A pilot study of inter-regional phase amplitude coupling as comorbid depression biomarker in temporal lobe epilepsy

Abstract

Temporal lobe epilepsy (TLE) is associated with depression, but the neurobiological mechanisms underlying their bidirectional relationship remain unclear. Phase-amplitude coupling (PAC) refers to the interaction between the phase of low-frequency oscillations and the amplitude of high-frequency oscillations within the same brain signal, reflecting coordination between different brain rhythms for communication and cognitive processes. While prior studies, including our own, have explored PAC within individual brain regions, inter-regional PAC (irPAC) has not been studied in TLE and comorbid depression. We investigated irPAC in 17 TLE patients with electrodes implanted in the hippocampus, amygdala, and four cortical and subcortical regions (superior temporal, superior frontal, mesial orbitofrontal, and rostral anterior cingulate) in Papez circuit and default mode network. Modulation indices for directional brain region pairings were computed using a data-driven approach. Our analysis revealed a distinct delta–beta coupling signature that differentiated depressed from non-depressed TLE patients and correlated significantly with Beck Depression Inventory scores (Spearman’s ratio ~ 0.5), with similar correlation strengths observed for seizure frequency in the Epilepsy Monitoring Unit. These findings suggest that aberrant inter-regional oscillatory interactions within limbic–cortical circuits may contribute to depression in epilepsy. The delta-beta irPAC signal may represent depression-related neural signatures that are distinct from general epilepsy network dysfunction. This work provides new insights into the interplay between epilepsy and depression in Papez circuit and default mode network.

1. Introduction

Temporal lobe epilepsy (TLE) is the most common type of focal epilepsy and focal epilepsy accounts for 60 % of all epilepsy cases [1,2]. TLE patients have recurrent seizures and aberrant oscillations originating in the temporal lobes of the brain and they have significant comorbid depression [1,3,4]. Despite the well-documented bidirectional relationship between TLE and depression, the underlying neurobiological mechanisms remain poorly understood [4,5].

Among the various neural networks implicated in TLE and depression, the Papez circuit and the default mode network show significant overlap, particularly limbic structures[6–9]. The limbic-cortical dysregulation model suggests that disrupted communication between limbic structures and cortical regions plays a central role in depression [10,11]. Specifically, hippocampus (HIP) and amygdala (AMG) are key limbic structures with established roles in both epilepsy and depression pathophysiology [12–14]. AMG is critical for emotional processing and fear conditioning and has demonstrated hyperactivity and altered connectivity in major depressive disorder [12,15]. Similarly, HIP is essential for memory formation and has ample evidence of volumetric reductions and functional abnormalities in both depression and epilepsy [12,15–18].

We hypothesized that neural communication patterns between limbic and cortical regions would differ between TLE patients with high versus low depressive symptoms. To test this hypothesis, we examined how the hippocampus and amygdala interact with broader cortical networks in the Papez circuit and default mode network. We selected four specific cortical regions: superior temporal gyrus (STG), superior frontal gyrus (SFG), mesial orbitofrontal cortex (mOFC), and rostral anterior cingulate cortex (rACC) based on established anatomical and functional connections between the two types of brain regions in imaging studies using fMRI, diffuse tensor imaging and the Human Connectome Project atlas [6,19,20]. These regions were also prioritized due to their frequent inclusion in traditional temporal lobe epilepsy investigations, which ensured adequate data availability.

These cortical regions demonstrated established structural connectivity with AMG and HIP, exhibited well-documented functional abnormalities in depression, and have been implicated in the epileptic networks in TLE patients: The STG showed volume reduction in patients with major depressive disorder [21] and a cortico-cortical evoked potential study demonstrated effective connectivity between temporal neocortex and amygdala and hippocampus in TLE patients. The SFG has connections to limbic regions [22] and has been shown to have altered functional connectivity with limbic structures in TLE patients with depression and neuroticism [23]. The mOFC is shown to be connected to amygdala and hippocampus [24,25] while it is implicated in TLE and depression [26]. Finally, the rACC, which has been shown to have aberrant neural activity in depression [27], is also anatomically and functionally connected to amygdala and hippocampus [28].

Intracranial electroencephalography (iEEG) in TLE patients offers a unique opportunity to study these six brain regions interaction. Unlike traditional imaging methods, iEEG provides direct recording of neural activity from within the brain, allowing for more precise characterization of neural dynamics and connectivity patterns. Our group and others have applied phase-amplitude coupling (PAC) to iEEG data from epilepsy patients to study neural connectivity and dysfunction in epilepsy and related mood disorders [29,30]. PAC refers to the interaction between the phase of low-frequency neural oscillations and the amplitude of high-frequency oscillations [31]. Functionally, PAC is thought to be a mechanism for integrating neural activity and plays a role in various cognitive functions, including memory, attention, and language processing [32]. Disruptions in this communication mechanism have been implicated in several neurological and psychiatric disorders [32]. While local PAC is calculated using the low- and high-frequency signal within the same brain region, the inter-regional phase-amplitude coupling (irPAC) extends this concept to interactions between different brain regions by establishing the direction and strength of rhythmic neural transmission between different brain networks [33]. This multi-scale approach to examining both regional and inter-regional PAC is particularly important for understanding depression in epilepsy, as it captures complementary aspects of network dysfunction that may share mechanistic features with seizure propagation. While regional PAC reflects local circuit abnormalities within individual limbic and cortical nodes, irPAC reveals how pathological synchronization patterns propagate across the broader limbic-cortical network, similar to how seizure activity spreads through epileptic networks. This hierarchical analysis directly tests predictions of the limbic-cortical dysregulation model by examining both local node dysfunction and inter-node communication failures at electrophysiologically relevant timescales.

Importantly, both seizure activity and depression in TLE may share common underlying neural mechanisms involving disrupted oscillatory patterns and connectivity. Previous studies have demonstrated that seizure frequency and depression severity are correlated in TLE patients, suggesting overlapping pathophysiological substrates [34,35]. The limbic-cortical networks implicated in depression are also central to seizure generation and propagation in TLE, with the hippocampus and amygdala serving as common seizure foci [12,36]. Furthermore, aberrant neural oscillations have been associated with both ictal activity and interictal mood disturbances in epilepsy patients [37]. Therefore, examining the relationship between irPAC patterns and seizure frequency provides a crucial validation of whether the observed neural oscillatory changes represent general network dysfunction or are specifically related to mood symptomatology.

The aim of this study is to investigate the role of irPAC in brain regions commonly affected in TLE, and their relationship to depression severity as measured by Beck Depression Inventory (BDI) scores [38]. In contrast to our previous study of local PAC in both TLE and neocortical epilepsy patients, we utilized a unique dataset of 17 TLE patients with intracranial electrodes implanted in six regions mentioned above to study the inter-regional PAC involvement in depression and epilepsy networks [29]. We hypothesized that irPAC patterns across different frequency bands would differ between TLE patients with high and low Beck Depression Inventory (BDI) scores. To test this hypothesis, we employed a data-driven approach to identify significant frequency ranges, brain regions, and their corresponding phase and amplitude interactions. Specifically, we studied the irPAC of temporal lobe regions AMG and HIP and the four cortical regions to test predictions of the limbic-cortical dysregulation model for depression development [10,11,39]. As a validation step, we also examined the relationship between irPAC patterns and seizure frequency in the epilepsy monitoring unit (EMU) to determine whether observed neural oscillatory changes represent mood-specific alterations or broader network dysfunction common to epilepsy pathophysiology. This approach allows us to disentangle depression-specific neural mechanisms from general epilepsy-related network perturbations, providing greater specificity to our findings regarding the neurobiological basis of comorbid depression in TLE.

2. Materials and methods

2.1. Data acquisition

The consent process, subject enrollment, and collection of epilepsy and mood assessment data were conducted as described in a previous study [29]. For this study, data from temporal lobe epilepsy patients from our dataset were analyzed in order to examine a more homogenous group. Demographic data are shown in Table 1 [29]. BDI-II scores were divided into low, those < 19 (minimal or mild depression) and high, those > 19 (moderate to severe depression). Electrode placement is shown in Fig. 1.

Table 1.

Demographic information for the depressed (BDI > 19) and non-depressed (BDI ≤ 19) groups. The parentheses for ages, BDI and number of seizures in the EMU represent 95 % CI calculated by bootstrap (2000 repetitions with replacement).

	BDI > 19 Group	BDI ≤ 19 Group
Demographics
N (Sex)	2 Male, 6 Female	4 Male, 5 Female
Handedness	7 Right, 1 Ambidextrous	9 Right
Mean Age at Study	43.2 (35.5–50.9)	33.7 (28.7–38.0)
Mean Onset Age	29.9 (18.9–40.4)	19.1 (11.4–26.8)
Mean Duration	13.4 (7.2–23.4)	14.6 (7.9–25.6)
Study Laterality	2 BL, 3 L, 3 R	4 BL, 2 L, 3 R
Mean BDI	30.0 (25.5–35.4) *	9.6 (7.2–12.7) *
Number of seizures in the EMU	9.0 (6.1–12.0)	9.1 (5.4–14.7)
MRI Findings
Lesional (N)	4/8	3/9
Non-lesional (N)	4/8	6/9
Mesial temporal sclerosis (N)	1/8	1/9
Other structural lesions (N)	3/8	2/9
iEEG Seizure Onset Localization
Mesial temporal lobe epilepsy (N)	5/8	3/9
Neocortical temporal lobe epilepsy (N)	0/8	2/9
Multifocal/Other (N)	3/8	4/9
Seizure Onset Laterality
Left (N)	0/8	4/9
Right (N)	5/8	1/9
Bilateral/Diffuse/Multifocal (N)	3/8	4/9

^*represents statistically significant differences (p < 0.05) between groups. Wilcoxon rank sum test was used to compare between two groups. BDI = Beck Depression Inventory; EMU = Epilepsy Monitoring Unit; iEEG = intracranial electroencephalography;

Fig. 1.

Approximate electrode positions on an average brain for the 17 subjects. Each of the six brain regions (hippocampus (HIP) – light blue, amygdala (AMG) – purple, superior temporal gyrus (STG) – yellow, superior frontal gyrus (SFG) – dark blue, mesial orbitofrontal cortex (mOFC) – green, rostral anterior cingulate cortex (rACC) – red) is represented by a single color.

For regions with multiple electrode contacts, we selected the contact with the smallest depth (closest to gray matter) to ensure optimal signal quality, as described in our previous work [29]. In large cortical regions, such as the superior temporal gyrus, sampling variability was minimized by prioritizing contacts within 4 mm of gray matter and using standardized DKT40 atlas parcellation for consistent anatomical assignment.

2.2. irPAC analysis

Prior to PAC analysis, all iEEG data underwent standardized preprocessing procedures as established in our previous work [29]. Raw iEEG data were downsampled to 256 Hz to optimize computational efficiency while preserving relevant frequency content. Line noise removal was performed using a 100-order FIR filter (Hamming window, 400 ms) to eliminate 60 Hz and 120 Hz AC noise. The activity was then re-referenced to the average of all channels. Contacts with poor signal quality were excluded from analysis following established procedures from our previous work [29].

Regarding epileptic activity, ictal periods were avoided by ensuring all recordings were obtained at least 2 h before a seizure to prevent possible PAC perturbation. Interictal epileptiform discharges were intentionally included in the analysis as they may constitute part of the neural biomarker for emotion. High-frequency oscillations (HFOs) were naturally excluded from our analysis since we only examined phase-amplitude coupling up to 100 Hz, which is below the typical HFO frequency range (>100 Hz).

We extracted 5 min of raw iEEG data from each electrode of each patient. These 5-minute segments were at least 2 h before the first seizure to avoid possible PAC perturbation caused by the seizure [29,40]. These recordings were then paired across different brain regions, incorporating laterality and directionality of irPAC. Given our hypothesis that HIP and AMG will exhibit unique irPAC features with SFG, STG, mOFC, and rACC, we systematically paired HIP and AMG from each hemisphere with each of the cortical regions. This generated 72 unique directional pairings (6 regions × 6 regions × 2 directions), where each region pair was analyzed bidirectionally for phase-to-amplitude coupling (e.g., AMG phase to HIP amplitude, and HIP phase to AMG amplitude).

Next, irPAC was computed using a customized MATLAB code that utilizes EEGLab and its PAC toolbox [41]. irPAC was computed in 1-minute intervals, as prior studies utilizing irPAC in task-based states used 30-second intervals and 50–1000 s for a state-based investigation [18,33], while a previous study from our group used 1-minute intervals [29]. It was reasoned that an average of 5 min would provide a representative measure of resting-state activity. The upper bound of PAC analysis was limited to 5 min to optimize computational efficiency.

Subsequently, the irPAC of each 1-minute interval was averaged over the 5-minute period. Given previous findings that the left and right hemisphere PAC behave differently [29,30], we analyzed irPAC separately by laterality and then put them as two values for a subject if the subject had electrodes from both hemispheres before comparing high and low BDI groups.

2.3. Data-driven analysis

Our approach combined hypothesis-driven region selection with data-driven signal discovery within theoretically-motivated brain circuits. While we selected the six brain regions (HIP, AMG, SFG, STG, mOFC, and rACC) based on established literature implicating limbic-cortical circuits in depression, we systematically explored all possible frequency band combinations and directional connectivity patterns without a priori hypotheses about which specific frequency coupling pairs or directional region pairings would be relevant to depression in epilepsy.

There is no consensus on the optimal frequency bands for PAC analysis, and studies have reported varying frequency ranges with different names and physiological interpretations [31,42]. Additionally, the directionality of phase-to-amplitude coupling remains uncertain, and both increases and decreases in PAC have been linked to mood disorders [29,43].

To address these uncertainties, the ranges of 1–30 Hz for phase and 13–100 Hz for amplitude were selected for a data driven analysis. This ensure that we covered delta-beta [44–47], delta-gamma [48–50], theta-beta [51,52] and theta-gamma [16,53] PAC, which have been studied in previous studies for different neuropsychiatric conditions on intracranial and scalp EEGs. This data-driven approach allowed the signal to reveal which frequency coupling combinations and connectivity patterns were significantly associated with depression severity, rather than testing predetermined hypotheses about specific frequency bands or directional connections.

For irPAC computation, we used the Kullback-Leibler distance method as proposed by Tort et al [54]. This analysis produced a 30 × 87 modulation index (MI) matrix for each directional irPAC pairing, where each entry reflects PAC strength for a specific phase (1–30 Hz) and amplitude (13–100 Hz) frequency combination, using 1-Hz resolution. We identified regions of interest (ROIs) within this matrix corresponding to four frequency band-pairs (delta-beta, delta-gamma, theta-beta, theta-gamma). Because these bands span different frequency ranges, the number of bins included in each ROI varied accordingly. For each ROI, we summed the MI values across all relevant frequency bins to capture the total PAC strength within that frequency range which is a metric of PAC strength [54]. We selected summation over averaging or median to ensure that all contributing frequency bins within the band were fully represented [54]. Supplementary Table 1 summarizes the ROI values for both the high and low BDI groups. To compare irPAC between groups while also accounting for multiple comparisons, we applied the Wilcoxon rank-sum test, with false discovery rate (FDR) correction to control Type I errors [55,56].

2.4. Exploratory analysis

To assess the potential of irPAC as a biomarker for depression severity, we analyzed the correlation between the magnitude of significant irPAC patterns and patients’ BDI scores. Rather than categorizing BDI scores into discrete groups, we treated them as a continuous variable to better capture the relationship through correlation analysis.

Previous literature review has shown seizures are correlated to coupling/decoupling of different frequency bands [40] and dynamic stability changes [29,30]. Thus, seizures can be considered as a surrogate for irPAC perturbation in the brain network and we explored correlations between irPAC and the number of seizures captured in the epilepsy monitoring unit (EMU).

Here, we used Spearman correlation and FDR correction to assess the relationships between irPAC magnitude, BDI scores, and number of seizures in the EMU. We also used a correlogram to summarize the correlations among irPAC, BDI scores and seizure count.

3. Results

3.1. Group differences in irPAC: elevated delta-beta coupling in depression

Out of all possible PAC combinations examined, only delta-beta coupling within six specific brain regions and directions exhibited statistically higher irPAC (p_FDR < 0.05) in the high, compared to the low, BDI groups as shown in Fig. 1. These regions included AMG delta to HIP beta (p_FDR = 0.046), AMG delta to SFG beta (p_FDR = 0.046), ACC delta to HIP beta (p_FDR = 0.046), HIP delta to SFG beta (p_FDR = 0.049); and the reverse direction: SFG delta to HIP beta, p_FDR = 0.046), and HIP delta to mOFC beta (p_FDR = 0.046) PAC.

3.2. irPAC correlates of depression severity (BDI)

Among all region pairings with statistically significant irPAC group differences, four exhibited a statistically significant (p_FDR < 0.05) correlation with patients’ BDI scores for delta-beta PAC, as shown in Fig. 2. These included AMG-HIP (ρ = 0.48; p_FDR = 0.033), SFG-HIP (ρ = 0.50; pFDR = 0.033), HIP-rACC (ρ = 0.58; pFDR = 0.033), and AMG-SFG (ρ = 0.54; pFDR = 0.033).

Fig. 2.

Heatmap of Phase-Amplitude Coupling in Significant Brain Regions. It shows side-by-side comparison of inter-regional phase-amplitude coupling (irPAC) strength across six brain regions between two subject groups—High BDI (top row) and Low BDI (bottom row). For each subplot, heatmaps display the coupling strength as a function of delta phase (1–4 Hz) and beta amplitude (13–30 Hz). A global color scale reflects the delta-beta coupling strength on the right.

3.3. irPAC and seizure frequency in EMU

While none of the irPAC pairs showed statistically significant (pFDR < 0.05) correlation with patients’ number of seizures in the EMU as shown in Fig. 3, several pairs exhibited correlation magnitudes similar to those observed with depression severity. Specifically, seizure count correlations included rACC-HIP (p_FDR = 0.38), AMG-HIP (p_FDR = 0.33), mOFC-HIP (pFDR = 0.43), SFG-HIP (pFDR = 0.33), SFG-AMG (pFDR = 0.56), and HIP-SFG (p_FDR = 0.06). To statistically compare the strength of correlations between depression and seizure count, we conducted Fisher’s Z tests for each significant irPAC pair. The Fisher’s Z tests revealed that depression correlations were not significantly stronger than seizure correlations for any of the four pairs: AMG-HIP (Z = 0.360, p = 0.719), SFG-HIP (Z = 0.582, p = 0.561), HIP-rACC (Z = −0.484, p = 0.628), and AMG-SFG (Z = 0.182, p = 0.855). All p-values were non-significant, indicating that the correlation strengths are not significantly different between these two clinical measures.

Fig. 3.

Irpac strength correlation to bdi scores. the y axis title shows the irpac pairing. the blue dots indicate data points and a red regression line illustrating the spearman correlation trend. each subplot includes the spearman correlation coefficient (ρ) andFalse Discovery Rate (FDR)-corrected p-value.

A correlogram showing correlations of BDI scores and seizure counts with irPAC is shown in Fig. 4. To address potential confounding effects of epilepsy severity on our depression-related findings, we analyzed the duration of drug-resistant epilepsy (DRE) between depressed and non-depressed patients. We found no statistically significant difference in DRE duration between the two groups (p > 0.05), suggesting that both groups had comparable epilepsy severity by this measure. (Fig. 5).

Fig. 4.

Irpac strength correlation to number of seizures in emu. The y axis title shows the irPAC pairing. The blue dots indicate data points and a red regression line illustrating the Spearman correlation trend. Each subplot includes the Spearman correlation coefficient (ρ) and False Discovery Rate (FDR)-corrected p-value.

Fig. 5.

Correlogram of BDI Scores and Seizure Counts with Brain Region PAC. This correlogram depicts the Spearman correlation coefficients between Beck Depression Inventory (BDI) scores, seizure counts, and delta-beta interregional phase amplitude coupling measures of six specific brain region pairs. Statistical significance, assessed via Spearman correlation and corrected for multiple comparisons using the False Discovery Rate (FDR) procedure, is indicated by bold font.

4. Discussion

In this study, we identified an irPAC signal that is characterized by its specificity in terms of spatial regions, frequency bands, coupling direction (i.e., which low-frequency phase modulates which high-frequency amplitude), and magnitude. Notably, this signal differentiated between depressed and non-depressed temporal lobe epilepsy patients in a way that differs from our previously reported elevated local PAC signals and other studies [27,29,30]. First, the signal was evident in regions that are spatially more distant yet remain functionally integrated within the depression–temporal lobe epilepsy network [16,33,57]. Second, the coupling predominantly involved delta phase and beta amplitude which has not been reported in the intracranial EEG literature on depression in epilepsy or in primary depression, and it added to the existing corpus of iEEG studies on depression [30,58]. Third, when treating BDI scores as a continuous variable, four out of six irPACs showed a significant positive correlation with depression severity. Lastly, no statistically significant correlations between irPAC and seizure count were observed.

Our findings of irPAC frequency ranges are consistent with the general pattern observed in previous literature, where lower frequency oscillations couple with higher frequency oscillations. While prior studies have reported theta (4–8 Hz) phase coupling with gamma (30–100 Hz) amplitude in memory formation [18,33,57], our findings demonstrate delta (0.1–4 Hz) phase coupling with beta (13–30 Hz) amplitude in the context of comorbid depression within epilepsy networks.[18,33,57].

Our findings, which indicated that delta–beta coupling can distinguish depressed from non-depressed epilepsy patients, aligned with recent scalp EEG literature. One study showed that elevated delta-beta coupling is associated with adolescents who suffer from depression and potentially have more difficulty regulating emotions [59]. In addition, earlier studies also suggest that increased delta–beta power coupling may reflect a maladaptive overcontrol and rigidity in emotion regulation [46,47]. Similarly, another study found that high delta–beta coupling is associated with social anxiety and behavioral inhibition [45].

In addition, the limbic-frontal connectivity disruptions shown in our study aligned with the neuropsychiatric profile observed in temporal lobe epilepsy (TLE) patients, characterized by increased neuroticism and depression [23]. Such profile was shown to be corelated to hyposynchronous connectivity between right hippocampus and both Brodmann area 9 (prefrontal cortex) and area 47 m (orbitofrontal cortex), notably overlapping with our observed AMG delta to HIP beta, HIP delta to SFG beta, and HIP delta to mOFC beta coupling patterns. This suggests that the elevated BDI scores in our depressed cohort may reflect this underlying fronto-limbic disconnection and emotional dysregulation.

Taken together, findings from the scalp and imaging literature and our intracranial EEG study suggest that the dysregulation of delta–beta interactions may be putative in the model of limbic-cortical dysfunction [39]. Particularly, these irPAC aligned with both the classic Papez circuit (via ACC–HIP pathways) and extended limbic-prefrontal loops (AMG–HIP, HIP–SFG/mOFC) embedded within the default-mode networks [7–9,60]. The hippocampus connects to the medial and orbitofrontal cortices via the fornix and uncinate fasciculus, tracts shown to be structurally and functionally impaired in TLE patients with depression in functional magnetic resonance imaging (fMRI) studies [6]. Our irPAC in AMG-HIP, AMG-SFG, HIP–SFG/mOFC, ACC-HIP findings correlated to other tracts investigation studies [61,62]. Our work provided an electrophysiological substrate for the overlapping network dysfunction seen in epilepsy and depression and highlighted amygdala–hippocampus delta–gamma coupling as a potentially novel mechanism for emotional regulation. The multi-scale nature of our findings reveals how inter-regional PAC dysfunction compounds existing regional PAC abnormalities to disrupt the entire limbic-cortical network. While regional PAC abnormalities (5–25 Hz × 80–100 Hz) represent compromised local neural coordination within individual limbic and cortical nodes, the addition of aberrant inter-regional delta-beta coupling (0.1–4 Hz × 13–30 Hz) creates pathological synchronization patterns that propagate across and disrupt the broader network. This compounding dysfunction resembles seizure propagation mechanisms, where abnormal oscillatory coordination spreads through neural circuits and disrupts normal network function. The result is a cascade of network failure where compromised individual nodes become further destabilized by pathological inter-regional connectivity, leading to breakdown of the entire limbic-cortical system. This hierarchical disruption provides electrophysiological evidence for how depression in epilepsy involves both local circuit dysfunction and network-wide communication failures that compound each other, supporting the limbic-cortical dysregulation model at multiple levels of neural organization.

As there is no consensus on how to interpret the direction of irPAC, we used a data-driven approach to identify frequency ranges as sub-rectangular regions of interest in the modulation indices for both directions between 2 brain regions. It is interesting that HIP and SFG have delta beta coupling in both directions, i.e. phase of HIP coupling with SFG’s amplitude and vice versa. This closed loop feature aligns with previous literature showing involvement of the HIP and SFG neural activities in emotional processing in subclinical depression [17]. While there is no surprise that AMG, HIP, ACC, SFG, mOFC are all involved in the depression network [15], their specific direction of irPAC needs further investigation to determine their physiological and functional significance.

We also argue that this data-driven approach may enable future investigations into previously reported frequency ranges and help refine specific neural oscillatory activities. For instance, we observed that the strongest signal in 0.1–2 Hz delta range for phase across all regions and the strongest signal in ACC-HIP irPAC. Future work may include refining the specific frequencies in the delta and beta ranges for irPAC signals and their specific regions. Additionally, we observed not only the strongest signals in the delta-beta phase ranges but also a distinctive pattern of irPAC magnitude index (MI) elevations and reductions. Specifically, there were variations where certain phase-frequency bands show increased coupling strength (elevations) while others exhibit decreased coupling (reductions), creating a spatially heterogeneous signature of irPAC modulation index. With the advent of machine learning and image recognition algorithms, a potential avenue for further dimensionality reduction of irPAC signals could involve unsupervised recognition of these modulation index patterns [63].

In this cohort of patients with temporal lobe epilepsy, we observed that delta–beta irPAC strength in four interregional circuits (AMG–HIP, SFG–HIP, HIP–rACC, and AMG–SFG) was positively associated with BDI scores (Spearman ρ ≈ 0.5). Since our recordings and clinical assessments were limited to discrete sessions in the EMU, we can only interpret these irPAC measures as trait-level markers of depressive symptomatology, not as dynamic biomarkers that fluctuate with mood state over days or within-day-night cycles. An important methodological consideration is that our 5-minute samples were selected from daytime recordings and subsequently evaluated by a board-certified epileptologist to determine brain state. The majority of patients were in awake states, with only two patients in sleep states. Importantly, these two sleep cases were equally distributed between the depressed (n = 1) and non-depressed (n = 1) groups, eliminating systematic bias between conditions. While different brain states have distinct EEG characteristics, the fact that our connectivity measures differentiated between depressed and non-depressed patients in both awake and sleep states provides preliminary evidence that the delta-beta irPAC signal may be detectable across different brain states. However, given the small number of sleep cases in our cohort, we cannot draw definitive conclusions about state-independence. Interestingly, previous work has demonstrated that local PAC biomarkers for depression show stronger signals during sleep states compared to awake states [29], suggesting that cross-frequency coupling measures may be more pronounced during sleep. While our irPAC measure represents a fundamentally different signal (inter-regional rather than local connectivity), this prior finding raises the intriguing possibility that our biomarker might also demonstrate enhanced detectability during sleep states. Further investigation with larger, systematically collected sleep-wake datasets would be necessary to fully characterize the robustness of this biomarker across different brain states.

However, the superior frontal gyrus (SFG) lies at the superior aspect of the prefrontal cortex overlapping functionally with dorsolateral prefrontal cortex (DLPFC) [64]. We speculate that enhanced PAC in SFG–HIP and AMG–SFG pathways might reflect fronto–hippocampal and amygdalo–frontal loops that can be modulated by DLPFC–targeted stimulation. Indeed, transcranial magnetic stimulation over DLPFC attenuates its functional connectivity with both hippocampus and amygdala in depression, and can transiently enhance PAC [65,66]. Future work should test whether continuous PAC monitoring in SFG–HIP and AMG–SFG pathways predicts clinical response to DLPFC–targeted or depth–electrode stimulation in comorbid depression with responsive neurostimulation.

Our exploratory analyses showed no statistically significant association between irPAC and the number of seizures across all participants. We acknowledge that there is no well-established biomarker for epilepsy brain network stability and seizures have shown to be associated with increased and decreased PAC [40,67]. The number of seizures patients had in the EMU serves as a rudimentary surrogate for brain network stability in theory, but our sample size does not capture a wide range of seizures and the patient-to-patient variation in their networks. The BDI scores and seizure counts were minimally correlated (Spearman ρ = 0.1), yet both contributed independent variance to irPAC patterns. This suggests that delta-beta irPAC may reflect overlapping but distinct neurophysiological processes, one related to depressive symptomatology and another to seizure-related network dynamics, rather than representing purely depression-specific or seizure-specific signatures.

We therefore propose that future work systematically examine seizure–induced irPAC changes to probe its promise as a biomarker capturing the bidirectional interplay between epilepsy and depression. The absence of a direct seizure–irPAC correlation, coupled with the selective sensitivity of delta–beta irPAC to depressive symptoms (but not seizure frequency), suggests that while this metric is particularly attuned to depression, it nonetheless embodies both depression– and seizure–related network influences via shared oscillatory pathways. Finally, our finding of no significant difference in drug–resistant epilepsy duration between depressed and non–depressed patients indicates comparable epilepsy severity across groups. Together, these observations tentatively suggest that epilepsy network dysfunction may not be the primary driver of our observed irPAC differences, leaving open the possibility that delta–beta irPAC serves as a depression–related neural signature in epilepsy.

The limitations of our study are as follows. First, our sample size is limited, and we included both hemispheric data in our regional analysis while accounting of left–right difference by coupling left and right regions separately in analysis. Our sample size precluded adequately powered analyses of lateralization-specific differences in irPAC patterns, which represents an important area for future investigation with larger cohorts. Exploratory analyses comparing left vs. right TLE patients showed no statistically significant differences in irPAC values, though this analysis was underpowered for lateralization-specific effects. Second, we interpreted BDI scores from pre-surgical neuropsychological evaluation as depression severity. More timely and other mood assessment tools can be employed to capture the complexity of depression in epilepsy [68,69]. Especially with delta-beta literature on emotional regulation and cognitive inhibition, future direction may use assessment tools in that area [59,70,71]. Third, comorbid depression in epilepsy can worsen or improve with seizures [68,69]. Further characterization of individuals, post-ictal mood assessments and irPAC measurements may allow more refined analysis in the future. Fourth, future studies would benefit from more comprehensive epilepsy severity assessments including seizure frequency data for preceding months, lifetime seizure counts, and history of status epilepticus, which were not available in our dataset. Additionally, our study did not systematically account for antiseizure medication (ASM) status during recordings. While patients are typically weaned off ASMs during EMU monitoring, the variable timing and completeness of withdrawal across patients represents a potential confounding factor that should be addressed in future prospective studies with standardized ASM documentation protocols. Fifth, we did not perform a formal test–retest analysis to evaluate the stability of irPAC estimates across different sample durations. While our use of 1-minute segments follows precedent in the literature [29], and we incorporated 5 min of data per condition to reduce temporal variability, we recognize that quantifying estimate stability as a function of sample length would enhance methodological rigor. However, due to the computational demands of such an analysis across multiple conditions and subjects, this was beyond the scope of the present study and represents a valuable direction for future work. Lastly, we do not have multimodal brain network data for this group of patients, such as fMRI. An interesting study to further our understanding of epilepsy and depression would be to correlate delta-beta irPAC signal to different antiseizure medications and functional connectivity measurements in a region, direction and magnitude specific approach [72].

5. Conclusions

We used a data driven approach to identify a delta-beta inter-regional phase amplitude coupling signal in AMG delta to HIP beta, AMG delta to SFG beta, ACC delta to HIP bet, HIP delta to SFG beta; and the reverse direction: SFG delta to HIP beta, and HIP delta to mOFC beta. The signal includes specific frequency ranges, direction of phase to amplitude and limbic-cortical brain regions pairs in Papez circuit and default mode network. This signal can differentiate the depressed state of temporal lobe epilepsy network and positively correlate the severity of depression in patients. This signal may provide a new tool to study the relationship between depression and epilepsy that is complementary to imaging studies.

Appendix A. Supplementary data

Supplementary data to this article can be found online at https://doi.org/10.1016/j.yebeh.2025.110728.