frequency bands using Fourier transform.

2

x(t) denotes the EEG signals acquired through the channel selection. is the spectral function computed from the EEG signal, where is the frequency. Although there have been studies [37, 38] describing different frequency bands with different evaluation performance, against our data, the EEG frequency is almost distributed in the delta band. As a result, we will only discuss the delta band.

If we want to further explore possible heart-brain synchronization phenomena, we also need to quantify EEG activity in the major bands. At the same time, quantification of cardiac activity is also necessary. We calculated EEG band power by Eq. 3 as well as HRV to quantify heart-brain activity by Eq. 4.

3

4

RR is the interval of each R wave. We divided the time windows by minutes and calculated P and HRV. It is worth mentioning that some of the time windows are notably oscillating. We utilized this special activity to define synchronization metrics as

5

where represent the number of events for delta power oscillations. , and represent the delta power oscillating first, HRV oscillating first, and delta power and HRV oscillating simultaneously, respectively. + + . Only simultaneous oscillations within the lag interval are recorded as a single synchronization event N. represents the synchronization percentage(SP). The specific algorithm pseudo-code is presented in the supplementary materials.

Since the predictive interoceptive coding model is adjusted for prediction error, it requires a high level of heart-brain alignment. The interoceptive predictive coding framework provides a mechanistic rationale for expecting rapid interactions between cortical state fluctuations and cardiac/autonomic responses. This theoretical account is supported by empirical studies showing heartbeat-related cortical modulations (heartbeat-evoked potentials and related EEG-ECG covariations [39]). Therefore, using a compact lag window for event-based EEG-ECG synchronization is physiologically plausible and consistent with prior theoretical and empirical work.

However, we do not know whether this bottom-up correction is affected by TBI. Specifically, if there is a disruption in cortical-subcortical connectivity, coma patients are unable to effectively resolve prediction errors through higher cognition. The system oscillates in a cycle of ”error accumulation → autonomic modulation → temporary relief → accumulation again”. As a result, there may be a lag. Therefore, we relaxed the event detection criterion under the condition of high alignment required by the predictive interoceptive coding model, allowing a minimal delay to accommodate subtle physiological phase shifts.

Additionally, our approach is conceptually consistent with event-based heart-brain coupling paradigms [40]. In that work, EEG delta events were detected and HR/HRV fluctuations were examined around those events, suggesting that small temporal offsets are acceptable in physiological coupling analyses. Importantly, we did not tolerate large lag; only a very small lag was permitted.

Meanwhile, due to the self-regulation of the heart, there are more oscillations that do not involve synchronization. Bottom-up correction of prediction errors can result in HRV oscillations that are not always accompanied by delta oscillations. Delta oscillations are more likely to be accompanied by HRV oscillations, so the number of events with delta oscillations is used to assess the synchronization percentage.

Further explain the reasons for using only EEG features for normalization:

First, HRV signals exhibit complex yet stable oscillations over long timescales, reflecting multiple layers of self-regulatory physiological and psychological processes [41]. Due to this intrinsic complexity of interdependent cardiac regulatory systems, many HRV oscillations occur independently of heart-brain synchronization. A symmetric or bidirectional normalization would therefore introduce excessive non-synchronous components, reducing the specificity of the synchronization metric.

Second, in severe traumatic brain injury (TBI), HRV may display abrupt changes prior to clinical deterioration [42], which are not necessarily aligned with cortical delta events. Thus, directly using HRV temporal dynamics for event-based normalization could be misleading.

By contrast, EEG slow waves (delta activity) have been identified as a sensitive cortical biomarker for both diagnosis and prognosis in TBI [43]. Accordingly, we used delta events as the normalization reference to ensure that the synchronization metric primarily reflects cortex-driven heart-brain coupling.

In summary, normalization only based on EEG features is physiologically justified and methodologically robust in this context.

represents the synchronization direction(SD). Equal to 0 denotes full synchronization, greater than 0 denotes delta power oscillating first, and less than 0 denotes HRV oscillating first.

Synchronization metrics need to be mapped to clinical metrics, including, but not limited to, GCS scores, and clinical outcome. Due to the abnormal variability of physiologic data in TBI patients, Sync do not always map to clinical metrics. Specifically, the brainstem and thalamus may retain basal endosensory predictive mechanisms (e.g., maintenance of heartbeat, breathing rhythms) even in coma patients with loss of consciousness, and these regions interact with the heart via the vagus nerve to form a prediction-error loop. However, self-regulation of the heart or its coupling to breathing [44, 45] would lead to more abnormal HRV oscillations, especially in non-healthy individuals. Although we consider both synchronization percentage and synchronization direction in Sync, similar synchronization metrics do not always map to similar clinical metrics due to the abnormal oscillations present. As described in Sect. 2, we use contrastive learning to find potential relationships in multimodal data to complement the synchronization metrics.

Multimodal synchronization feature extractor architecture

To enable contrastive learning, we designed a multimodal feature extractor, as illustrated in Fig. 2. This extractor operates as a dual-stream architecture with two independent channels: one for delta power of EEG and the other for heart rate variability (HRV).

The EEG stream (Channel 1) includes both temporal and spatial convolutional layers to separately capture time-dependent and spatial features from the delta power signal. The temporal convolution module is designed to capture the local temporal dynamics of EEG signals. The EEG input is first transposed to align the time and channel dimensions, followed by a 1D convolution layer that extracts temporal features. An average pooling layer then performs temporal downsampling to obtain a compressed temporal representation, effectively modeling short-term temporal dependencies. The spatial convolution module aims to learn inter-channel spatial relationships. The features output by the temporal module are processed through two successive pointwise (11) convolution layers and an adaptive average pooling layer to unify the temporal dimension. This design enables linear transformations across feature channels, facilitating spatial integration of EEG information.

Given the input , where T is time and C is the number of channels, extracting temporal features helps capture rhythmic oscillations and patterns of neural synchrony with the time. Notably, we retained all eight EEG channels in the feature extractor, rather than limiting them based on channel selection. This decision reflects a distinction between synchronization analysis–where dominant channels are emphasized–and feature extraction, which benefits from integrating diverse channel information. The relative contribution of each channel in this context is intentionally left unconstrained.

In the HRV stream (Channel 2), the HRV sequence passes through an independent temporal convolution path to extract time-dependent patterns. The subsequent convolution layer increases the channel dimensionality, and an average pooling layer produces a stable HRV representation. This channel captures the temporal trends of HRV and provides physiological context for subsequent cross-modal interaction. The hyperparameters of model is shown in Supplementary Table S1.

Both feature streams are subsequently passed through a pooling layer, producing two vectors: (EEG features) and (ECG features). These vectors are then input into a cross-attention layer designed to model interdependencies between the two modalities. The cross-attention layer is defined [46] as

6

where , and . are learnable parameters. are the outputs of and through channel 1 and channel 2. In this model, the ECG-derived vector () serves as the query because the study discussed a possible bottom-up predictive interoception framework, where more stable signals (like ECG) are used to modulate more variable ones (like EEG). The attention matrix captures potential synchronization patterns between heart and brain activity. Finally, the resulting representation is passed through a projection head to map it into a shared heart-brain feature space:

7

where BN denotes batch normalization, and , are fully connected layers. This joint representation serves as the foundation for downstream contrastive learning and synchronization analysis.

Clinically consistent contrastive learning

The crucial point of contrastive learning is to construct positive sample pairs, and we constructed a pre-training model for selecting positive sample pairs that carry information about clinical metrics.

Corresponding to the given patient data, contains the delta power and . Data augmentation is applied to the sample to obtain → .This allows the model to accept more complex patient data to construct more robust pairs of positive and negative samples. Specifically, corresponding to each sample, we add Gaussian noise with the same size. The noise follows a Gaussian distribution with mean 0 and standard deviation 0.2.

The multimodal data and its augmented samples are combined as input (, ). The features (, ) were obtained after the encoder. (, ) are positive sample pairs since they are original and augmented data from the same patient at the same moment while other samples in the same batch are negative sample pairs [47]. We use the normalized temperature-scaled cross entropy loss, which is commonly employed in contrastive learning [47]. It defined as follows

8

where sim denotes the cosine similarity function. denotes the temperature parameter, which is used to control the relative scaling of the probability distributions for positive and negative pairs. and denote the latent feature vectors of the EEG-ECG pair and its augmented version, respectively, obtained from the multimodal encoder. represents other negative samples within the same mini-batch, and N denotes the total number of sample pairs in the batch. This loss follows the InfoNCE formulation commonly used in contrastive learning.

Despite considering the interactions of ECG and EEG and fusing their features, this approach is missing clinical information. Specifically, patients with large differences in clinical metrics may have similar data features at different moments. It would not naturally be grouped into positive sample pairs and would be clinically unexplained. It is necessary to take clinical metrics into account when computing our loss function. It is also the task of our pretrained CCCL.

We made label for the patients’ GCS scores, which was divided into four categories based on the score. However, as we described in Sect. 2, GCS scores are timed to be completed daily and do not provide real-time feedback on patient status. Clinical outcomes, on the other hand, will always reflect the result, regardless of changes in the patient’s status during the monitoring process. Specifically, GCS scores may change when not assessed, but clinical outcomes must remain the same.

We made label for the patients’ clinical outcomes. These represents death, persist coma states, daily care, and good prognosis, respectively. represents the j-th segment of data for the i-th patient and . It is worth mentioning that we defined all label for the same patient as the same value, which depends only on his or her clinical outcome. We change the output of the fully-connected layer to match the label information and compute the loss via cross-entropy as

9

where represents the features of the input through encoder which has changed full connectivity layer. We define the loss function of the model as

10

is the variable hyperparameter that represents the percentage of clinical information in the model. In this study, was set to 0.1, was set to 0.2. Finally, the features obtained by inputting y into the trained model. Label carrying features of clinical and physiological information are obtained by using Leiden clustering. Leiden is a community discovery algorithm for network or graph data. Its improves on the shortcoming that Louvain may find arbitrarily poorly connected communities by introducing a finer termination condition [48]. Record when the clustering results in 4 classes. Construct the similarity matrix using and record the other sample with the highest similarity for each sample .

It is important to clarify that GCS scores and clinical outcomes are not used as supervision labels in the final heart-brain synchronization model(MHCL). Because both of them are assessed intermittently and does not provide continuous feedback, we do not use it for synchronization supervision. Instead, they are employed only in the CCCL to guide pretraining via clinical stratification. The goal is to enhance the representation space by selecting clinically semantically consistent positive sample pairs of. MHCL is independently optimized using synchronization metrics labels, which are more temporally aligned with the input signals.

Multimodal heart-brain contrastive learning

CCCL embeds clinical metrics into the construction of sample pairs. It allows positive sample pairs to carry clinical information. Specifically, a positive sample pair (() ) is constructed from the samples selected by the pre-trained model. is the most similar sample selected based on feature similarity after training the CCCL model. We define the loss function for MHCL as

11

where N represents the number of samples, represents the output from input to Multimodal Synchronization Feature Extractor. As Eq. (8), we did not exclude all positive samples in the denominator. The reason for this is that even the data in the same positive sample pair does not always have the same heart-brain synchronization. It can force the similarity of the positive sample pairs to be not only higher than the negative samples, but also to be more significant among the whole samples to obtain a more robust learning. However, it just keeps the model from confusing the heart-brain synchronization information in the feature space. We still need the input of heart-brain synchronization information during the loss calculation.

Combining Sync with GCS and clinical outcome, we presents heart-brain labels. Specifically, , , , and others, respectively. We give the reason in section EXPERIMENT. We still use the cross-entropy loss function to obtain . Finally, the loss function for MHCL is defined as

12

integrates clinical metrics with heart-brain synchronization information. was set to 0.75 to emphasize the contribution of the feature space.

Experiment

Data selection

Figure 3 presents correlation matrices derived from EEG recordings of selected patients. Regardless of direction, correlations closer to ±1 tend to convey more clinically relevant information. Based on this observation, we calculated the mean absolute correlation value for each EEG channel and identified the channel with the highest average as the most informative.

Fig. 3.

Correlation matrix and corresponding GCS scores. Patients with the highest and lowest GCS scores were selected.1T1 patients were two different terrible clinical outcome (S6,S11). A clear pattern: patients with higher GCS scores exhibit stronger positive correlations, whereas those with lower scores show more negative correlations

Heart-brain synchronization metrics

Results for Sync

We calculated the delta power and heart rate variability (HRV) for each patient and visualized a representative case in Fig. 4. The figure illustrates that while synchronization exists between the two signals, it is not complete. As outlined in the Methods and Materials section, a temporal lag is present, and its direction varies across synchronization events. Sync defined in Sect. 3.2.2 was applied to all patients. As shown in Fig. 4, the HRV signal contains a higher proportion of oscillations that are not synchronized with the EEG, further supporting our design of the Sync metric. In other words, the delta power series has better robustness.

Fig. 4.

At the top are visualizations of delta power and HRV series, respectively. At the bottom is a synchronization visualization of the two. The figure at the bottom also shows three events: brain lag synchronization, full synchronization, and heart lag synchronization. PSD_4 denotes the delta power of the data of EEG channel 4. std_rr denotes HRV

We then applied the Sync calculation across all patients, with results summarized in Table 2. The synchronization percentage is defined as , while the directionality ratio is given by /. As seen in Table 2, patients with a Sync value of zero generally had better clinical outcomes or higher GCS scores. In contrast, patients with Sync values exceeding ±0.1 tended to exhibit poorer outcomes or lower GCS scores. These findings form the basis for assigning the heart-brain label .

Table 2.

Synchronization Information

	GCS	Clinical Outcome	SP (%)	SD	Sync
S1	1T1	Daily Care	84.6	0.33	− 0.938
S2	1T4	Death	56.2	7.00	1.094
S3	4T4	Daily Care	41.7	1.00	0
S4	1T1	Good Prognosis	80.0	1.00	0
S5	2T3	Persist Coma States	58.8	0.50	− 0.408
S6	1T1	Death	36.8	2.00	0.255
S7	1T1	Death	52.4	0.65	− 0.226
S8	3T4	Daily Care	39.7	0.80	− 0.089
S9	1T1	Good Prognosis	76.3	1.00	0
S10	3T1	Daily Care	83.3	2.00	0.577
S11	1T1	Persist Coma States	87.5	2.00	0.607

One exception is patient S10, whose metrics deviate from this pattern. Although higher synchronization percentages and brain-leading synchronization directions are more frequently associated with favorable clinical outcomes, the exception cannot be avoided. The observed associations between synchronization and clinical metrics suggest a potential link, but one that warrants further investigation to understand its underlying mechanisms.

Comparison with other synchronization metrics

Sync also needs to be compared with the other metrics for assessing heart-brain synchronization. We chose to perform the following methods on EEG delta power and HRV:

• Phase synchronization analysis (PLV) which can assess synchronization by comparing the phase difference between two signals. Larger values represent better phase synchronization.

• Pearson’s correlation(P_cor) whcih can assess global synchronization by quantifying the co-variation of the two signals through time. Larger values represent better correlation.

• Mutual information(MI) which can capture the relationship between two signals, including synchronization, through joint and marginal probability distributions. Larger values represent stronger dependence.

• Cross-correlation(C_cor) which can assess the similarity of two signals by searching for the best alignment position of the two signals. In this study the limiting lag is 1. Larger values represent better correlation.

• Granger causality(GC) which can assess the synchronous coupling between signals by describing the bidirectional causality transfer between them. Due to the proposed of predictive interoceptive prediction coding model, the brain’s prediction of organ persistence, we limited our direction to EEG-to-ECG prediction. Larger values represent stronger causality.

• Dynamic time warping (DTW) which can assess synchronization by nonlinearly aligning two signals to search for their minimum distance. The smaller distance represents the better synchronization.

The result is shown in Table 3. Although each of these metrics assesses the synchronization of the two signals in different ways, the clinical mapping is poor. The only pattern is that the three patients with bad clinical outcomes, S6,S7,S11, have negative values for both P_cor and C_cor. Although Sync is limited by event detection and does not fully map to clinical metrics, it has a more meaningful interval division than other metrics. And the effectiveness of Sync as a label is demonstrated in the ablation study below. Futhermore, we observed that Sync consistently showed clearer separability between patients with different clinical outcomes. This suggests a stronger potential of Sync to reflect clinically relevant physiological states.

Table 3.

Comparison with Other Synchronization Metrics

	PLV	P_cor	MI	C_cor	GC	DTW	Sync
S1	0.37	0.12	0	127.50	3.35	455.17	− 0.938
S2	0.34	0.03	0.0007	159.07	18.37	388.73	1.094
S3	0.12	0.05	0.0014	56.21	0.01	732.92	0
S4	0.61	0.03	0	129.19	0.31	223.33	0
S5	0.63	0.13	0.007	94.11	4.19	300.55	− 0.408
S6	0.08	− 0.02	0	− 1.43	0.03	298.13	0.255
S7	0.47	− 0.20	0.0575	− 143.35	25.82	895.61	− 0.226
S8	0.27	0.14	0.0929	99.01	1.44	476.06	− 0.089
S9	0.29	0.12	0.0052	252.61	250.88	379.88	0
S10	0.25	0.17	0	168.59	135.10	210.37	0.577
S11	0.10	− 0.08	0.0415	− 27.11	0.17	317.46	0.607

Clinically consistent contrastive learning

Given that synchronization metrics do not fully capture clinical outcomes, the goal of contrastive learning is to bridge this gap. To achieve this, we applied CCCL to identify positive samples that reflect clinical assessment information, which are then used in MHCL. Since CCCL is designed to group features based on clinical similarity, the resulting clusters should contain features that correspond to similar clinical states.

Figure 5 visualizes the hourly feature embeddings for each patient. While CCCL does not always group all of a patient’s data into a single cluster, this variation is expected due to fluctuations in the patient’s microstate with the time. The critical point is that the majority of data points within each cluster share the same clinical outcome, indicating that CCCL effectively captures relevant clinical information.

Fig. 5.

Visualization of the hourly data features for each patient. Each of the four clusters corresponds to one of four clinical outcomes

It is also notable that only patients S4 and S9 appear in the ’Good Prognosis’ cluster. This suggests that CCCL is particularly sensitive to detecting features associated with better outcomes.

Multimodal heart-brain contrastive learning

We employed the pre-trained CCCL model to obtain a feature vector for each data segment. Using cosine similarity, we constructed a similarity matrix and selected the most similar samples as positive pairs for contrastive training. Subsequently, MHCL was trained using these pairs, and the resulting feature is visualized in Fig. 6, with various label annotations provided for interpretability.

Fig. 6.

Visualization of MHCL output features with different labels. a: unsupervised clustering into three groups via the Leiden algorithm. b: clusters are labeled by clinical outcomes. c: further simplifies the classification by relabeling ”Daily Care” and ”Good Prognosis” as Better, and ”Death” and ”Persist Coma State” as Worse. d: clustering results based on the synchronization label , where Sync1, Sync2, and Sync3 represent distinct synchronization patterns. (e): label the clusters based on the presence or absence of heart-brain synchronization, as defined in Fig. 4. (f): combine the labels in (c) and (e)

Figure 6a, each cluster is indexed numerically. In Fig. 6b, Cluster 2 includes both ”Daily Care” and ”Good Prognosis” patients, with those receiving daily care exhibiting higher GCS scores. Cluster 0 contains a mix of patients with good prognosis and those in a coma state, while Cluster 1 contains fewer favorable outcomes. In Fig. 6c, although the differences between Clusters 0 and 1 are less distinct, the presence of favorable outcomes exclusively in Cluster 2 suggests it reflects a better clinical state. Notably, no patients in Cluster 2 had a poor prognosis.

Interestingly, in Fig. 6d, Cluster 2 now includes Sync3, which is associated with poor prognosis. This apparent discrepancy with Fig. 6b, c warrants further analysis. One key patient contributing to this conflict is S11. Although S11 is labeled as persist coma states in Fig. 6b, this label is omitted in (c) for three reasons: (1) MHCL grouped S11’s data with other Better patients, (2) the patient’s family was pursuing active treatment, and (3) persist coma states are clinically complex. For example, S5, also in a persist coma state, was not clustered with Better patients, suggesting that S11’s classification is influenced by treatment intensity and possible neurological improvement. Thus, MHCL may have captured exploratory patterns in S11’s data that could reflect latent signals of recovery not fully captured by traditional clinical labels.

In Fig. 6e, Cluster 2 primarily contains synchronized data, supporting the notion that synchronization may indicate better outcome. Cluster 0 contains fewer synchronized samples compared to Cluster 1. It is worth mentioning that in (f) we labeled the data of S11 in cluster 2. Similarly, only synchronization data is present in Cluster 2. Despite the fact that there are some Worse patients in there, it is still reasonable to assume that the synchronized data signify a better state for the patients. Because these data are closer to cluster 2. As a result, the exploratory analysis of heart-brain synchronization by MHCL may offer preliminary insights that are consistent with the interoceptive prediction coding model in brain-injured patients.

Prediction error signals are generated when the interoceptive system detects that the body state (e.g., heart rate, blood pressure) does not match the brain’s prediction. With impaired brain function, the inability to effectively minimize the error by acting (e.g., adjusting posture) or updating the prediction results in the accumulation of error signals in subcortical structures. This manifests as delta power oscillations(It is because the EEG frequency distribution in brain-injured patients is mainly focused on delta power.). Prediction error triggers the autonomic nervous system (especially the vagus nerve) to try to adjust physiological states to match the prediction. For example, the brain may predict ”need to keep heart rate steady”, but actual heart rate fluctuations trigger the error. By increasing parasympathetic activity, the vagus nerve tries to reduce the heart rate fluctuations and minimize the error. In TBI patients, the brain tries to maintain the stability of the internal environment through the most basic prediction-error cycle, but cortical deficits lead to inefficient regulation, which is manifested as synchronized oscillations of delta and HRV.

Therefore, the presence of significant heart-brain synchronization, particularly if captured in Cluster 2 by MHCL during monitoring, may indicate residual interoceptive regulatory function. This could serve as a marker of the brain’s remaining capacity for predictive coding ability, and thus may hold value as a prognostic indicator in patients with TBI.

Ablation study

Ablation study on MHCL

Then, we performed ablation study on different modules. Clustering performance evaluated by Calinski-Harabasz Index (CHI). CHI evaluates the quality of clustering by calculating the ratio of the inter-cluster scatter matrix to the intra-cluster scatter matrix.The larger the value of CHI, the better the clustering is and the more effective the clustering result is. Specifically, CHI integrates the separation between clusters and the compactness within clusters, which can reflect the quality of clustering well. It was calculated as

13

where K is the number of clusters, n is the number of samples in the data set, x is the sample in cluster G, is the central point of cluster , and c is the central point of the entire dataset. Larger values indicate better clustering and more valid clustering results.

Our clustering is accomplished through Leiden algorithm, which optimizing graph modularity and community connectivity. It does not assume Euclidean geometry, cluster centroids, or compact spherical clusters. Consequently, commonly used internal clustering indices such as the Silhouette Index and the Davies-Bouldin Index , which are defined in metric spaces and rely on distance-to-centroid or nearest-cluster assumptions, are not fully aligned with the Leiden framework. CHI, which evaluates between-cluster versus within-cluster dispersion at a global level, is more suitable for our exploratory analysis.

Meanwhile, we report graph modularity(Q) as the additional internal index. It is the objective function optimized by Leiden and is therefore a method-consistent internal quality measure for graph-based clustering. Q is calculated by dividing communities through the adjacency matrix [48].

As shown in Fig. 7, the complete MHCL has the best performance. Module B, the chosen by CCCL, contributes more to the clustering. Module A, on the other hand, proves its validity by outperforming C by A+C. Specifically, ablation study demonstrated the effectiveness of CCCL and heart-brain synchronization label.

Fig. 7.

Clustering results for different combinations of modules. Modules A, B, C represent ,, respectively. A+B+C is the complete MHCL, B+C is with the heart-brain synchronization optimization removed, A+C is with the clinical positive sample optimization removed, and C is with only the augmented sample computational loss retained

In addition, the related ablation experiment results appear in Supplementary Fig. S1, which are respectively obtained by both EEG and ECG, merely EEG or ECG. The results show that the simultaneous use of EEG/ECG can achieve the best interpretability and good clustering effect at the same time. Although the clustering effect seemed good in the results using only ECG, the clustering results were completely irrelevant to the clinical information. Specifically, it causes completely confusion among patients with different clinical outcomes and GCS scores, providing no reference at all.

Ablation study on the attention layer

We chose to assign ECG signals as the query in the attention layer because heart rate and heart rate variability (HRV) are typically more stable and thus serve as a more reliable “reference signal” compared to the more variable EEG signals.

However, we admit that while HRV is more stable, it is also more complex, with more frequent oscillations on the temporal axis. This added complexity is a factor that we had not fully considered in terms of heart-brain interaction.

Therefore, we conducted an ablation experiment comparing the performance of ECG-query, EEG-query, Bidirectional-query in the attention layer. The results was shown in Supplementary Fig. S2. The results demonstrated that using ECG as the query signal yielded better consistency in this dataset and task. Additionally, the ECG-query setup resulted in superior clustering performance, supporting its efficacy as the query signal in our model.

Comparison with other augmentation techniques

We conducted experiments comparing multiple augmentation strategies, including Gaussian noise, flip,time-shift, scale and freq, and reported both clustering quality (Calinski-Harabasz Index, CHI) and classification performance based on synchronization labels. These augmentation methods were described in studies [49, 50]. Supplementary Fig. S3 showed that, compared to the augmentation method adopted in the manuscript, alternative approaches yielded lower CHI, indicating poorer cluster separability. Although these augmentations retained partial interpretability, we prioritized methods that enhanced discriminability between outcome-related labels, as this eventually supports clearer prognostic differentiation.

Notably, while moderate Gaussian noise injection improved global feature robustness, excessive noise could disrupt the subtle temporal information. To validate this problem, we further examined clustering performance under different noise standard deviations. The results confirmed that our constraint on Gaussian noise amplitude is appropriate, balancing robustness enhancement with signal integrity preservation.

Discussion

Validity of data selection

To generate the topographic maps in Supplementary Fig. S4, we calculated the average voltage and average inter-channel correlation for each of the four patients, consistent with those presented in Fig. 3. In patients with TBI, localized damage often results in abnormal voltage fluctuations within specific brain regions. If a channel affected by such abnormal activity is selected, the resulting signal may reflect only localized damage or noise, rather than meaningful physiological patterns. Consequently, focusing on these channels could limit our ability to capture global changes.

In contrast, channels selected based on higher inter-channel correlation tended to show voltages within a normal range. This suggests that these channels are less affected by localized artifacts or noise and are more likely to carry global information about overall brain dynamics. Notably, for the second patient, the selected channel displayed larger voltage variability than in the other cases. This may be due to two factors: (1) while most channels exhibited positive correlations, channel 3 showed a distinct negative correlation; and (2) the patient underwent a left-sided contusion resection with hematoma removal, without bilateral surgical intervention. It is possible that channel 3 became physiologically distinct as a result of this localized treatment.

Another point worth noting is that the correlation may be contaminated by noise and volume conduction. It is necessary to compare with other metrics used for channel selection. We have incorporated and compared three additional approaches, as suggested in study [51]: Partial Correlation (PC), Phase Lag Index (PLI), and Granger Causality (GC). Our results which was shown in Supplementary Fig. S5 indicate that for patients with higher GCS scores, the channels selected by PC and Pearson correlation were consistent. For patients with lower GCS scores, discrepancies were observed, suggesting that correlation-based selection is less reliable in more severe TBI cases.

Moreover, we calculated each channel’s Signal-to-Noise Ratio (SNR) and weighted Phase Lag Index (wPLI) to quantify the influence of noise and volume conduction. The results showed that no single method yielded channels that were consistently least affected by both factors. It is worth noting that although the channels selected by the correlation coefficient may be contaminated by noise and volume conduction, for some patients, the channels selected by this method slightly better in terms of pollution indicators.

Our data collection is still ongoing. Although channel selection is not the main objective of this study, we are conducting subsequent multi-center prospective studies. It help us to evaluate the generalizability of these methods on larger datasets in future work to establish a more robust channel selection framework suitable for TBI patients.

Discussion of persist coma states

As we described above, S11 performs specially in the MHCL results and we have explained it. However, the specificity of the persist coma states made us focus on S5, which has the same clinical outcome as S11. The data of S5 never existed in cluster 2. We show data with heart-brain synchronization in S5 and compare it with data classified as Cluster 2 in S11 in Supplementary Fig. S6.

Compared to S5, the two data segments of S11 have lower frequency and have amplitudes between − 250 and 250 . Cluster 0 has a lower amplitude and cluster 1 has a higher amplitude. Abnormally high amplitudes may be due to epileptic episodes. It is worth noting that although additional data is not shown here, most of the data has a similar pattern. By querying the patient’s case information, perhaps it was the more positive interventions S11 received that gave him signs of recovery. Specifically, in the persist coma state, EEG data that has the prerequisite as below can be characterized similarly to those of a patient with a good prognosis. (1) voltage intervals of − 250–250 , (2) relatively higher frequency, (3) occurrence of heart-brain synchronization.

Supplement to MHCL

We provide a further supplement to Fig. 6d, specifically addressing the presence of the Sync3 label within Cluster 2. In addition to S11, Cluster 2 also includes data from S10. At the time of initial data labeling, S10’s clinical outcome was categorized as Daily Care. However, a recent follow-up revealed that S10 had returned to the hospital and no longer required daily care. Subsequent clinical evaluation reclassified the outcome as Good Prognosis.

While the Sync label does not fully capture the complexity of clinical outcomes, this observation provides exploratory evidence that MHCL’s feature extraction and clustering may capture meaningful patterns. Notably, S10 was not labeled as Good Prognosis during CCCL training, yet MHCL still grouped its data with other Good Prognosis patients. This may be attributed to its GCS score and the enhanced feature representation learned through MHCL.

In summary, Cluster 2 in Fig. 6 serves as an illustrative example of patients who may be showing positive recovery trajectories, based on preliminary observations. This includes both S10, who experienced clinical improvement post-discharge, and S11, who received intensified therapeutic intervention. These results highlight the potential of MHCL to uncover latent patterns in heart-brain dynamics that may precede or complement traditional clinical assessments.

Meanwhile, the clustering results show that higher interpretability and higher CHI do not necessarily coincide with higher Q. The original proposed method does not always yield the highest modularity, but it consistently yields the most interpretable structure and the highest CHI.

This observation highlights that two indices capture different aspects of structure. In this study, CHI and qualitative interpretability better reflect our intended goal: identifying non-random, recovery-aligned structure in the learned representation space, rather than maximizing graph partition optimality.

It is also important to supplement that we did not compare the performance of MHCL with other prognostic models in this study. The reason is that:

• The aim of this study was to integrate heart-brain synchronization with the clinical presentation of patients through MHCL and to suggest possible assistance to clinical monitoring, assessment and intervention. We explore physiological synchronization between EEG and ECG as a complementary monitoring signal, rather than to replace traditional prognostic models such as IMPACT or CRASH.

• In existing prognostic models for TBI [52], the most common predictors were GCS score, age, and pupillary reactivity. More variables not covered in this study, the neurologic examination, anatomical and physiological changes identified with CT and MRI and so on, were added to impove the performence of model [53–56]. As a result, comparability study is inconvenient to conduct without equal conditions. Future work will consider multimodal integration once sufficient structured clinical data are collected.

• Existing methods that focus on interval to assess patients. We aim to assess patient status in real-time by analyzing biosignals that can be continuously monitored. This study focuses on real-time synchronization, and differences in task objectives compared to other studies limit the ability to make model comparisons.

• Traditional models like IMPACT and CRASH rely on static admission data (e.g., age, GCS, and CT features) to provide a one-time short-term prognosis. Their predictions reflect the initial severity of injury but cannot capture physiological changes during ongoing treatment. In contrast, our proposed MHCL framework is based on continuous EEG/ECG recordings, allowing real-time updates of prognostic estimates as the patient’s physiological state evolves. For instance, in patient S10, the initial label was ’Daily Care’, but later follow-up confirmed ’Good Prognosis’. Compared with other patients, S10 received more treatment. While traditional tools would not adjust their predictions, MHCL detected subtle recovery-related changes during ongoing therapy and correctly anticipated clinical improvement.

Statistical analysis for sample size and sync

It is necessary to perform statistical tests to justify the sample size and assess whether Sync offers better separability than alternative metrics. We conducted a post-hoc power analysis based on the effect sizes of each Synchronization metric between groups with different clinical performences (good vs. poor). The good outcome group was defined as patients with favorable clinical recovery and a GCS score > 6 (n = 4), while the poor outcome group included the remaining patients (n = 7). The Sync index was categorized according to , since its interval representation is more meaningful than absolute values.

Meanwhile, we applied Mann–Whitney U test to compare synchronization metrics between groups with good and poor clinical performences, and reported effect sizes to quantify discriminative strength. The results is shown in Table 4.

Table 4.

Statistical Analysis of Different Synchronization Metrics on Clinical Performence

Metric	n_good	n_poor	p	Cohen’d	Cliff’s delta	Posthoc_power
Sync	4	7	0.0028	− 6.062	− 1	1
PLV	4	7	1	0.009	0	0.05
Pearson	4	7	0.6485	0.559	0.214	0.126
MI	4	7	0.6987	0.294	0.179	0.071
C_cor	4	7	0.4121	0.758	0.357	0.192
GC	4	7	0.6485	0.442	− 0.214	0.097
DTW	4	7	0.6485	0.195	0.214	0.059

The results showed that: (1) The Sync metric exhibited a very large effect size (Cohen’s d = − 6.06), yielding a statistical Posthoc_power of 1.0 with the current sample size, indicating robust significance; (2) In contrast, other metrics (PLV, Pearson, MI, GC, etc.) showed smaller effect sizes (Cohen’s d < 0.8) and lower Posthoc_power (0.05–0.19), suggesting that the current sample size may be insufficient to detect weaker effects.

Meanwhile, the results demonstrated that ’Sync’ significantly differentiated between good and poor clinical performences patients (p = 0.0028), with a very large effect size (Cohen’s d = − 6.06, Cliff’s delta = − 1), indicating strong signal discrimination within this exploratory cohort. In contrast, other commonly used synchronization or coupling measures (PLV, Pearson correlation, Mutual Information, cross-correlation, Granger causality, and DTW) did not show significant group differences (p > 0.4), and their effect sizes ranged from small to moderate. These findings support our decision to use Sync as the primary synchronization metric for this study. Nevertheless, we have also acknowledged that larger sample are required to further validate the robustness of this statistical distinction.

In addition, this study aims to explore the feasibility and potential clinical significance of combined EEG-ECG synchronous indicators in patients with severe TBI, which is inherently a preliminary methodological and proof-of-concept study.

While the large effect size and high post-hoc power of the ’Sync’ metric suggest a underlying signal within the observed data, we must emphasize that these statistical metrics are based solely on a small cohort () that is not large enough to discuss the generalizability. The limited sample size inevitably results in wider confidence intervals for the estimated effect sizes, indicating a lack of precision, and thus requiring these findings to be interpreted conservatively. A small cohort significantly restricts the generalizability and robustness of the resulting machine learning models, including our MHCL framework.

Therefore, our findings regarding distinct synchronization patterns and the predictive utility of the MHCL model must be interpreted conservatively and are explicitly framed as exploratory and preliminary. Specifically, any reported classification performance and observed effect sizes should be viewed as providing a proof-of-concept for the utility of heart-brain synchronization. It is necessary to validate the proposed method in a larger, independent and ideally multi-center cohort. Thus, extending the sample size to verify the wide applicability and predictive stability of the model is also the ongoing work.

The role of theta band

According to the fourier transform results, our EEG data were predominantly distributed within the delta band, which is expected in the coma phase. Theta-band analysis was not originally included because theta oscillations are more frequently discussed in mild TBI [57, 58].

However recent studies have highlighted the potential relevance of theta activity even in severe TBI patients. Following this suggestion, we additionally analyzed both the absolute theta power(ATP) and relative theta power(RTP), which have been validated as useful indicators in previous works [59–61]. The result was shown in Supplementary Table S2. To more clearly demonstrate the correlation between ATP and RTP in clinical perfermance, we simultaneously present the GCS score and clinical outcomes again in the table.

Our results showed that patients with higher GCS scores or partial recovery generally exhibited higher absolute theta power, whereas relative theta power displayed weaker clinical correlations. In summary, (1) EEG activity during coma is primarily dominated by the delta band; (2) Absolute theta power may serve as a potential indicator of consciousness recovery and prognosis. In future study, we will investigate the dynamic activity of theta-band to better track neural recovery pattern in TBI patients.

End-to-end MHCL monitoring

Accurate diagnosis plays a critical role in guiding clinical interventions for brain-injured patients, facilitating timely treatment, promoting recovery of consciousness, and improving overall prognosis. Real-time monitoring is essential to support dynamic clinical decision-making. However, continuous monitoring remains challenging in routine clinical practice. In this context, MHCL offers a promising solution, as it can identify patterns in patient data indicative of favorable recovery trajectories.

In its current form, our model is primarily designed to support early risk stratification and dynamic monitoring in patients with severe TBI. The combined EEG-ECG Synchronization metric can reveal early and sensitive signs of autonomic dysregulation or secondary brain deterioration, which may be more responsive than changes in conventional GCS scores. Such information could inform clinical decisions regarding sedation depth adjustment, optimization of hemodynamic management, and escalation of ICU monitoring intensity.

It is important to emphasize that the current Sync (EEG-ECG Synchronization) metric is not intended to serve as an independent prognostic marker or to directly determine treatment decisions. However, the framework (MHCL&Sync) already provides a practical basis for bedside assessment and monitoring optimization. Compared with traditional evaluations that rely on structural imaging or delayed neurological scoring, dynamic monitoring of EEG-ECG physiological Synchronization better reflects real-time pathological processes, potentially shifting TBI management from “delayed response” to “Intervene in advance”.

Importantly, our study may offer a potential framework for future real-time monitoring studies capable of continuously evaluating patients’ physiological recovery trajectories (although it needs further validation). As data collection expands, integration with bedside clinical systems could support early rehabilitation planning or intensity adjustment, particularly in cases where overt behavioral signs are delayed or ambiguous. For instance, patients identified with stable heart-brain synchronization (cluster 2) could be considered for early initiation of sensory stimulation or physical therapy. Meanwhile, if cluster 2 data appears after intervention (such as S10 and S11) , it may implies that the patient has recovered.

In parallel, we are actively collecting data from brain-injured patients in the ICU. These efforts aim to enhance the robustness and generalizability of MHCL, ultimately supporting its deployment in a more general and complex clinical setting. Although the findings highlight a potential association between EEG-ECG synchronization and recovery trajectories, the present study remains exploratory. The proposed framework is not intended for clinical deployment at this stage. Larger, prospective, and multi-center datasets will be necessary to validate its robustness, real-time applicability, and clinical interpretability.

Conclusion

This study presents an exploratory investigation into the role of heart-brain synchronization in patients with severe TBI. While not designed as a prognostic model, our findings reveal latent physiological patterns, particularly EEG-ECG synchronization, that may reflect residual interoceptive regulatory capacity. This insight offers a potential physiological basis to complement existing neurorehabilitation assessments.

In this work, we introduced a novel multimodal learning framework that integrates EEG and ECG signals to explore latent physiological signatures in patients with TBI. We proposed a heart-brain synchronization metric that captures coupling between HRV and delta power of EEG, and designed a two-stage contrastive learning approach, Clinically Consistent Contrastive Learning (CCCL) and Multimodal Heart-Brain Contrastive Learning (MHCL), to embed clinical relevance into the learned feature space. The observed clustering structure is consistent across repeated runs of the representation learning stage.

Our findings demonstrate that heart-brain synchronization patterns are associated with clinical outcomes and may provide additional insights into patient prognosis beyond traditional metrics such as GCS. The MHCL model enables unsupervised clustering that aligns with clinical outcomes. This suggests its potential as a real-time monitoring tool to support customization of neurorehabilitation treatments and clinical decision-makingin neurocritical care.

However, this work still has limitations:

• Sync still needs to be expressed more accurately, which requires not only improvements in the detection of events number, but also refinements in the expression of the synchronization metrics.

• CCCL and MHCL still need more samples for performance improvement and validation. For this, we still continue to collect data from TBI patients in ICU.

• Although we discussed clusters 0, 1, 2 with different labels, clusters 0 and 1 still need further clinical discussion.

For above limitations, our work still focuses on the early methodological investigation rather than a clinically ready system. The results provide preliminary evidence supporting the feasibility of examining heart-brain synchronization in severe TBI, but further large-scale validation is required before any clinical translation can be considered.

Future studies will explore this link between acute neurophysiological patterns and rehabilitation. Meanwhile, we will focus on expanding the dataset. It helps us to validate the proposed methods across multiple centers, perform systematic lag-window and event-window sensitivity analysis, conduct systematic stability analysis, investigate the dynamic activity of theta-band, and develop a real-time clinical decision support system based on MHCL.

Author contributions

LXL and THB contributed to the literature and study methods; CP, LZY and THB contributed to study selection and methodological quality assessment; LXL and YYZ contributed to data extraction and experiments; LXL, THB and CP contributed to study design and writing; LPC, SML and YJX contributed to data collection; LXY and LK contributed to the feasibility study and assessment of future work practices. All authors read and approved of the final manuscript.

Funding

The work has been financially supported by the Sichuan Science and Technology Programme (Grant Nos. 2023YFH0037 and 2024YFFK0033), Foundation of Sichuan Research Center of Applied Psychology (CSXL-25231), 135 project for disciplines of excellence, West China Hospital, Sichuan University (Grant Nos. ZYYC21004 and ZYJC21081), Key Research and Development Project of Science and Technology Department of Sichuan Province (2024YFFK0234).

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

This work involved human subjects or animals in its research. Approval of all ethical and experimental procedures and protocols was granted by the Biomedical Ethics Committee of West China Hospital, Sichuan University under Application 2023(1977), and were performed in line with the ethical standards of the World Medical Association (Declaration of Helsinki).

Competing interests

The authors declare no competing interests.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12984-025-01869-5.