Gamma-band cortical functional network abnormalities in late-life depression with suicidal ideation: insights from EEG graph theory and machine learning

Abstract

Background:

Late-life depression (LLD) is the major risk factor for elderly suicide, and suicidal ideation (SI) is a crucial stage for prevention. However, LLD are less likely openly express SI. Gamma oscillations, closely linked to cognition and mental processes, may contribute to the pathophysiology of LLD and suicidal behavior through their dysregulation within large-scale brain networks. The aim of our research was to investigate the cortical functional networks in the gamma band to better understand the neurobiological mechanisms underlying SI in LLD.

Methods:

Electroencephalography (EEG) was recorded from 30 LLD with SI (LLD-SI), 32 LLD without SI (LLD-NSI), and 34 normal controls. We applied source-level graph theory based on functional connectivity in gamma band and utilized machine learning to differentiate between LLD-SI and LLD-NSI groups using network features.

Results:

Significant diminished gamma functional connectivity, particularly involving the orbitofrontal cortex, was observed in both subtypes of the LLD group. In graph theory analysis, LLD-SI showed decreased average clustering coefficient (p < 0.001) and characteristic path length (p = 0.021), along with increased global efficiency (p = 0.015) compared to LLD-NSI. Compared to NC, LLD-SI also demonstrated reduced average clustering coefficient (p = 0.004), characteristic path length (p = 0.004), and higher global efficiency (p = 0.004). We also found several nodal metrics, which suggested potential hubs related to SI. The graph theorical method effectively distinguished SI in LLD, with an accuracy of 69.35%, sensitivity of 73.33%, and specificity of 65.63% based on gamma-band network features.

Limitations:

The sample sizes are relatively small. Higher-density EEG systems and interventional study designs should be included in future research. Future studies should incorporate external validation datasets to confirm the clinical utility of the proposed classification framework.

Conclusion:

Our research provides valuable insights into the brain connectome in gamma band of SI in LLD. Gamma-band network indices may serve as potential biomarkers for detecting SI and offer frequency-specific targets for neuromodulation in suicide prevention and treatment strategies for LLD patients.

1. Introduction

Globally, more than 700 000 people die by suicide each year.¹ With the global population aging, suicide among the elderly has become a growing concern. Suicide among older adults differs from other age groups, characterized by higher lethality and more explicit planning.² Late-life depression (LLD) is one of the most important risk factors for elderly suicide.³ Suicidal ideation (SI) is the first stage of the complex suicide continuum, which serves as an important stage of suicide prevention.⁴ However, identifying SI in LLD is challenging since older adults are less likely acknowledge SI.⁵ Therefore, investigating the underlying neurophysiological mechanism of late-life SI is crucial for improving its detection.

Previous magnetic resonance imaging (MRI) studies have found that suicidality in patients with depression is associated with decreased gray matter volume in the orbitofrontal cortex (OFC), anterior insular, and prefrontal cortex.⁶ Functional MRI studies have also revealed that reduced intrinsic functional connectivity between the rostral anterior cingulate cortex (rACC) and the orbitomedial prefrontal cortex is associated with SI in patients with major depressive disorder (MDD).⁷ Numerous neuroimaging studies suggest that investigating the dysregulation of specific large-scale brain networks may provide insights into the neurobiological mechanisms of suicidality in depression.⁸ The default mode network (DMN), which is involved in self-referential thinking and rumination, has shown dysfunction in individuals with MDD,⁹ and emerging evidence links this dysfunction to suicidal behaviors in MDD.¹⁰

Gamma oscillations are closely associated with various cognition and mental processes. Recently, we investigated the oscillatory patterns in LLD with different depressive severity, suggesting that abnormal gamma oscillations may potentially serve as distinguishing feature between LLD and MDD in younger individuals.¹¹ Furthermore, gamma oscillations play a crucial role in identifying suicidal behaviors. Previous research has found altered gamma power in the frontal, central, and temporal regions of MDD with suicidal behaviors.¹² Abnormal gamma power has also been observed in the visual cortex and insula of these patients.¹³ Additionally, decreased inter-network gamma-band connectivity between the executive control network and the DMN has been associated with increased suicide risk in MDD with depression.¹⁴ According to previous studies of gamma-band brain network, exploring the brain topological structure in the gamma band may enhance our understanding of the mechanisms underlying SI in LLD.

Electroencephalography (EEG), with its superior temporal resolution and accessibility, offers a promising alternative for probing frequency-specific oscillatory dynamics and network abnormalities. EEG graph theory analysis has made progress in depression research, offering valuable insights into the topological properties of brain networks. Several studies revealed that network metrics of EEG graph theory such as clustering coefficient and global efficiency could distinguish MDD from normal control.^15,16 Additionally, EEG graph theory parameters have been identified as potential antidepressant-responsive phenotypes for MDD.¹⁷ Researchers have investigated cortical functional networks in MDD with suicide attempt (SA) and SI based on resting state EEG graph theorical analysis, which revealed patients with SA showed lower clustering coefficient and higher path length (PL) in alpha-band brain network.¹⁸ These findings suggest that abnormal functional networks identified through EEG graph theory may reveal biomarkers for suicidal behavior in MDD. However, most studies have only focused on adolescent or middle-aged MDD patients. MDD and LLD differ in large-scale brain network abnormalities, including the fronto-parietal network, dorsal attention network, and visual network.¹⁹ Compared to younger counterparts, the mechanisms of late-life suicide risk are far less understood.

The purpose of this cross-sectional study on LLD was to investigate the cortical functional networks of LLD-SI in gamma band. We also employed a machine learning approach to explore the role of EEG graph theory and functional connectivity analysis as potential marker for identifying SI in LLD. We hypothesized that LLD-SI would display altered cortical functional networks in gamma band compared to those without SI and NC. We also hypothesized that EEG-based brain networks could effectively distinguish between LLD patients with and without SI.

2. Methods and materials

2.1. Participants

We recruited 62 LLD patients in The Affiliated Brain Hospital of Guangzhou Medical University, and 34 normal control (NC) from communities. The dataset partially shared participants with our previous microstate analysis.²⁰ Three additional participants were recruited for both the LLD-NSI and NC group. Although we aimed to increase the LLD-SI sample, gamma-band source-level analyses require strict EEG quality standards since this frequency band is highly susceptible to electromyographic artifacts.²¹ Three LLD-SI participants were excluded due to elevated high-frequency muscle artifacts. The final sample comprised 30 LLD-SI, 32 LLD-NSI, and 34 NC. The study was approved by the Ethics Committees of the Affiliated Brain Hospital of Guangzhou Medical University (2014, 078). All participants received free clinical examinations, and the results for hospitalized patients were given to their attending physicians to support diagnosis and treatment.

For the LLD groups, we included patients: (1) aged 60–85; (2) met the DSM-5 criteria of MDD with current major depressive episodes confirmed by two trained psychiatrists; (3) we included NC whether they (1) demonstrated normal cognition; (2) aged 60–85, and (3) had no history of depressive episodes. The exclusion criteria for both LLD and NC were the same: (1) other DSM-5 psychiatric disorders; (2) stroke, delirium, Parkinson's disease, brain tumors, dementia, and other neurological diseases; (3) hypothyroidism; (4) history of substance abuse; (5) history of head injury with loss of consciousness >30 min. All inclusion and exclusion criteria were assessed and confirmed by two experienced psychiatrists.

2.2. Demographic and clinical assessments

We collected demographic information, including age, sex, education, and current medication for all participants. Before the EEG recording, all subjects underwent a comprehensive assessment by two trained psychiatrists. Depressive severity was measured using the HAMD, with a cut-off of 17 to classify current depressive episodes. We also recorded clinical details, such as the age of onset and disease duration. The Activities of Daily Living Scale (ADL), which is used to evaluate an individual's ability to perform basic self-care tasks necessary for independent living.

SI was defined by meeting both of the following criteria: (1) a score of ≥ 2 on item 3 of the 17-item Hamilton Rating Scale for Depression (HAMD),²² which assesses “wishes to be dead or has any thoughts of possible death to self”, and (2) a score > 1 on item 4 or 5 of the Beck Scale for Suicide Ideation (SSI),²³ which evaluates active and passive suicidal thoughts. Both conditions had to be met for a diagnosis of SI. In the SSI, patients with a score > 1 on items 4 or 5, indicating the presence of SI, were required to complete the remaining 14 items. Those with a score of 0 on both items 4 and 5, indicating no SI, could skip the remaining items. The SSI score was calculated by summing the 19 items.²⁴ Patients with current or past suicide attempts were excluded from the study.

2.3. Neuropsychological assessments

We used Clinical Dementia Rating (CDR) to exclude dementia in our research. The CDR is a widely used clinician-rated scale that assesses cognitive performance in domains such as memory, orientation, judgment, and daily activities, providing a global measure of dementia severity.²⁵ Participants with a CDR score > 0.5 were excluded from the study.²⁶

Global cognitive functioning was evaluated using the Mini-Mental State Examination (MMSE), with normative cutoffs adjusted according to education level based on established Chinese population standards: 17 points for participants without formal education, 20 points for elementary education completion, and 24 points for those achieving middle school or higher educational levels.²⁷

However, the LLD groups were not required to meet the standard MMSE cutoff values, as MMSE scores in LLD are often influenced by depressive symptoms, which may result in lower scores.²⁶ Therefore, we used only the CDR to exclude participants with dementia in the LLD groups.

2.4. EEG recording and preprocessing

We instructed all participants to abstain from alcohol, sleep medications, caffeine, and nicotine for at least 24 h prior to the examination. We collected EEG data between 8 a.m. and 11 a.m. to control circadian rhythm effects and promote a state of wakefulness. To ensure clean electrode placement, participants washed their hair upon arrival. We collected at least 5 min resting-state EEG under eyes-closed conditions for all participants using the Electrical Geodesics Inc. 64-channel system at a sampling rate of 1000 Hz. We used Cz electrode as the online reference and the band-pass was set within 0.1–70 Hz. All electrode impedance remained below 50 kΩ.

EEG data were preprocessed offline using the EEGLAB v14.1.1 toolbox in MATLAB R2016b. The data were band-pass filtered (0.1–70 Hz) with a notch filter (48–52 Hz) and segmented into 2000 ms epochs. We interpolated noisy electrodes using spherical spline interpolation as well as removing artifacts from eye movements, blinks, and muscle activity using independent component analysis. We excluded data with artifacts exceeding ±100 mV. Finally, we re-referenced the final data to the average.

2.5. Source localization

We implemented the source reconstruction in the Brainstorm toolbox (http://neuroimage.usc.edu/brainstorm) in MATLAB R2022b.²⁸ We used a three-layer boundary element model from the ICBM152 (Montreal Neurological Institute 152) anatomy template to compute the lead field matrix, with a cortical mesh of 15 002 vertices. We used the standardized low resolution electromagnetic tomography (sLORETA) method as an inverse solution, which approximated the electrical neuronal activity on the cortex. Regions of interest (ROIs) were extracted based on Desikan–Killiany atlas, which consisted of 68 regions.²⁹

2.6. Resting-state functional connectivity analysis

To improve the EEG spatial resolution and overcome the volume conduction effect, we performed resting-state functional connectivity (Rs-FC) analysis based on the source reconstruction described above. We calculated the phase locking value (PLV) to index the extent of phase relationship between two brain regions. We applied the Fourier transform to extract phase information within the band of interest (gamma, 30–50 Hz) for PLV computation.^12,30 In addition, based on the findings of Kim and colleagues,¹⁸ we also included the high alpha band (10–12 Hz) for PLV computation. Both frequency bands were subsequently used for the calculation of graph theoretical parameters.

2.7. Network analysis

Graph theory can mathematically quantify brain network connectivity. Graph theory analysis involves a set of nodes (brain regions) and edges (functional connectivity between nodes). We conducted graph theory analysis by MATLAB R2022b. The graph consists of 68 nodes within the Desikan–Killiany atlas connected to each other by edges. We also assessed the edge using the PLV measure mentioned above, which has been extensively applied in network analysis.³¹ The brain network topological characteristics were measured by following graph parameters:

Average clustering coefficient (CC): the clustering coefficient of a node (i.e., nodal CC) quantifies the degree of its neighbors in a network tend to cluster together. Accordingly, the CC is the mean of nodal CCs across the entire network. A lower CC indicates a network with decreased segregation.

Characteristic PL: measuring the average shortest distance between all pairs of nodes in a network. The shortened PL represents a well-integrated network.

Global efficiency (E_glob) of a network is the average inverse shortest PL, which measures the overall efficiency of information transfer in a network. Nodal efficiency (E_i) were also assessed for each node. The enhanced E_glob can also imply a well-integrated network.

2.8. Machine learning for classification

We further tested the classification efficacy of several parameters to distinguish SI from NSI, including the connected pair with significant group differences in gamma PLV, which indicated topologically focal connectivity; global metrics that indicated impaired segregation and enhanced integration in the brain functional organization; and nodal metrics that suggested potential hub related to SI. We conducted this analysis using the support vector machine (SVM) approach, which has been applied in several studies on neural networks.^6,32 We conducted the SVM analysis using LIBSVM based on MATLAB ³³ (https://www.csie.ntu.edu.tw/∼cjlin/libsvm). We employed a linear kernel function in SVM, as it demonstrates lower overfitting risk than nonlinear SVM. We set the regularization parameter C to its default value (C = 1). We validated the SVM model using leave-one-out cross-validation and a 10 000-iteration permutation test to verify stability and generalizability. Our analysis calculated weight factors to quantify each selected feature's contribution to classification. Finally, we used the accuracy, specificity, sensitivity, and the value of the area under characteristic (AUC) of the receiver operating characteristic (ROC) curve to indicate classification performance.

2.9. Statistical analysis

For demographic characteristics, we compared continuous variables using 1-way Analysis of variance (ANOVA) or Student's t test and analyzed categorical variables with χ² tests to examine group differences.

We compared Rs-FC differences across 68 brain regions using Analysis of covariance (ANCOVA), with age, sex, and years of education as covariates. To control for type I errors from multiple comparisons, we adjusted the significance of connected pairs using the network-based statistic (NBS) approach, implementing 5000 permutations via NBS toolbox 1.2 in MATLAB.³⁴ The primary threshold was set at a significant level of p < 0.001 to detect edges with the most localized differences. The secondary threshold for family-wise error-corrected cluster testing was set at p < 0.05.

We performed univariate ANCOVA to compare global cortical network characteristics across the three groups, controlling for age, sex, and education years. We also applied repeated measures ANCOVA to assess differences at nodal level, with the same covariate above, with groups as between-subject factors and 68 nodes as within-subject factors. If Mauchly's sphericity test indicated violation, we applied Greenhouse Geisser correction. Once the interaction effect between nodes and groups was significant, we conducted post-hoc univariate ANCOVA with Bonferroni multiple comparison correction.

Pearson correlation analyses were conducted between the gamma-band network measures that significantly differed between the LLD-SI and LLD-NSI groups (CC, PL, and E_glob) and clinical/neuropsychological test scores, with correction for multiple comparisons using the false discovery rate.

3. Results

3.1. Demographic and clinical characteristics

Table 1 shows the demographic and clinical characteristics. Our analysis revealed no significant differences in age, sex distribution, years of education, total ADL and CDR scores among the three groups. Additionally, no differences in medication use were found between the LLD-SI and LLD-NSI groups. Both LLD groups had worse HAMD scores compared to the NC group, but no significant difference was observed between the LLD-SI and LLD-NSI groups. Furthermore, both LLD-SI and LLD-NSI groups showed poorer cognitive performance than the NC group.

Table 1.

Demographic and clinical data of all participants.

Parameter	LLD-SI (A)	LLD-NSI (B)	NC (C)	t/F/χ2	p value	Post-hoc
Demographic (M, SD)
Number of subjects	30	32	34	NA	NA
Age in years	67.60 (4.56)	68.22 (5.96)	69.65 (4.74)	1.375	0.262b
Sex (male/female)	7/23	10/22	9/25	0.501	0.778a
Age of onset (years)	55.17 (10.91)	57.94 (12.02)		t = 0.948	0.347
Duration of disease (years)	12.00 (9.64)	10.26 (10.41)		t = 0.620	0.538
Years of education	9.23 (2.54)	9.28 (2.55)	10.26 (3.05)	1.505	0.227b
Assessments (M, SD)
HAMD-17 scores	22.83 (2.89)	21.59 (2.26)	2.94 (2.33)	652.688	<0.001 b	A, B > C
SSI	15.10 (4.05)	0.66 (0.94)	–	t = 19.651	<0.001	A > B
ADL	13.53 (0.73)	13.66 (0.83)	13.76 (0.70)	0.751	0.745
MMSE scores	22.73 (2.10)	23.19 (2.14)	27.18 (1.01)	60.848	<0.001 b	A, B < C
CDR	0.40 (0.20)	0.38 (0.22)	0.31 (0.25)	1.425	0.246
Medications (N, percent)
Unmedicated	12 (40.00%)	14 (43.75%)	–	0.089	0.765a
SSRI	7 (23.33%)	8 (25.00%)	–	0.023	0.878a
SNRI	6 (20.00%)	7 (21.88%)	–	0.033	0.856a
NASSA	5 (16.67%)	4 (12.50%)	–	0.217	0.642a
Antipsychotic	8 (26.67%)	7 (21.88%)	–	0.194	0.660
BZDs	12 (40.00%)	17 (53.13%)	–	1.071	0.301a

Note: Significant p values were emphasized as bold.

Abbreviations: A: LLD-SI: late-life depression with suicidal ideation; B: LLD-NSI: late-life depression without suicidal ideation; C: NC: healthy control; M: mean; SD: standard deviation; N: number; HAMD: Hamilton Rating Scale for Depression; SSI: Beck's Scale for Suicide Ideation; ADL: activities of daily living scale; MMSE: Mini-Mental State Examination; CDR: Clinical Dementia Rating; SSRI: selective serotonin reuptake inhibitors; SNRI: serotonin–norepinephrine reuptake inhibitors; NASSA: noradrenergic and specific serotonergic antidepressants; BZDs: benzodiazepines.

^aχ² test.

^bOne-way ANOVA analyses; t for Student's t test.

3.2. Resting-state functional connectivity in gamma frequency band

After source reconstruction and NBS correction, we observed significant diminished gamma connectivity in both subtypes of LLD groups, focally within bilateral subregions of the OFC (LOF.R-LOF.L, F = 8.493, p < 0.001, η² = 0.159; LOF.R-MOF.L, F = 9.982, p < 0.001, η² = 0.182; MOF.R-MOF.L, F = 9.285, p < 0.001, η² = 0.171), between subregions of the right OFC, and bilateral rostral anterior cingulate (LOF.R-RACC.L, F = 11.412, p < 0.001, η² = 0.202; LOF.R-RACC.R, F = 11.355, p < 0.001, η² = 0.201; MOF.R-RACC.L, F = 8.903, p < 0.001, η² = 0.165; MOF.R-RACC.R, F = 7.960, p < 0.001, η² = 0.150). Additionally, both LLD subgroups showed decreased gamma-band connectivity between the right lateral OFC and superior frontal gyrus compared to the NC group (LOF.R-SFG.R, F = 11.599, p < 0.001, η² = 0.205). (see Table 2 and Fig. 1)

Fig. 1.

Connected pairs with significant group differences in gamma band after correction using the network-based statistic. Figure presents the results after correction. The edges indicate the significant decreased gamma connectivity observed in both groups of LLD patients compared to NC. The thickness and color of the edges indicate the magnitude of the F-statistic, with thicker or brighter edges signifying stronger gamma connectivity between two brain regions. Abbreviations: LOF.R: right lateral orbitofrontal cortex; LOF.L: left lateral orbitofrontal cortex; MOF.R: right medial orbitofrontal cortex; MOF.L: left medial orbitofrontal cortex; RACC.R: right rostral anterior cingulate; RACC.L: left rostral anterior cingulate; SFG.R: right superior frontal gyrus.

Table 2.

PLV of connected pairs with significant group differences in the gamma band across LLD-SI, LLD-NSI, and NC groups.

Brain regions	LLD-SI (A) mean (SD)	LLD-NSI (B) mean (SD)	NC (C) mean (SD)	F value	p value	Effect size (partial η2)	Post-hoc
LOF.R-LOF.L	0.17 (0.10)	0.23 (0.17)	0.32 (0.17)	8.493	<0.001	0.159	A, B < C
LOF.R-MOF.L	0.30 (0.12)	0.36 (0.17)	0.46 (0.17)	9.982	<0.001	0.182	A, B < C
LOF.R-RACC.L	0.43 (0.10)	0.48 (0.15)	0.58 (0.14)	11.412	<0.001	0.202	A, B < C
LOF.R-RACC.R	0.47 (0.11)	0.52 (0.14)	0.62 (0.14)	11.355	<0.001	0.201	A, B < C
LOF.R-SFG.L	0.21 (0.12)	0.23 (0.14)	0.36 (0.14)	11.599	<0.001	0.205	A, B < C
MOF.R-MOF.L	0.73 (0.06)	0.76 (0.09)	0.80 (0.08)	9.285	<0.001	0.171	A, B < C
MOF.R-RACC.L	0.85 (0.05)	0.86 (0.07)	0.90 (0.05)	8.903	<0.001	0.165	A, B < C
MOF.R-RACC.R	0.88 (0.05)	0.90 (0.05)	0.93 (0.05)	7.960	0.001	0.150	A, B < C

Note: Abbreviations: LLD-SI: late-life depression with suicidal ideation; LLD-NSI: late-life depression without suicidal ideation; NC: normal control; PLV: phase locking value; LOF.R: right lateral orbitofrontal cortex; LOF.L: left lateral orbitofrontal cortex; MOF.R: right medial orbitofrontal cortex; MOF.L: left medial orbitofrontal cortex; RACC.R: right rostral anterior cingulate; RACC.L: left rostral anterior cingulate; SFG.R: right superior frontal gyrus; SD: standard deviation.

3.3. Global level differences in cortical network characteristics

There were significant differences in the three global level parameters of the gamma frequency band. Post-hoc ANOVA revealed that the CC (p < 0.001) and PL (p = 0.021) in LLD-SI were significantly lower than those in LLD-NSI. Meanwhile, E_glob (p = 0.015) was significantly higher in LLD-SI compared to LLD-NSI. Similarly, CC (p = 0.004) and PL (p = 0.004) in LLD-SI were significantly lower than in NC, while E_glob (p = 0.004) was significantly higher (see Fig. 2). No significant group differences were observed in these parameters within the high alpha frequency band (details see Table S1).

Fig. 2.

Group differences of global network topological parameters in gamma band among LLD-SI, LLD-NSI, and NC. Abbreviations: CC: average clustering coefficient, PL: characteristic path length, E_glob: global efficiency, LLD-SI: late-life depression with suicidal ideation, LLD-NSI: late-life depression without suicidal ideation, NC: normal control.

3.4. Nodal level differences in cortical network characteristics

Based on the significant differences global level indices observed in the gamma band among three groups. We decided to explore the nodal parameters.

In the analysis of nodal CC, the interaction effect of group ^* ROI was also significant on nodal CC (F = 2.411, p = 0.000, η² = 0.051, Greenhouse Geisser correction). Post-hoc ANCOVA showed that the nodal CCs of LLD-SI were significantly lower than that observed in LLD-NSI in gamma band in right caudal anterior cingulate (p = 0.005), left isthmus cingulate (p = 0.030), left precuneus (p = 0.037), left (p = 0.043), and right (p = 0.046) lateral orbitofrontal. Meanwhile, when compared to NC, LLD-SI exhibited lower nodal CC in right caudal anterior cingulate (p = 0.013), left (p = 0.008) and right (p = 0.001) insula, left isthmus cingulate (p = 0.019), left (p = 0.004) and right (p = 0.001) lateral orbitofrontal, left (p < 0.001) and right (p < 0.001) medial orbitofrontal, left (p < 0.001) and right (p < 0.001) rostral anterior cingulate, right rostral middle frontal (p = 0.012), and left (p < 0.001) and right (p < 0.001) superior frontal. Besides, when compared to NCs, LLD-NSI exhibited significantly lower nodal CC in right insula (p = 0.031), right medial orbitofrontal (p = 0.002), left (p < 0.001) and right (p < 0.001) rostral anterior cingulate, and left (p = 0.003) and right (p = 0.010) superior frontal. However, nodal CC in left precuneus (p = 0.037) of LLD-NSI was significantly higher than NC (See Table 3 and Fig. 3 for details).

Fig. 3.

Group difference in nodal clustering coefficient among LLD-SI, LLD-NSI, and NC. Blue nodes indicate decreased nodal clustering coefficient in LLD-SI when compared to LLD-NSI or NC. Green nodes indicate decreased nodal clustering coefficient in LLD-NSI when compared to NC, while yellow nodes indicate enhanced nodal clustering coefficient in LLD-NSI when compared to NC.

Table 3.

Nodal clustering coefficients in brain regions with statistically significant differences in the gamma band across LLD-SI, LLD-NSI, and NC groups.

Brain regions (left/right)	LLD-SI (A) mean (SD)	LLD-NSI (B) mean (SD)	NC (C) mean (SD)	Effect size (partial η2)	F	p value (corrected)	Post-hoc
Caudal anterior cingulate R	0.186 (0.005)	0.207 (0.005)	0.205 (0.004)	0.122	6.266	0.003	A < B, C
Insula L	0.200 (0.006)	0.217 (0.006)	0.228 (0.006)	0.096	4.772	0.011	A < C
Insula R	0.192 (0.007)	0.205 (0.007)	0.230 (0.006)	0.150	7.950	0.001	A, B < C
Isthmus cingulate L	0.167 (0.006)	0.186 (0.005)	0.187 (0.005)	0.097	4.846	0.010	A < B, C
Lateral orbitofrontal L	0.220 (0.006)	0.239 (0.005)	0.246 (0.005)	0.116	5.924	0.004	A < B, C
Lateral orbitofrontal R	0.213 (0.005)	0.232 (0.005)	0.214 (0.005)	0.135	7.014	<0.001	A < B, C
Medial orbitofrontal L	0.216 (0.005)	0.228 (0.004)	0.244 (0.004)	0.181	9.932	<0.001	A < C
Medial orbitofrontal R	0.219 (0.005)	0.227 (0.004)	0.249 (0.004)	0.215	12.328	<0.001	A, B < C
Precuneus L	0.194 (0.006)	0.215 (0.006)	0.190 (0.006)	0.113	5.733	0.005	A < B, B > C
Rostral anterior cingulate L	0.225 (0.004)	0.230 (0.004)	0.255 (0.004)	0.262	16.002	<0.001	A, B < C
Rostral anterior cingulate R	0.219 (0.004)	0.228 (0.004)	0.253 (0.004)	0.273	16.873	<0.001	A, B < C
Rostral middle frontal R	0.223 (0.008)	0.232 (0.007)	0.255 (0.007)	0.095	4.739	0.011	A < C
Superior frontal L	0.225 (0.007)	0.234 (0.006)	0.265 (0.007)	0.188	10.410	<0.001	A, B < C
Superior frontal R	0.230 (0.008)	0.251 (0.007)	0.282 (0.007)	0.215	12.339	<0.001	A, B < C

Note: P_corrected (Bonferroni corrected).

In the analysis of E_i, the interaction effect of group ^* ROI was also significant on nodal clustering coefficient (F = 2.575, p = 0.000, η² = 0.054, Greenhouse Geisser correction). Compared to LLD-NSI, patients with LLD-SI showed significantly higher E_i in gamma band in the left parahippocampal (p = 0.028) and left precuneus (p = 0.027). In comparison with NC, LLD-SI exhibited significantly increased E_i in left (p = 0.027) and right (p < 0.001) parahippocampal, left (p = 0.004) and right (p = 0.001) precuneus, and left supramarginal (p = 0.004). Furthermore, E_i in right lateral orbitofrontal (p = 0.028) and medial orbitofrontal (p = 0.047) of LLD-SI were significantly lower than that observed in NC. Meanwhile, LLD-NSI exhibited significantly lower E_i in right lateral orbitofrontal (p = 0.024) compared to NCs. (Table 4 and Fig. 4 show the details).

Fig. 4.

Group difference in nodal efficiency among LLD-SI, LLD-NSI, and NC. Red nodes imply increased nodal efficiency in LLD-SI compared to LLD-NSI or NC. Blue nodes indicate decreased nodal efficiency in LLD-SI compared to NC. Green node indicates decreased nodal efficiency in LLD-NSI when compared to NC. Abbreviations: L: left, R: right, CAC: caudal anterior cingulate, ISCG: isthmus cingulate, PCUN: precuneus, LOF: lateral orbito frontal, MOF: medial orbito frontal, INS: insula, RACC: rostral anterior cingulate, RMF, rostral middle frontal, SFG: superior frontal, PHG: parahippocampal, SMG: supramarginal. LLD-SI: late-life depression with suicidal ideation, LLD-NSI: late-life depression without suicidal ideation, NC: normal control.

Table 4.

Nodal efficiency in brain regions with statistically significant differences in the gamma band across LLD-SI, LLD-NSI, and NC groups.

Brain regions (left/right)	LLD-SI (A) mean (SD)	LLD-NSI (B) mean (SD)	NC (C) mean (SD)	Effect size (partial η2)	F	p value (corrected)	Post-hoc
Lateral orbitofrontal R	0.252 (0.004)	0.252 (0.004)	0.266 (0.004)	0.097	4.842	0.010	A, B < C
Medial orbitofrontal R	0.263 (0.003)	0.265 (0.003)	0.271 (0.002)	0.068	3.290	0.042	A < C
Parahippocampal L	0.281 (0.003)	0.269 (0.003)	0.269 (0.003)	0.094	4.650	0.012	A > B, C
Parahippocampal R	0.284 (0.004)	0.275 (0.004)	0.263 (0.003)	0.148	7.822	0.001	A > C
Precuneus L	0.272 (0.004)	0.257 (0.004)	0.253 (0.004)	0.120	6.118	0.003	A > B, C
Precuneus R	0.263 (0.004)	0.254 (0.003)	0.243 (0.003)	0.143	7.535	0.001	A > C
Supramarginal L	0.249 (0.006)	0.230 (0.005)	0.223 (0.005)	0.116	5.894	0.004	A > C

Note: P_corrected (Bonferroni corrected).

3.5. Classification results

We evaluated three machine learning models to assess the classification performance of significant measures in classifying LLD-SI versus LLD-NSI. Model 1 included only the strong focal gamma connectivity indexed by PLV, model 2 was trained using graph theorical parameters, and model 3 incorporated both significant gamma connectivity and graph theorical parameters. Models 2 and 3 outperformed model 1, with minimal differences between models 2 and 3 (accuracy: 69.35% vs. 70.97%; AUC: 0.78 vs. 0.73). In model 2, the top three features were the global PL (40.60%), clustering coefficient of the right caudal anterior cingulate (12.55%), and nodal efficiency of the left precuneus. In model 3, the top three features were functional connectivity between the right lateral OFC and right medial OFC (13.62%), the global PL (12.29%), and the nodal efficiency of the left precuneus (9.43%). (see Table 5 and Fig. 5)

Fig. 5.

Receiver operating characteristic (ROC) curve and the area under the curve (AUC) of the classifier. The red line represents the ROC curve for model 1, trained solely on the PLV value of significant connectivity (AUC = 0.42). The blue line represents the ROC curve for model 2, which was trained exclusively on significant global metrics and nodal metrics based on graph theory (AUC = 0.78). The green line represents the ROC curve for model 3, which was trained by both significant connectivity and graph theory parameters (AUC = 0.73).

Table 5.

Performance of the support vector machine (SVM) classification model.

Model	Relative importance of the features (index by contribution, %)	Accuracy (%)	Sensitivity (%)	Specificity (%)	Model performance
1	FCLOF.R-RACC.L (31.05), FCLOF.R-RACC.R (19.46), FCMOF.R-RACC.L (15.14)	40.32	46.67	34.38	p = 0.0317
	FCMOF.R-RACC.R (11.87), FCMOF.R-MOF.L (9.56), FCLOF.R-MOF.L (9.46)
	FCLOF.R-SFG.L (2.62), FCLOF.R-LOF.L (0.84)
2	PLglob (40.60), CCCAC.R (12.55), EPCUN.L (12.26)	69.35	73.33	65.63	p = 0.0159
	CCglob (8.87), CCISCG.L (10.78), CCPCUN.L (5.69)
	EPHG.L (4.89), CCLOF.R (3.10), Eglob (0.91)
	CCLOF.L (0.36)
3	FCLOF.R-MOF.L (13.62), PLglob (12.29), EPCUN.L (9.43)	70.97	80.00	62.50	p = 0.0159
	FCLOF.R-RACC.R (9.41), FCLOF.R-RACC.L (8.21), Eglob (6.63)
	FCLOF.R-SFG.L (6.35), CCPCUN.L (6.19), CCCAC.R (5.84)
	FCLOF.R-LOF.L (4.61), FCMOF.R-RACC.R (3.35), FCMOF.R-RACC.L (3.30)
	FCMOF.R-MOF.L (3.24), CCLOF.L (2.80), CCISCG.L (2.05)
	CCglob (1.45), EPHG.L (0.93), CCLOF.R (0.28)

Note: LOF.R: right lateral orbitofrontal cortex; LOF.L: left lateral orbitofrontal cortex; MOF.R: right medial orbitofrontal cortex; MOF.L: left medial orbitofrontal cortex; RACC.R: right rostral anterior cingulate; RACC.L: left rostral anterior cingulate; SFG.R: right superior frontal gyrus; CAC.R: right caudal anterior cingulate; ISCG.L: left isthmus cingulate; PCUN.L: left precuneus; PHG.L: left parahippocampal; FC: functional connectivity; CC_i: the clustering coefficient of nodal i; CC_glob: the global (grand-average) clustering coefficient; E_i: the efficiency of nodal i; E_glob: the global (grand-average) efficiency; PL_glob: the global (grand-average) path length, measuring the average shortest distance between all pairs of nodes in a network. Contribution: the contribution can be interpreted as the absolute value of the weight for each feature in a linear kernel SVM. In the current study, the relative importance of features was computed through dividing each features’ contribution by the total sum of all contributions. The significance of the model performance was assessed using 10 000-bootstraping permutation tests.

3.6. Correlation analysis

No significant correlations were found between gamma-band network measures and neuropsychological or clinical scales (p > 0.05).

4. Discussion

In contrast to our prior EEG microstate analysis of large-scale temporal dynamics,²⁰ the current study investigated the frequency-specific EEG network among LLD-SI, LLD-NSI, and NC using functional connectivity and graph theory methods. Traditional EEG microstate analysis, based on broadband signals, primarily reflects large-scale brain network dynamics and does not capture network alterations and topological abnormalities within specific frequency bands. Based on our previous findings, the present study conducted a more targeted investigation of gamma-band network abnormalities. Rs-FC analysis revealed significant reductions in gamma connectivity focally within orbitofrontal areas in both LLD subtypes. Graph theory analysis showed decreased CC and PL, along with enhanced global efficiency (E_glob) in the gamma-band network of LLD-SI compared to LLD-NSI and NC. Several nodal metrics in several brain regions were identified, which can help differentiate between LLD-SI and LLD-NSI. In contrast to the gamma band, no significant differences in alpha-band graph theoretical metrics were observed between LLD-SI and LLD-NSI. Previous studies have shown that cortical alpha rhythms are strongly modulated by aging, with elderly adults exhibiting a gradual slowing of peak alpha frequency and reduced alpha power, as well as region-specific cortical alpha power abnormalities.³⁵ Given this age-related reduction in alpha activity, alpha-band network indices may have limited sensitivity for detecting SI within LLD.

In contrast, EEG network analysis focusing on the gamma band may more effectively capture the pathophysiological alterations associated with suicidality in LLD. Gamma oscillations play an important role in suicidal behaviors, with their generation associated with the activation of gamma-aminobutyric acid (GABA)ergic neurotransmitter systems.³⁶ GABAergic dysfunction is particularly pronounced in LLD, disrupting the balance between cortical inhibition and excitation.³⁷ This imbalance contribute to dysregulation of gamma oscillatory rhythms, potentially impairing cognitive functions and emotional regulation,^{38, 39} increasing the risk of suicidal behaviors.⁴⁰ Identifying disruptions in the large-scale brain network interactions of gamma oscillations might help predict suicide risk in LLD.⁴¹

The efficient human brain is characterized by a small-world network, with higher CC, shorter PL, as well as higher E_glob, indicating a more segregated and well-integrated network compared to a randomized network.^42,43 In our study, patients with SI exhibited decreased CC and PL, along with higher E_glob, suggesting a poor segregated and enhanced integrated brain network, with a trend toward the randomized network.⁴⁴ This trend has been linked to neuropsychiatry diseases, such as Schizophrenia, Alzheimer’s disease, and increased suicide risk in MDD.^44,45 The increased integration might indicate the compensatory mechanism to poor network segregation, which may serve as a protective factor against suicidal behaviors.⁴⁶

In the current study, significantly decreased nodal CCs were found in several brain regions in patients with LLD-SI. Among these nodal-level CCs, the cingulate cortex, including the isthmus cingulate, caudal and rostral anterior cingulate, belonging to the limbic system, which is responsible for emotion regulation.⁴⁷ The ACC has extensive functional interactions with cortical and subcortical regions, and abnormality in ACC has been regarded as neuro-biomarker for suicidal risk.^7,48 CCs in lateral and medial orbitofrontal (LOF and MOF) were significantly lower in LLD-SI. Studies have shown that abnormalities in the OFC may affect decision-making in MDD and increase the likelihood of impulsive behaviors, potentially leading to suicide attempts.^49,50 The precuneus has also been frequently investigated in studies of suicidal behavior in depression.⁵¹ Notably, the brain regions with abnormal CCs mainly belonged to the DMN, suggesting that impaired default brain function in the gamma band may be associated with SI in LLD.

Additionally, LLD-SI showed increased E_i in brain regions such as precuneus and parahippocampal. This increased E_i may act as compensatory mechanisms for the impaired network segregation caused by decreased nodal CCs in LLD-SI. Precuneus is one of the hubs of DMN, is involved in executive control, and has extensive connections across the cortex.⁵² Decreased functional connectivity in precuneus has been found in depressed individuals with suicide attempt.⁸ Researchers suggested that the antidepressant response of patients with depression is dependent on the high connectivity of the precuneus.^52,53 The parahippocampal gyrus supports the autobiographical memory, and impaired autobiographical memory has been found in depression with SI.⁵⁴ A task-based fMRI study revealed that prolonged reward delays in depressed individuals with suicide attempts were associated with reduced parahippocampal responses.⁵⁵ Previous research also indicates a connection between reduced parahippocampal activity and suicide attempts in depression.⁵⁶ Hence, we can infer that increased activity within these hubs may serve as a protective factor against the progression of LLD to suicidal behavior. However, hyperconnectivity within these hubs always incurs a metabolic cost in neural systems.⁵⁷ Once the overload persists or worsens without intervention, it may lead to a metabolic crisis, impairing node function and disrupting network organization.⁵⁸ However, it remains unclear whether decompensation in these hubs will exacerbate SI or lead to suicide attempts. Therefore, including LLD with suicide attempts in our future investigation is crucial.

Notably, reduced CC, decreased E_i, and diminished functional connectivity were both observed in LOF and MOF in LLD-SI. This may also indicate further deterioration within these regions, implying that OFC might no longer be capable of performing its compensatory role within the network in LLD-SI. The OFC plays a crucial role in the etiology of depression and has attracted considerable attention in research on SI in depression.⁵⁹ Therefore, we infer that impaired segregation and integration abilities in the OFC region may represent key mechanisms underlying LLD-SI.

Comparing model 1 (using graph theory metrics) and model 2 (using rs-FC metrics), we observed that the model using graph theory analysis in the gamma band exhibited significantly stronger accuracy, sensitivity, specificity compared to the rs-FC model. In model 3, adding connectivity parameters slightly improved classification sensitivity by 6.67%, but reduced specificity by 3.13% and AUC by 3%. These findings suggest that approaches revealing the topological properties of altered large-scale networks, such as graph theory, may be more effective in distinguishing SI in LLD patients.

Our findings have potential clinical implications, particularly in overcoming the challenges of identifying SI in LLD. Additionally, we identified several focal regions of abnormal activity in LLD-SI, such as the ACC, OFC, and precuneus. These results support the idea that these areas may serve as potential targets for neuromodulation to regulate emotions, improve cognition, and reduce suicide risk.^60,61

4.1. Limitations

There are some limitations in the present study. First, the sample size was relatively small. Second, we focused solely on LLD with SI. Future studies should include LLD with suicide attempts, allowing for a comprehensive investigation of the full continuum of suicidal behavior, from ideation to action, and facilitating a deeper understanding of this complex phenomenon. Moreover, although 64-channel EEG used in the current study meets the preliminary requirements for source-reconstructed graph theory analysis,⁶² future research could benefit from the use of higher-density EEG systems, such as 256-channel devices, to enhance the accuracy of spatial localization. Furthermore, we utilized the MNI152 template as head model for source localization. Future work will incorporate individualized head models for more precise and accurate source reconstructions. Finally, although we applied leave-one-out cross-validation and permutation tests to assess the robustness of the classification model, the absence of an independent test set raises concerns about potential overfitting. Future studies should incorporate external validation datasets to confirm the clinical utility of the proposed classification framework.

5. Conclusion

In the current study, we utilized EEG source-level graph theoretical measures, revealing abnormal network indices in LLD-SI compared to those without SI and NC. Specifically, we observed a trend from a small-world network toward a more randomized network in LLD-SI. Moreover, we identified that network indices in the gamma band could serve as potential biomarkers to differentiate between LLD-SI and LLD-NSI. These findings have implications for suicide prevention and intervention in older adults by offering potential markers for tracking SI and identifying network-based neuromodulation targets in LLD-SI. However, future research should involve larger sample sizes, higher-density EEG systems, and improved experimental designs to confirm these findings and fully understand the complex neurobiological mechanisms of suicidality in LLD.

Data availability

The raw and processed data generated in this study are not publicly available because of ethical restrictions and the need to protect participant privacy. However, these data can be made available upon reasonable request to the corresponding author for the purpose of verifying the research findings. Requests will be reviewed to ensure compliance with ethical and legal guidelines.

Author contributions

Conceptualization: YL

Data curation: YL, YZ, ZW, BC, MZ, GL, JL, QW, DX, KY, YC, YN, XZ

Formal analysis: YL, YZ

Investigation: YL

Methodology: YL, YN, XZ

Supervision: YZ, BC, MZ, GL, JL, QW, KY, YC, YN, XZ

Writing – original draft: YL, YZ

Writing – review & editing: YL, ZW, BC, DX, YN, XZ

Supplementary material

Supplementary data are available with the article at https://doi.org/10.1139/jpn-25-0068.