Temporal dynamic brain-heart interaction alterations associated with suicidal ideation in major depressive disorder

Abstract

Background

Previous studies have indicated abnormal brain activation or heart rate variability in patients with major depressive disorder (MDD). Suicidal ideation, as one of the main concerns of MDD, is a serious public health problem. However, the interactions between brain and heart of MDD patients with and without suicidal ideation remain largely unknown.

Methods

In this study, resting-state EEG and ECG data were simultaneously collected from 42 healthy controls (HCs), 69 MDD patients with suicidal ideation (MDDSI), and 39 MDD patients without suicidal ideation (MDDNSI). We proposed a novel methodology for analyzing temporal dynamic brain-heart interactions (TD-BHI). Then, we calculated the correlations between abnormalities of TD-BHI and cognitive performance, as well as sleep quality. Finally, we assessed the ability of TD-BHI to distinguish MDDSI patients.

Results

We found that the TD-BHI values in the MDDSI group significantly increased in the parietal lobes and decreased in the right parietal lobes, compared with those in the HCs group. Additionally, the attention/vigilance score of the MATRICS consensus cognitive battery (MCCB) and the Pittsburgh sleep quality index scores were significantly correlated with the TD-BHI values in the MDDSI group. The ensemble learning model based on TD-BHI achieved an AUC of 0.825 in distinguishing MDDSI from MDDNSI patients.

Conclusion

Distinct abnormalities of TD-BHI may mediate suicidal ideation in MDD patients and can serve as a reliable biomarker for the diagnosis of MDDSI patients.

Introduction

Major depressive disorder (MDD) is a very common psychiatric condition that affects more than 350 million people around the world [1]. As the same time, suicide has also become a global problem, and approximately one million people commit suicide per year [2]. Depression and hopelessness are the most commonly cited risk factors for suicide ideation, attempt and death [3]. Previous studies have demonstrated that the memory of suicidal thoughts in MDD patients can become strengthened, thus lowering the threshold for suicide and potentially leading to suicide attempts [4]. Thus, early detection and timely intervention are crucial for MDD patients with suicidal ideation (MDDSI). However, the diagnosis of MDDSI rely on scales and questionnaires that needs the objective physiologic measures to improve the ability to identify suicidal ideation.

Several studies have used electroencephalography (EEG) to investigate band power and brain network alternations in depression subjects with suicidal ideation [5]. In the alpha frequencies, major depressive disorder patients with suicidal ideation have more activity compare to the ones with low suicidal ideation [6]. Furthermore, high gamma band may also serve as an identified marker for depressive patients with suicidality [7]. Brain-heart interactions govern the homeostatic regulation of interoception through anatomical and functional connections, which have been confirmed to play a pivotal role in cardiovascular diseases (CVD) [8, 9], emotional states [10, 11], behavioral controls [12], and neurological disorders [13, 14]. Depression is recognized as a disorder that involves the interaction of the brain, heart and blood [15, 16]. Furthermore, a meta-analysis revealed that abnormalities in brain structure and function are highly correlated with suicidal ideation [17, 18]. There was found that an increase in suicidal ideation among patients diagnosed with heart disease [19]. Brain and heart changes might be not only the cause but also the consequence of a depressive disorder or suicidal ideation [20]. Brain-heart interactions may be not only related to the depression, but also connected with the suicidal ideation. The mechanisms of how these interactions influence suicide or depression are still unclear. Moreover, no study has revealed the differences between healthy controls and MDD patients with or without suicidal ideation through simultaneous detection of central and autonomic nervous activity.

Accumulating evidences have shown that MDD patients experience impairments in various cognitive domains, such as working memory, processing speed, executive function, and attention [21, 22]. Some studies have pointed out that cognitive dysfunctions are more frequently detected in MDD patients with suicidal ideation [23, 24]. Additionally, there is an urgent need in this field to establish an indicator for identifying suicidal ideation.

Bidirectional information exchange between the brain and heart has been extensively studied and recognized as a complex and dynamic mechanism rather than a static input [25]. The temporal dynamic brain-heart interactions (TD-BHI) method proposed in this paper can more clearly describe the normal interaction patterns between the brain and heart. By controlling for other confounding variables, we attempt to explain the relationship between the TD-BHI and depression, as well as the presence or absence of suicidal ideation.

We hypothesized that MDD patients with suicidal ideation (MDDSI) would exhibit altered TD-BHI compared to MDD patients without suicidal ideation (MDDNSI) and healthy controls. In this study, MDD patients were categorized into MDDSI and MDDNSI groups based on the presence or absence of suicidal ideation. EEG and ECG signals were simultaneously recorded for all participants, and TD-BHI values were calculated (Fig. 1). Correlation analyses were performed to examine the associations between TD-BHI and clinical symptoms, including suicidal ideation severity. Finally, we evaluated the discriminative performance of TD-BHI in identifying MDDSI patients.

Fig. 1.

The calculation process of the TD-BHI. Dynamic EEG power spectral density (PSD) was estimated using short-time fourier transform (2s hanning window, 50% overlap) and integrated into five frequency bands (delta, theta, alpha, beta, and gamma). ECG-derived RR intervals were interpolated to an equidistant time series, and their PSD was computed to extract low-frequency and high-frequency components. Temporal dynamic brain-heart interactions (TD-BHI) was quantified by Pearson’s correlation coefficients (PCC) between dynamic EEG and ECG PSDs across corresponding frequency bands

Methods

Participants

One hundred and eight patients diagnosed with FeMDD and forty-two healthy controls were included in this study. All participants comprehended the study procedure and signed informed consent forms. The entire study process was reviewed and approved by the Ethics Committee of the Affiliated Brain Hospital of Guangzhou Medical University (Number AF/SC-02/02.1) and conducted in accordance with the most recent version of the Declaration of Helsinki (2013). The diagnosis of MDD was made by a trained clinician who conducted a structured clinical interview with all participants according to the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5).

The inclusion criteria for all patients were as follows: (1) aged between 16 and 40 years, (2) had a first-episode of MDD for less than 2 years, and (3) did not take any medication regularly or use medication for more than 2 weeks. HCs matched for age and education level were recruited from the general population through advertising. HCs with a history of psychiatric disorders, such as depressive disorder, bipolar disorder, or substance abuse/dependence, were excluded. Those with current or past significant medical and neurological illnesses were also excluded.

The Bech Scale for Suicide Ideation-Chinese Version (BSI-CV) was used to assess the severity of patients’ specific attitudes, behaviors, and plans related to suicide within the past week [26]. The total BSI-CV score ranges from 0 to 38, with higher scores reflecting a more severe risk of suicide. Item 4 and Item 5 (Have you had any active or passive thoughts of suicide in the past seven days?) assess the presence of active or passive suicidal thoughts. In this study, participants who select a rating of 1 or 2 for either statement are identified as having suicidal ideation [27–29]. The FeMDD patients were divided into two groups based on the presence or absence of suicidal ideation. The group with suicidal ideation was referred to as MDDSI (n = 69), while the group without suicidal ideation was referred to as MDDNSI (n = 39). Healthy control group participants were confirmed by a trained clinician to have had no suicidal ideation in the past seven days.

Clinical and cognitive evaluation

The severity of depressive symptoms was assessed using the Hamilton Depression Rating Scale (HAMD-17) [30]. Additionally, the score on HAMD-17 without including the suicide item (Item 3) was calculated. Sleep quality was assessed using the Pittsburgh sleep quality index (PSQI), a self-assessment scale consisting of 7 dimensions, with each dimension scored on a scale of 0–3, related to sleep quality. A higher total score indicates poorer sleep quality [31]. Cognitive performance of all participants was evaluated using the MCCB [32]. Five dimensions of the MCCB were selected based on previous studies [33, 34]: the speed of processing domain (SOP), AV, working memory (WM), verbal learning (VRB) and visual learning (VIS).

EEG and ECG data acquisition

The 32-channel EEG and a single-lead ECG (Neuracle, Inc.) connected through a synchronous control box were used to record physiological data from all participants. Resting-state physiological data were collected for 10 minutes and sampled at a rate of 1000 Hz. All participants were instructed to sit comfortably in a chair, close their eyes, stay awake, and try to clear their minds during the data acquisition procedure.

EEG and ECG data preprocessing

EEG data was preprocessed using the Python-MNE toolbox [35]. The preprocessing procedures included 1) bandpass filtering the data to 0.5–70 Hz and applying a notch filter at 50 Hz to remove power line noise, 2) baseline correction and reference of the data using REST [36], 3) resampling of the data to a sampling rate of 500 Hz to reduce the data length, 4) applying independent component analysis to remove other artifacts, and 5) removing abnormally fluctuating time-period signals.

ECG data was preprocessed using the MNE and NeuroKit2 toolboxes in Python. The preprocessing procedures included the following steps: 1) bandpass filtering of the data from 0.02 to 40 Hz, 2) baseline correction using the wavelet function, and 3) resampling of the data to a sampling rate of 500 Hz for alignment with the EEG data.

Calculation of the TD-BHI

The power spectral density (PSD) of EEG signals represents the energy distribution of the signals across the entire frequency spectrum and can be used to characterize the level of brain activation. As shown in Fig. 1, the dynamic PSD was calculated using a short-time Fourier transform (STFT) with a Hanning taper over a sliding time window of 2 seconds with a 50% overlap. The time-series data were then integrated across five frequency bands (delta: 1–4 Hz, theta: 4–8 Hz, alpha:8–13 Hz, beta: 13–30 Hz, gamma: 30–70 Hz).

The ECG signal was analyzed using a QRS complex detection algorithm to extract the RR intervals, which represent the HRV. Subsequently, cubic spline interpolation was used to obtain the equidistant time series of the RR intervals. Finally, the PSD of the RR intervals was calculated using the same method as that used for the EEG data, and the low-frequency PSD (0.04–0.15 Hz) and high-frequency PSD (0.15–0.4 Hz) were extracted from the data.

TD-BHI demonstrated the relationship between brain activation and heartbeat by calculating the Pearson’s correlation coefficient (PCC) between the dynamic PSD of EEG and ECG in various frequency bands. Specifically, there were 320 groups of TD-BHI that represented the relationships between the brain and heart in various brain regions and frequency bands. Considering the resting-state condition, the time-series PSD of the EEG and ECG for each participant was segmented into non-overlapping 20-second epochs. TD-BHI values were calculated for each epoch, and the mean TD-BHI across all epochs was then computed to obtain a subject-level coupling feature. This multi-epoch averaging strategy was applied to enhance the stability and reliability of the brain-heart interaction estimates.

Statistical analysis

The analysis separated the three groups into three distinct components as follows:

1. First, one-way analysis of variance (ANOVA) and chi-square tests were used to assess group differences in demographic characteristics (e.g., gender, age, level of education) among the three groups. The general linear model and ANOVA were used to compare clinical characteristics (e.g., SOP, AV, WM, VRB, VIS of the MCCB) among the three groups, with demographic factors serving as covariates. Bonferroni correction was applied for group-level multiple comparisons at the group level. Finally, the two-sample t-tests were used to examine differences in clinical characteristics between MDDNSI and MDDSI (e.g., HAMD-17, HAMD-17 without suicide, PSQI), while also including gender, age, and level of education as covariates. The aforementioned methods were carried out using SPSS 24.0 (IBM Corporation).

2. Initially, the Kruskal-Wallis test was used to compare TD-BHI values among the three groups, as the data consisted of continuous variables with a nonnormal distribution. Next, significant differences in the TD-BHI values were extracted to determine group-level differences using the Mann-Whitney U test.

3. A partial correlation analysis was performed to assess the relationship between the significant differences in the TD-BHI values from the previous step and cognitive performance (e.g., five dimensions of the MCCB) in the three groups. The correlation between significant differences in TD-BHI values and clinical characteristics (e.g., HAMD-17, HAMD-17 without suicide, PSQI) was also evaluated in two groups of MDD patients.

Machine learning analysis

A supervised ensemble learning framework was implemented to distinguish MDD vs. HCs and MDDNSI vs. MDDSI using TD-BHI features. To reduce dimensionality and prevent overfitting, embedded feature selection was performed using L1-regularized logistic regression, retaining features with non-zero coefficients. Model development was conducted within a stratified five-fold cross-validation framework to ensure balanced class distributions across folds. The classification framework was based on an ensemble learning strategy comprising three base classifiers: LightGBM, CatBoost, and logistic regression. For each classifier, hyperparameters were optimized using Bayesian optimization with the area under the receiver operating characteristic curve (AUC) as the objective function. The three optimized base learners were subsequently integrated using a soft-voting scheme, in which predicted class probabilities were combined to generate the final decision. Model performance was evaluated across cross-validation folds using AUC, accuracy, F1 score, and confusion matrix analysis. This ensemble-based pipeline integrates feature selection, Bayesian hyperparameter optimization, and multi-classifier fusion to enhance model robustness and generalization capability.

Results

Demographic and clinical characteristics

Table 1 shows the demographic and clinical characteristics of all subjects. In particular, demographically, the MDDNSI, MDDSI and HCs groups did not differ significantly in terms of gender ( = 5.03, p = 0.081, df = 2), but they differed significantly in age (F = 7.43, p = 0.001, df = 2) and level of education (F = 7.10, p = 0.001, df = 2). These differences should be considered covariates to eliminate potential confounding effects. Furthermore, the scores of HAMD-17 (t = 3.10, p = 0.081), HAMD-17 without Suicide (t = 0.61, p = 0.442), and PSQI (t = 0.01, p = 0.961) did not differ significantly between the MDDNSI group and the MDDSI group after controlling for the confounding effects of gender, age, and level of education.

Table 1.

The demographic and characteristics of participants

Characteristics mean(S.D.)	MDDNSI (n = 39)	MDDSI (n = 69)	HCs (n = 42)	F/t/(p)
Gender  (Female/male)	25/14	43/26	18/24	5.03 (0.081)
Age  Year	25.79 (4.61)	22.91 (4.95)	22.36 (2.86)	7.43 (0.001)
Level of education  Year	15.13 (2.67)	14.12 (2.42)	15.90 (2.37)	7.10 (0.001)
HAMD-17	20.79 (5.32)	22.87 (4.63)	-	3.10 (0.081)
HAMD-17 without Suicide	20.05 (5.36)	20.94 (4.25)	-	0.61 (0.442)
BSI-CV	1.53 (7.62)	11.61 (7.74)
PSQI	12.38 (3.60)	12.22 (3.42)	-	0.01 (0.961)

Abbreviations: MDDNSI, first-episode major depressive disorder patient without suicidal ideation; MDDSI, first-episode major depressive disorder patient with suicidal ideation; HCs, healthy controls; HAMD-17, Hamilton Depression Rating Scale; BSI-CV, Bech Scale for Suicide Ideation-Chinese Version; PSQI, pittsburgh sleep quality index

Additionally, we observed significant differences in the four dimensions of the MCCB among the three groups, with gender, age, and level of education as covariates (Table 2). Specifically, MCCB-SOP (F = 14.48, p < 0.001, df = 2), MCCB-AV (F = 5.68, p = 0.004, df = 2), MCCB-WM (F = 5.5, p = 0.005, df = 2), and MCCB-VRB (F = 8.04, p < 0.001, df = 2) were significantly different among the three groups. Moreover, post hoc analysis was used to identify significant differences in the MCCB-SOP, MCCB-AV, and MCCB-VRB scores between the MDD group and HCs group. There were no differences between the MDDNSI group and the MDDSI group in the three dimensions of the MCCB. Interestingly, a significant difference was observed only between the MDDSI group and the HCs group in MCCB-WM (t = 3.32, p = 0.003).

Table 2.

Cognitive performance of participants

MCCB mean (S.D.)	MDDNSI (①, n = 39)	MDDSI (②, n = 69)	HCs (③, n = 42)	F (p*)	Post-Hoc
						t (p*)
SOP	34.92 (11.74)	32.29 (10.32)	46.21 (11.3)	14.48 (<0.001)	③>①	3.50 (0.002)
					③>②	5.36 (<0.001)
					①>②	1.34 (0.547)
AV	36.14 (11.81)	34.65 (9.83)	42.14 (8.32)	5.68 (0.004)	③>①	2.74 (0.021)
					③>②	3.18 (0.005)
					①>②	0.02 (1)
WM	39.82 (13.79)	38.72 (11.37)	48.48 (11.33)	5.55 (0.005)	③>①	2.22 (0.085)
					③>②	3.32 (0.003)
					①>②	0.78 (1)
VRB	34.13 (9.33)	33.29 (10.73)	42.88 (8.24)	8.04 (<0.001)	③>①	2.47 (0.044)
					③>②	4.01 (<0.001)
					①>②	1.18 (0.72)
VIS	40.49 (8.27)	40.9 (8.50)	46.05 (7.21)	2.89 (0.059)	③>①	2.03 (0.133)
					③>②	2.24 (0.08)
					①<②	0.10 (1)

*Adjusted values were calculated with gender, age, and level of education as covariates

Abbreviations: MCCB, MATRICS consensus cognitive battery; MDDNSI, first-episode major depressive disorder patient without suicidal ideation; MDDSI, first-episode major depressive disorder patient with suicidal ideation; HCs, healthy controls; SOP, speed of processing domain; AV, attention/vigilance; WM, working memory; VRB, verbal learning; VIS, visual learning

Intergroup differences in the TD-BHI values

First, the significant differences in the TD-BHI between the three groups are illustrated in Fig. 2. Most of the significant relationships we observed included all-ECG (LF-ECG and HF-ECG) data with theta-EEG, alpha-EEG and beta-EEG data. In other words, the areas with significantly different values were distributed in the right superior frontal gyrus, the temporal poles, and the parietal lobes. Then, the significantly different TD-BHI values were extracted to observe the intergroup differences (Fig. 3). Compared with HCs, the MDDNSI group exhibited a decrease in TD-BHI values in the left parietal lobe (all-ECG with alpha-EEG and beta-EEG) and in the right superior frontal gyrus (all-ECG with beta-EEG). In addition to the above differences, the MDDSI group also had decreases in the temporal lobes (all-ECG with beta-EEG) and the right parietal lobe (all-ECG with beta-EEG), as well as increases in the TD-BHI values in the parietal lobes (all-ECG with theta-EEG) and the right temporal lobe (all-ECG with theta-EEG). However, there were no differences between the MDDNSI and MDDSI group.

Fig. 2.

Spatial distribution of brain regions with significant TD-BHI differences among MDDSI, MDDNSI, and HCs. Abbreviations: LF, low-frequency PSD of ECG; HF, high-frequency PSD of ECG; means the rhythm PSD of EEG; PSD, power spectral density; TD-BHI, temporal dynamic brain-heart interactions; MDDSI, first-episode major depressive disorder patient (MDD) with suicidal ideation; MDDNSI, MDD without suicidal ideation; HCs, healthy controls

Fig. 3.

Spatial distribution of brain regions with significant TD-BHI differences between HCs and MDDNSI and between HCs and MDDSI. Abbreviations: LF, low-frequency PSD of ECG; HF, high-frequency PSD of ECG; means the rhythm PSD of EEG; PSD, power spectral density; TD-BHI, temporal dynamic brain-heart interactions; MDDSI, first-episode major depressive disorder patient (MDD) with suicidal ideation; MDDNSI, MDD without suicidal ideation; HCs, healthy controls

Correlations between TD-BHI and clinical characteristics

Demographic characteristics (gender, age, and level of education) were included as covariates in the correlation procedure. As indicated in Table 3, the correlation analysis of the MDDNSI group revealed that the VRB of the MCCB was negatively correlated with TD-BHI in the left parietal lobe (all-ECG with theta-EEG) (r = −0.48, p = 0.002) and was positively correlated with PSQI (r = 0.45/0.46, p = 0.014). In the MDDSI group, we observed a negative correlation between the TD-BHI values and the VRB score of MCCB (r = −0.26/-0.25, p = 0.044/0.046) in the left temporal lobe (HF-ECG with beta-EEG) as well as the VIS score of the MCCB (r = −0.28, p = 0.02/0.023) in the left temporal lobe (all-ECG with beta-EEG) and PSQI (r = −0.22, p = 0.01/0.011) in the right parietal lobe (all-ECG with beta-EEG). Furthermore, we found a positive correlation between the TD-BHI values and the AV score of MCCB (r = 0.17, p = 0.033) in the middle parietal lobe (LF-ECG with theta-EEG). However, no significant relationships were found between the TD-BHI values and the scores of the other dimensions in the MCCB, HAMD-17, or HAMD-17 without suicide.

Table 3.

Partial correlation analysis between TD-BHI and clinical characteristics

Groups	Clinical characteristics	TD-BHI	r (p*)
MDDNSI	MCCB-VRB	low-θ-Pz	−0.48 (0.002)
		high-θ-Pz	−0.48 (0.002)
		high--CP5	−0.34 (0.028)
	PSQI	low-θ-CP6	0.45 (0.014)
		high-θ-CP6	0.46 (0.014)
MDDSI	MCCB-AV	low-θ-PZ	0.17 (0.033)
	MCCB-VRB	high-β-C4	−0.26 (0.044)
		high-β-CP2	−0.25 (0.046)
	MCCB-VIS	low-β-CP2	−0.28 (0.023)
		high-β-CP2	−0.28 (0.020)
	PSQI	low-β-P3	−0.22 (0.011)
		high-β-P3	−0.22 (0.010)

*Adjusted values were calculated with gender, age, and level of education as covariates

Abbreviations: TD-BHI, temporal dynamic brain-heart interactions; MDDNSI, first-episode major depressive disorder patient without suicidal ideation; MDDSI, first-episode major depressive disorder patient with suicidal ideation; MCCB, MATRICS consensus cognitive battery; VRB, verbal learning; AV, attention/vigilance; VIS, visual learning; PSQI, pittsburgh sleep quality index; low, low-frequency PSD of ECG; high, high-frequency PSD of ECG; means the PSD of various bands in EEG; P3, Pz, T7, CP5, P4 means the position of electrode; low--P3 means interaction between low-frequency PSD of ECG and the PSD of in EEG on the scalp position of P3

Classification performance based TD-BHI

We developed 2 classifiers to distinguish between MDD patients and HCs and between MDDNSI patients and MDDSI patients using the ensemble learning model, respectively. For classifier to distinguish between MDD patients and HCs, using 5-fold cross-validation, we achieved a a mean AUC of 0.815 ± 0.079, with an accuracy of 0.753 and an F1 score of 0.838 (Fig. 4A). For classifier to distinguish between MDDNSI patients and MDDSI patients, using 5-fold cross-validation, we achieved a mean AUC of 0.825 ± 0.067, with an accuracy of 0.806 and an F1 score of 0.851 (Fig. 4B).

Fig. 4.

Classification performance of ensemble learning model based on TD-BHI. Receiver operating characteristic (ROC) curves of 5 cross-validation folds and confusion matrices for distinguishing between MDD patients and HCs (A), and between MDDNSI and MDDSI patients (B). Abbreviations: TD-BHI, temporal dynamic brain-heart interactions; MDDSI, first-episode major depressive disorder patient (MDD) with suicidal ideation; MDDNSI, MDD without suicidal ideation; HCs, healthy controls; AUC, area under the curve

Discussion

To our knowledge, this is the first study to demonstrate significant associations between TD-BHI and cognitive performance, as well as sleep quality, in MDDSI. The main findings were as follows: (1) the MDDSI group exhibited increased theta-band TD-BHI in the parietal lobe and decreased beta-band TD-BHI in the right parietal lobe; (2) the MCCB AV score was significantly positively correlated with TD-BHI values in the MDDSI group; (3) PSQI scores were positively correlated with TD-BHI in the MDDNSI group but negatively correlated in the MDDSI group; and (4) TD-BHI demonstrated stable discriminative performance for identifying MDDSI.

First, the significant differences in the TD-BHI values among the three groups indicate that brain-heart interactions could serve as the important potential biomarkers to predict suicidal ideation in FeMDD patients. Although the MDDSI group did not show significant differences in TD-BHI values compared with the MDDNSI group, they had more abnormalities in TD-BHI values in the MDDSI group. This finding suggested more pronounced disruptions in brain-heart interactions in MDDSI group. The present study revealed a decrease in TD-BHI in the alpha and beta bands of the parietal lobes, as well as the right frontal lobe, in FeMDD patients. These findings are consistent with previous studies indicating reduced activity in the prefrontal cortex (PFC), specifically the ventromedial PFC and dorsolateral PFC, and decreased high-frequency HRV [37–40]. Another study also reported a lower global field power in several regions of MDD patients, including the frontal cortices, parietal regions, and temporal regions, by evaluating heartbeat-evoked potentials [38]. The interoceptive neural circuit is responsible for coordinating communication between the central nervous system (CNS) and the autonomic nervous system (ANS), playing a vital role in regulating emotional experiences [41]. As William James famously stated, ‘emotion is the feeling of bodily changes’ [42]. A decrease in TD-BHI inhibited effective emotional processing, further emphasizing the importance of disrupted BHI in affecting emotional experiences. Several studies have examined the correlations between interoceptive awareness and suicidal ideation, and the results have consistently demonstrated a negative correlation between them [43, 44]. Based on these findings, our results suggested that reduced interoceptive awareness, as indicated by the decreased TD-BHI values in the MDDSI group, may contribute to the development of suicidal ideation in FeMDD patients. Interestingly, we observed significant increases in the TD-BHI of the theta band in the middle and right parietal lobes in the MDDSI group, compared with those in the HCs, but these differences were not found in the MDDNSI group. Previous studies have indicated that emotion elicitation, which involves an increased level of attention, is related to the interactions between EEG oscillations in the theta band and vagal modulations in high-frequency power [45]. Furthermore, previous studies have shown that emotion-related stimuli, particularly those involving highly aggression-related content in films, can provoke the most significant increases in EEG responses within the theta band [46, 47]. The TD-BHI of the theta band could be specifically related to impulsive emotions or behaviors, which may mediate the development of suicidal ideation. Additionally, heightened levels of arousal resulting from persistent emotional stimulation were found to be correlated with increased BHI over the parietal electrodes [48]. Abnormalities in parietal networks were consistent findings in MDD patients [49]. A recent study on treatment-resistant depression with suicidal ideation has shown that theta burst stimulation is an effective approach for reducing suicidal ideation and anxiety symptoms [50]. These findings indicate that TD-BHI in the theta band of the parietal lobe is closely correlated with emotional processing sensitivity [51], which is correlated with suicidal ideation in MDD patients [52].

In addition, dysfunction of the prefrontal cortex (PFC) has been consistently reported in patients with MDD and is closely associated with suicidal ideation [53]. The PFC, particularly the dorsolateral PFC (DLPFC), orbitofrontal cortex (OFC), and frontopolar cortex (FPC), plays a central role in emotional regulation and cognitive control. Moreover, prefrontal regions are critically involved in vagally mediated heart rate variability regulation, highlighting their role in brain-autonomic interactions [54]. Although the present study primarily identified TD-BHI alterations in the parietal cortex, these findings should not be viewed as discrepant from the established prefrontal model of suicidality. Instead, both regions are key components of the fronto-parietal control network. From a network perspective, prefrontal dysfunction may reflect impaired top-down regulatory control over emotional and interoceptive processes [55, 56], whereas altered parietal TD-BHI may represent abnormal bottom-up integration of emotionally salient bodily signals [57, 58]. The interaction between impaired prefrontal top-down control and aberrant parietal brain–heart coupling dynamics may together disrupt the balance between cognitive regulation and interoceptive signaling, thereby increasing vulnerability to suicidal ideation in MDD.

When compared with the HCs, MDD patients demonstrated significantly lower scores of the MCCB (e.g., SOP, AV, WM, VRB), which is consistent with the findings of a previous study [59]. It was confirmed that the functional impairment observed in MDD patients was mediated by dysfunction within the frontal cortex and the middle temporal lobe [60]. In this study, we observed significant correlations between the TD-BHI and VRB as well as the VIS of the MCCB in both the MDDSI and MDDNSI groups. There were similar significant differences in verbal learning performance among patients with non-suicidal self-injury [61]. In fact, VRB deficits are symptoms of memory loss in MDD patients. The exciting thing is that TD-BHI can describe certain cognitive states of MDD patients, which is consistent with the findings of previous studies. Furthermore, we found a positive correlation between only the TD-BHI values and the AV (attention/vigilance) score of the MCCB in the MDDSI group. AV is related to impulsive behavior, and impulsiveness has long been recognized as a key factor in suicide [62]. On the other hand, the infralimbic cortex plays a crucial role in cognitive flexibility and is also involved in the presence of suicidal ideation [63]. Cognitive impairments were associated with suicidal ideation in psychosis patients [64]. Poorer cognitive function was consistently associated with an increased risk of suicidal ideation [65, 66]. In summary, cognitive impairments were correlated with the emergence of suicidal ideation, and TD-BHI could specifically characterize suicidal ideation in FeMDD patients.

Circadian rhythms are modulated by bidirectional BHI, and disruptions in this interaction could contribute to sleep disturbances, ultimately leading to the development of depressive disorders [67]. It is important to note that sleep plays a critical role in mental disorders [68]. As in previous studies, our study also revealed lower sleep quality in FeMDD patients [69, 70]. However, we found that the correlation between the TD-BHI values and the PSQI score was opposite in the MDDSI and MDDNSI groups. These intriguing findings suggest contrasting or contradictory correlations between BHI values and sleep quality in the MDD subgroups. In particular, previous studies have indicated that the correlation between sleep quality and suicidal ideation may have relatively independent effects in comorbid mental disorders [68, 71]. Based on our findings, we hypothesized that there may be distinct neural mechanisms underlying sleep disturbances in MDD patients with or without SI. A recent meta-analysis supported this hypothesis, indicating that total sleep time was significantly lower in the MDDSI group than in MDDNSI group [72]. These findings suggest that differences in sleep quality could serve as biomarkers for identifying MDD subgroups with or without SI.

The ensemble learning analysis demonstrated a moderate but stable discriminative performance in distinguishing MDD patients from healthy controls as well as differentiating MDD patients with and without suicidal ideation. Although the classification accuracy was not exceptionally high, this may partly reflect the relatively limited sample size and class imbalance. Notably, the model showed strong capability in identifying patients with suicidal ideation, suggesting that TD-BHI features capture clinically meaningful neurophysiological signatures associated with suicidal vulnerability. Compared with previous studies that applied support vector machine learning strategies to differentiate MDD patients with and without suicidal ideation (reported AUC ≈0.79) [73], the present study achieved a slightly higher AUC of 0.825. While direct comparisons should be interpreted cautiously due to differences in datasets and feature construction, this improvement may indicate that dynamic brain-heart coupling features provide complementary and clinically relevant information for suicide risk stratification. Given that emotion involves complex interactions between central and autonomic systems, TD-BHI may reflect adaptive modulation between cortical regulatory regions and vagal-mediated autonomic activity. The observed associations between TD-BHI and clinical symptoms further support its relevance to affective and cognitive dysfunction in MDD. Collectively, these findings suggest that TD-BHI may represent a promising physiological marker for identifying suicidal risk within depressive populations. Future studies with larger cohorts and advanced deep learning approaches may further improve classification performance by modeling the complex nonlinear dynamics between central and autonomic nervous systems [74–76].

Limitations

Although our results are significant and provide insight into the underlying mechanisms of brain-heart interactions in MDDSI, it is important to acknowledge several limitations within the study. First, this study was limited by an imbalance in sample size and the sex ratio among the three groups. Although no significant difference in gender ( = 5.03, p = 0.081, df = 2) was found, it is important to consider that previous studies have indicated the possible influence of gender on brain networks and clinical symptoms in MDD patients [77, 78]. Second, we observed significant differences in age and level of education among the three groups. These differences may introduce bias, which can only be addressed by controlling for age and level of education as covariates. Third, we did not explore the specific mechanisms through which TD-BHI affected suicidal ideation in MDD patients. Future studies should incorporate more comprehensive and multimodal datasets to further elucidate the underlying neural mechanisms by which brain-heart interactions contribute to suicidal ideation in patients with MDD [79]. Longitudinal designs with larger sample sizes and richer clinical, behavioral, and physiological information would provide a more complete understanding of the dynamic brain–heart coupling process. Such efforts may ultimately facilitate the development of more precise and effective early intervention strategies for individuals at high risk.

Conclusion

In conclusion, our results revealed more pronounced alterations in TD-BHI in the MDDSI group compared with the MDDNSI and HC groups. These findings suggest that dysfunction of interoceptive neural circuits may contribute to impaired emotion regulation and, consequently, to the emergence of suicidal ideation in patients with MDD. Significant associations between TD-BHI and clinical symptoms, including cognitive performance and sleep quality, further support its clinical relevance. Notably, TD-BHI demonstrated stable discriminative performance in distinguishing MDDSI from MDDNSI, highlighting its potential utility as an auxiliary biomarker for identifying suicidal ideation in MDD. Overall, this study provides novel evidence for the involvement of dynamic brain-heart interactions in suicidal ideation and underscores the need for further mechanistic investigations to advance clinical translation.

Abbreviations

BHI

Brain Heart Interaction

TD-BHI

Temporal Dynamic Brain Heart Interaction

MDD

Major Depressive Disorder

FeMDD

First-episode Drug-naïve Major Depressive Disorder

SI

Suicidal Ideation

EEG

Electroencephalogram

ECG

Electrocardiogram

MCCB

MATRICS Consensus Cognitive Battery

CVD

Cardiovascular diseases

DSM-5

Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition

BSI-CV

Bech Scale for Suicide Ideation-Chinese Version

HAMD-17

Hamilton Depression Rating Scale

PSQI

Pittsburgh Sleep Quality Index

Authors’ contributions

Zhaobo Li and Jing Zhou completed the data analysis and wrote the main manuscript; Yuanyuan Huang acquired the data and critically revised the manuscript; Fengchun Wu and Kai Wu revised the manuscript. All authors read and approved the final manuscript.

Funding

This work was supported by the National Key Research and Development Program of China (2023YFC2414500, 2023YFC2414504), announcement and Leading Science and Technical Foundation of Guangzhou Civil Affairs (GCAAL2022001 to G.Z.), Guangzhou Planned Project of Science and Technology (2023B04J0106); the National Natural Science Foundation of China (82271953, 82301688), the Key Research and Development Program of Guangdong (2023B0303020001, 2023B0303010003), the Natural Science Foundation of Guangdong Province (2024A1515013058), the Guangdong Key Laboratory of Battery Safety at Guangzhou Institute of Energy Testing (2019B121203008-KJ-2024-040/KJ-2024-041), the Science and Technology Program of Guangzhou (2025A03J3357), clinical Collaboration Project on Integrated Traditional Chinese and Western Medicine for Major and Difficult Diseases (Bipolar Disorder, ZDYN-2024-A-121), the Research capacity improvement project of Guangzhou Medical University (2024SRP200), and Guangzhou Key Clinical Specialty (Clinical Medical Research Institute).

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

Prior to enrollment, all participants demonstrated a comprehensive understanding of the study procedures and provided written informed consent. All participants were aged 16 years or older. For participants aged 16 years, written informed consent was obtained from both the participants themselves and their parents or legal guardians. For participants older than 16 years, parental or legal guardian consent was not required. The study protocol was reviewed and approved by the Institutional Review Board of the Affiliated Brain Hospital of Guangzhou Medical University (ethical approval number: AF/SC-02/02.1) and was conducted in accordance with the ethical standards of the 2013 revision of the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.