Comparison of overnight trends in relative power for specific frequency bands, sleep stages, and brain regions between patients with depressive disorder and matched control subjects

Abstract

Most EEG studies of MDD examined mean differences in frequency-band-specific power between MDD patients and control subjects; however, a few studies looked for differences in temporal trends in power. We focused on overnight time trends in relative power using polysomnography studies of 544 patients with MDD and 1,662 age- and sex-matched controls. We sought to replicate our finding on MDD with an additional 653 patients with depressive disorder (DD) and 1,959 age- and sex-matched controls. For each subject, we estimated trends as regression slopes separately for 180 features defined by six frequency bands, five sleep stages, and six brain regions. Relative theta power during stage 2 (N2) non-rapid eye movement (NREM) sleep in the frontal and central regions of the brain increased more rapidly through time in MDD patients than in controls. Similar upward trends of relative theta power were also statistically significant in DD patients. If validated in a longitudinal study, the time trend in relative theta power in the N2 stage of the NREM sleep could potentially serve as a surrogate biomarker for monitoring the responses of patients with depressive disorders to treatment.

1. INTRODUCTION

Major depressive disorder (MDD) is one of the most prevalent and debilitating mental illnesses worldwide (Hasin et al., 2018; Monroe and Harkness, 2022; Otte et al., 2016; Slavich and Sacher, 2019). MDD is characterized by depressed mood, lack of interests, impaired cognitive function accompanied by physical changes such as weight gain or loss (American Psychiatric Association, 2022). Patients with MDD experience loss of interest or pleasure in daily activity and often have problems with sleep, eating, concentration, energy, and self-worth. MDD affects hundreds of millions of people world-wide (Collaborators, 2022). According to the National Institute of Mental Health (NIMH), in the United States alone, in 2021, an estimated 21.0 million adults (8.3% of adults) had at least one major depressive episode. The etiology of MDD is multifactorial and has both genetic and environmental components (Verduijn et al., 2015). MDD affects twice as many women as men (Hasin et al., 2018; Monroe and Harkness, 2022; Otte et al., 2016; Slavich and Sacher, 2019).

Electroencephalography (EEG) has been used to study MDD. EEG records brain oscillatory activities from which characteristics such as power, symmetry, coherence, and connectivity are derived (for recent reviews, see (Fernandez-Palleiro et al., 2020; Klooster et al., 2023; Mumtaz et al., 2015; Olbrich and Arns, 2013; Tsai et al., 2023)). Electrical signals in the brain have six main oscillatory rhythms – delta (0.25–4.0 Hz), theta (4–7 Hz), alpha (7–12 Hz), sigma (12–16 Hz), beta (16–30), and gamma (30–70 Hz), although these definitions vary somewhat from study to study. One aim of EEG studies is to identify non-invasive and cost-effective EEG-based biomarkers that not only can distinguish between patients and healthy controls but also are discriminative enough to help monitor response to treatment (Fernandez-Palleiro et al., 2020; Klooster et al., 2023; Mumtaz et al., 2015; Olbrich and Arns, 2013; Tsai et al., 2023). For MDD in particular, altered theta activity is emerging as one of the common themes, though not yet an established biomarker (de la Salle et al., 2020; Fernandez-Palleiro et al., 2020; Fitzgerald and Watson, 2018; Klooster et al., 2023; McLoughlin et al., 2022; Mumtaz et al., 2015; Olbrich and Arns, 2013; Widge et al., 2019; Wu et al., 2021).

Patients with MDD often have sleep-related disorders such as insomnia (Fang et al., 2019; Murphy and Peterson, 2015; Nutt et al., 2008; Riemann, 2020). Polysomnography (PSG) is a technique for studying sleep-related disorders overnight in a clinical setting. Various PSG studies showed that patients with MDD had altered rapid-eye movement (REM) sleep with shorter REM latency and increased duration of REM sleep along with decreased duration of slow-wave sleep (Armitage et al., 2001; Cheng et al., 2015; Emslie et al., 1994; Solelhac et al., 2023). MDD patients also exhibited abnormal spindle activity (Bovy et al., 2022; Nishida et al., 2016).

Few EEG studies on MDD have been carried out during sleep. Most EEG studies of MDD used routine EEG in which subjects are either at rest, with eyes open or closed, or performing a task. Such studies typically involved a short recording time ranging from minutes to an hour. Results from those studies suggest that the theta oscillations (4–7 Hz) are altered in MDD patients compared to healthy controls (de la Salle et al., 2020; Fernandez-Palleiro et al., 2020; Fitzgerald and Watson, 2018; Klooster et al., 2023; McLoughlin et al., 2022; Mumtaz et al., 2015; Olbrich and Arns, 2013; Widge et al., 2019; Wu et al., 2021). In contrast, overnight PSG studies include several hours of EEG recording, making it possible to study temporal changes in EEG power during sleep. The few existing sleep EEG studies of MDD primarily focused on temporal trends in slow-wave EEG power using an exponential relationship that had been proposed earlier (Borbely, 1982; Daan et al., 1984). The authors showed that the ratio of slow-wave activity in the first non-REM (NREM) period relative to the second NREM period (referred to as dissipation) was lower in patients with MDD than non-patients (Armitage et al., 2000; Cheng et al., 2015; Goldschmied et al., 2019; Hoffmann et al., 2000). These authors did not obtain patient-specific estimates of trend but, instead, provided an overall regression estimate by synchronizing the sleep data to the same number of NREM periods or hours of sleep. Moreover, little is known about temporal trends for other EEG frequency bands in MDD patients during sleep.

Consequently, we analyzed the patient-specific temporal trends of six different EEG frequency bands (delta, theta, alpha, slow sigma and fast sigma, and beta) and tested whether the temporal trends differ between patients with MDD and their matched controls. Recent reviews that focused on EEG features as non-invasive biomarkers for depression in general, not only for MDD, have suggested that both gamma and theta band power show promise as diagnostic biomarkers (de Aguiar Neto and Rosa, 2019; Tsai et al., 2023). These reviews together with our preliminary results for MDD prompted us to expand our investigation to include patients with DD and their matched controls.

Our goal in this paper is to examine temporal trends in relative power for different frequency bands, sleep stages, and regions of the brain, comparing patients with depression and control subjects. We conducted two separate analyses: one that compared trends between MDD patients and their matched controls and as second that in patients with depressive disorder (DD), distinct from the MDD patients, and their controls. We found that the main differences in overnight temporal trends in relative power between MDD patients and control subjects were restricted to the theta band during the N2 stage of NREM sleep. This same finding emerged from our analysis patients with DD and their matched control subjects.

2. METHODS

2.1. In-laboratory PSG data

We downloaded de-identified EEG data from 9,641 in-laboratory PSG studies carried out between January 2019 and March 2023 in an American Academy of Sleep Medicine (AASM) accredited sleep laboratory at the University of North Carolina at Chapel Hill. Each study included an electroencephalogram with at least six channels (frontal F3 and F4, central C3 and C4, and occipital O1 and O2, plus two reference channels M1 and M2) (Figure 1), two electrooculograms, submental and bilateral tibialis surface electromyograms, and an electrocardiogram. The multi-channel polysomnograms were recorded digitally and stored using a Natus polygraph (Natus^® SleepWorks^™ PSG Software). The Natus system uses a 32-channel amplifier with amplifier gain minimum of 100,000. For our PSG recording, the sampling rate was 256 Hz; bandwidth was from 0.1 Hz to 400 Hz. Studies were manually scored following guidelines from the AASM scoring manuals and interpreted by physicians certified by the American Board of Sleep Medicine.

Figure 1.

Schematic overview of the data and analysis. (A) Data processing and feature extraction. EEG band power for the six rhythms (delta, theta, alpha, slow sigma, fast sigma, and beta) were extracted from multitaper spectrogram analysis for each of the six channels (frontal F3 and F4, central C3 and C4, and occipital O1 and O2). (B) Linear regression analyses using log-transformed relative band power, absolute band power, and total absolute power as response variables and time as the dependent variable to estimate temporal trends. (C) Real examples of scatter plots of relative band power with respect to sleep time together with fitted trend lines (blue: MDD; red: controls). (D) Two-sided Wilcoxon signed-rank tests to compare temporal trends between MDD patients and controls for each feature defined by frequency band, sleep stage, and channels. We adjusted for multiple comparisons using the Bonferroni approach and regarded adjusted p values <0.05 as statistically significant.

2.2. Study subjects

2.2.1. MDD patients

Through retrospective review, we identified 686 PSG studies from 564 patients diagnosed with MDD with recurrent episodes. MDD diagnosis was extracted from electronic health records by the sleep technician at the time of the PSG study. The diagnosis was made by each patient’s physician before the sleep study. We expect that all diagnoses were based on criteria in the then current Diagnostic and Statistical Manual of Mental Disorders (DSM) (American Psychiatric Association, 2022), although we have no explicit confirmation of the criteria used for each patient. For patients with multiple PSG studies, we used the one with the youngest age (first study diagnosed with MDD). This resulted in 564 PSG studies from 564 MDD patients. After eliminating those with fewer than 120 minutes of sleep time, 554 PSG studies for 554 patients were selected for the subsequent analysis. We did not exclude any MDD patients based on comorbidities.

2.2.2. DD patients

We also identified 1,535 studies from 1,140 patients diagnosed with depressive disorder (DD). As with MDD, diagnosis of DD was made by each patient’s physician before the sleep study. None of the 1,140 patients were included in the MDD patients. Similarly, for each DD patient, we chose the PSG study with the first recorded diagnosis of DD and eliminated studies with fewer than 120 minutes sleep time. We did not exclude any DD patients based on comorbidities. We were able to use only 653 of these DD patients in our analysis due to limitations in the availability of matching controls (see below).

2.2.3. Control subjects

From the collection of PSG studies, we also identified 3,490 studies from 3,152 subjects without any diagnosis of MDD or DD as potential controls and required each study to have at least 120 minutes of sleep time. For subjects with multiple PSG studies, we used the one where the subject was youngest. We excluded any PSG studies for which the subjects were diagnosed with psychiatric illness including depression, bipolar, autism, schizophrenia, and schizoaffective disorders. We also excluded subjects diagnosed with other major medical issues such as cancer, diabetes, brain disease, and cardiopulmonary disease.

From the resulting list of 3,152 subjects, we randomly selected three matching control subjects for each MDD patient, i.e., 1,662 controls for MDD. Controls were, matched on sex and age (±3 yr) and selected without replacement – none was used for more than one MDD patient. Similarly, we obtained three matching controls for 653 DD patients. i.e., 1,959 controls for DD. We also wanted to minimize the number of control subjects who served as a control for both MDD and DD patients, but the number of available controls was limited. As a result, we had 1,225 control subjects who were matched with both an MDD patient and a DD patient. We chose to match three controls to each patient (instead of only one) to increase the power of statistical tests.

This study was approved by the Institutional Review Board of the University of North Carolina at Chapel Hill (UNC-CH) (IRB #21-1984).

2.3. EEG data processing

We used the Python MNE package to extract the six channel EEG data (Gramfort et al., 2013). Each channel was re-referenced with respect to its opposite mastoid channel M1 or M2 (F3-M2, F4-M1, C3-M2, C4-M1, O1-M2, and O2-M1) (Duce et al., 2014). To each channel, we applied a zero-phase notch filter to remove a power-line noise at 60Hz and a zero-phase bandpass filter to retain signals in the frequency range of 0.25 to 25.0Hz. We removed epochs (30 seconds) with excessive flat signal; specifically, we removed epochs with an amplitude standard deviation less than 1 μV or less than 0.2 μV for 5 consecutive seconds. To minimize movement artifacts, we further removed epochs with signal amplitude greater than 2,000 μV (slow wave EEG signals can have high amplitude, e.g., > 1000 μV, especially in children). We also excluded epochs without sleep-stage annotations.

For the remaining epochs, we obtained a spectrogram for each channel by using 30-second windows without overlap. We constructed each spectrogram using the multitaper technique in MNE (Prerau et al., 2017) with a time-half bandwidth product NW=2, which provided a frequency resolution of 0.13Hz. Subsequently, we extracted the band power in the following six bands: 0.25–4.0 Hz (delta), 4.0–7.5 Hz (theta), 7.5–11.0 Hz (alpha), 11.0–13.0 Hz (slow sigma), 13.0–15.0 Hz (fast sigma), and 15–25.0 Hz (beta). Next, we obtained the epoch-wide total power by summing power across the six bands. We applied additional outlier removal procedure to the data (supplementary text).

For every epoch that passed this outlier filter, we have a value for total absolute power. Similarly, we also have absolute band power in each of the six frequency bands for each epoch. For every epoch, we normalized band-specific absolute power by dividing band-specific absolute power by total absolute power. We refer to this normalized power as ‘relative power.’

2.4. Statistical Analysis

2.4.1. Analyses of sleep characteristics

We extracted the total sleep time, sleep efficiency, and sleep latency for both MDD and DD patients and their respective matched controls. For each matched set, we averaged the values of the three controls. Separately, for MDD and for DD, we carried out a Wilcox signed rank test on the patient minus control differences for each sleep characteristic.

2.4.2. Estimating temporal trends in EEG power

Each 30-second epoch can be indexed uniquely by a sleep time measured in hours after sleep initiation. We set the minimum number of epochs to be 10 for regression analysis. To estimate temporal trends, we assumed an exponential relationship (Armitage et al., 2000) between relative EEG power and sleep time, that is, y_x = e^α+βx, where y_x is relative power at sleep time x, α is an intercept parameter, β is the slope coefficient that we regard as a measure of trend, and x is sleep time in hours. After log-transformation, the resulting linear relationship between relative EEG power and time is:

$$\text{log}\left(y_x\right)=\alpha +\beta x.$$ (1)

If we fit regression models with the same form but with either band-specific absolute power or total absolute power as the response, we see that the trend in relative power can be represented by the difference in trends between the band-specific and total absolute power. Briefly, consider EEG power data for a single feature and a single subject. Let y_b,x be the EEG absolute power of band b at time x and y_T,x be the total absolute EEG power at time x. Accordingly, $\frac{y_{b,x}}{y_{T,x}}$ defines the relative EEG power of band b at time x. Adapting eq(1) to these additional response variables, we have:

$$\text{log}\left(y_{b,x}\right)={\alpha }_b+{\beta }_{b}x$$ (2)

$$\text{log}\left(y_{T,x}\right)={\alpha }_T+{\beta }_{T}x$$ (3)

where the subscripts on the parameters indicate the corresponding response. Then

$$\text{log}\left(\frac{y_{b,x}}{y_{T,x}}\right)=\text{log}\left(y_{b,x}\right)-\text{log}\left(y_{T,x}\right)=\left({\alpha }_b-{\alpha }_T\right)+\left({\beta }_b-{\beta }_T\right)x$$ (4)

Comparing eq. (1) to eq. (4), one can see that: α = α_b – α_T and β = β_b – β_T.

We estimated the coefficients in these three regression models using ordinary least squares separately for each subject and each feature. For each subject, we obtained 180 regression slope coefficients (5 sleep stages × 6 bands × 6 channels) for the relative power (eq. 1) and absolute band-specific power (eq. 2) but only 30 slope coefficients (5 sleep stages × 6 channels) for the absolute total power (eq. 3).

Our interest is in assessing temporal trends as quantified by the regression slope coefficients. We use the estimated coefficient to summarize trend while recognizing that the actual trajectory of relative or absolute power through time may be a complicated non-linear function that is not fully captured by this simple model (see supplementary text).

2.4.3. Analysis of estimated slopes

We used the estimated regression slopes as a measured outcomes when comparing temporal trends between MDD or DD patients and their respective matched controls. We conducted two analyses, one for MDD and one for DD, using the same procedure for each. For each patient, we have three age- and sex-matched controls. For each feature separately, we averaged the regression slope coefficients from the three controls to provide a single trend value for the controls for that patient. Next, for each of the 180 features, we calculated the patient-minus-control-average differences and performed a two-sided Wilcoxon signed-rank test. We adjusted for multiple comparisons using the Bonferroni approach and regarded adjusted p values <0.05 as statistically significant. We also computed 95% confidence limits for the mean patient-minus-control-average differences for each feature; these intervals were based on the t-distribution and were not adjusted for multiple testing. We followed the same procedure for the other power responses: band-specific absolute power and total absolute power, though the latter involved only 30 features.

3. RESULTS

3.1. Demographics of study subjects

We used data from 554 PSG studies from 554 unique patients our analyses. Among the 554 patients, 75 were listed as in (partial) remission (Table S1). For simplicity, we referred to the 554 patients with MDD with recurrent episodes as patients with MDD. The median and mean ages of the MDD patients were 53.3 and 51.8, respectively (Table 1). Because of matching, controls had the same sex distribution and a similar age distribution as the MDD cases (Table 1).

Table 1.

Sex and age distributions of the patients with MDD with recurrent episodes and DD and their matched controls

Category	Total	Sex	N	Min.	1st Qu.	Median	Mean	3rd Qu.	Max.
MDD	554	M	177	9.2	40.6	53.8	52.4	64.6	91.5
		F	377	13.4	40.8	53.2	51.5	63.8	86.5
Control	1,662	M	531	8.7	40.1	53.3	52.0	65.9	94.9
		F	1,131	9.1	39.9	52.0	51.3	63.4	88.8
DD	653	M	298	7.7	34.8	50.8	48.3	63.7	87.4
		F	355	7.6	35.3	49.0	47.7	61.3	88.8
Control	1,959	M	895	5.4	34.9	49.4	47.8	63.0	87.1
		F	1,064	4.6	35.5	49.0	47.4	60.2	89.1

Among the 554 patients, 487 (87.9%) had prescription for antidepressants/antipsychotics. A majority (57.7%) of patients had prescription for antipsychotics. Nearly half (46.7%) of the patients had prescription for selective serotonin reuptake inhibitors (SSRIs) and 26.8% for serotonin and norepinephrine reuptake inhibitors (SSNRIs). Ninety-one (16%) patients had prescription for serotonin antagonist and reuptake inhibitors (SARIs). About 12% of patients had prescription for tricyclic antidepressants (TCAs). Nine patients had prescription for serotonin modulators and stimulators (SMSs) and one for monoamine oxidase inhibitor (MAOI). The remaining 67 patients were not found to be prescribed any antidepressants/antipsychotics. Nearly half (44.9%) of the MDD patients were diagnosed with anxiety. A small fraction of the MDD patients (2.8%) had bipolar disorder. Detailed information for each MDD patient is provided in the supplementary Table S2A.

The mean and median ages for DD patients were 3–5 younger than the corresponding MDD mean and median ages (Table 1). For the DD patients, 103 (15.7%) had prescription for antipsychotics, 91 (13.9%) for SSRIs, 34 (5.2%) for SSNRIs, 24 (3.6%) for TCAs, 23 (3.5%) for SARIs, and 2 (0.3%) for SMSs. None had prescription for MAOIs. Sixty-six of the DD patients also had anxiety. None had bipolar disorder (Table S2B).

3.2. Total sleep time, sleep efficiency, and sleep latency

We saw no significant differences in total sleep time (p=0.40), sleep efficiency (p=0.10) or sleep latency (p=0.70) between MDD patients and their matched controls (Table S3). We did, however, see some differences in these sleep characteristics between DD patients and their matched controls: total sleep time (p=0.02), sleep efficiency (p=0.004), and sleep latency (p=0.10) (Table S3).

3.3. Temporal trends in EEG power: MDD

Among the 180 features, only five showed statistically significant differences in the temporal trend (slope) of relative power between patients with MDD and matched controls after Bonferroni adjustment for multiple testing (Table 2 and Table S4). All five features were for the relative theta power in the N2 stage of NREM sleep in the frontal (F3, F4), central (C3, C4), and left occipital (O1) regions of the brain (Figure 2A, Figure S1A). For these five features, the mean slope coefficients for relative power for both MDD and controls (e.g., Table 2, row 1, columns 3 and 4) were both positive, indicating that the relative theta power during the N2 stage of NREM sleep was trending upward both in MDD patients and in matched controls. The mean difference in slope coefficients (e.g., Table 2, row 1, column 5) was also positive, indicating that the upward trend was steeper in patients compared to controls (Table 2, supplementary text). To probe how the upward trend in relative theta power for these five features was related to trends in the numerator or in the denominator, we looked at the average estimated slopes both for absolute theta power and for total absolute power in patients with MDD and in controls. By and large, we saw little evidence of significant differences between patients with MDD and matched controls in either of these average slopes in the N2 stage of NREM sleep (Table 2). For all five features, both of these regression coefficients were negative on average for both patients with MDD and controls (e.g., Table 2, rows 2 and 3, columns 3 and 4), indicating that both absolute theta power and total absolute power trended downwards in N2 stage of NREM sleep regardless of MDD status. However, the magnitude of the average downward trend was smallest for absolute theta power among patients with MDD and largest for total absolute power also among patients with MDD; the magnitudes of average downward trends among controls were between the corresponding averages for patients. Consequently, the trends of absolute theta power and total absolute power diverged more rapidly across sleep time for patients with MDD than they did for controls (Table 2). In turn, the upward trend in relative theta power in these five features was larger in patients with MDD than in controls (Table 2, Table S4).

Table 2.

Average regression slope coefficients for the relative theta power, absolute theta power, and total absolute power in the N2 stage of NREM sleep in patients with MDD and controls.

Channel	Regression coefficient (slope)					p-value (Bonferroni)
	Mean of MDD	Mean of control	Mean difference (MDD-control)	95% confidence
			Relative theta power
C3	0.0212	0.0085	0.0127	0.0064	0.0190	0.0005
C4	0.0165	0.0053	0.0111	0.0049	0.0174	0.0003
F3	0.0294	0.0170	0.0124	0.0059	0.0190	0.0002
F4	0.0265	0.0135	0.0131	0.0064	0.0197	0.0001
O1	0.0232	0.0127	0.0105	0.0037	0.0173	0.01
O2	0.0225	0.0144	0.0082	0.0015	0.0148	0.07
Absolute theta power
C3	−0.0292	−0.0389	0.0097	0.0042	0.0152	0.4
C4	−0.0323	−0.0408	0.0085	0.0027	0.0143	0.8
F3	−0.0330	−0.0418	0.0088	0.0033	0.0143	1.0
F3	−0.0365	−0.0452	0.0087	0.0034	0.0140	0.5
O1	−0.0171	−0.0233	0.0062	0.0004	0.0119	1.0
O2	−0.0172	−0.0211	0.0038	−0.0027	0.0103	1.0
Absolute total power
C3	−0.0507	−0.0473	−0.0034	−0.0115	0.0047	1.0
C4	−0.0479	−0.0461	−0.0018	−0.0105	0.0070	1.0
F3	−0.0636	−0.0586	−0.0051	−0.0138	0.0036	1.0
F4	−0.0635	−0.0585	−0.0050	−0.0139	0.0039	1.0
O1	−0.0402	−0.0356	−0.0046	−0.0128	0.0037	1.0
O2	−0.0397	−0.0354	−0.0042	−0.0129	0.0044	1.0

Figure 2.

Violin plots of the slopes of the relative theta power (rate of change of the logarithm of the relative power per hour) in N2 sleep for the 544 MDD cases and the 1,662 matched controls (A) and for the 653 DD cases and the 1,959 matched controls (B). These distributions are truncated at the 10^th and 90^th percentiles to better display the quartiles; see Table S1 for corresponding distributions without truncation. All six channels (F3, F4, C3, C4, O1 and O2) are shown. For each channel, the distribution for patients (blue) is shown on the right whereas the distribution for the controls is shown on the left (red). The dashed line indicates the median whereas the dotted lines mark the 25^th and 75^th percentiles. * p ≤ 0.05, **p ≤ 0.01, *** p ≤ 0.001, **** p ≤ 0.0001.

3.4. Temporal trends in EEG power: DD

Fifteen of 180 features defined by frequency band, sleep stage, and channel showed statistically significant differences in temporal trend (slope) or relative power between patients with DD and matched controls after Bonferroni adjustment for multiple testing (Table 3 and Table S5). Five of those fifteen were five of the six theta band features during the N2 stage of NREM sleep (only channel O2 failed to reach statistical significance). As with MDD, patients with DD had, on average, a steeper upward trend in relative theta power during N2 stage of NREM sleep than matched controls. Both the absolute theta power and absolute total power showed downward trends in both patients with DD and their age- and sex-matched controls. Similar to MDD, the downward trend in absolute theta power was slower in patients with DD than controls and the download trend in total absolute power was faster in patients with DD than controls. Of course, this description reflects averages across subjects; slopes for individual subjects varied widely (Figure 2B and Figure S1B)

Table 3.

Average regression slope coefficients for the relative theta power, absolute theta power, and total absolute power in the N2 stage of NREM sleep in patients with DD and controls.

Channel	Regression coefficient (slope)					p-value (Bonferroni)
	Mean of DD	Mean of Control	Mean difference (DD-Control)	95% Confidence
Relative theta power
C3	0.0126	0.0056	0.0069	0.0009	0.0130	0.001
C4	0.0139	0.0034	0.0104	0.0052	0.0157	0.0001
F3	0.0232	0.0143	0.0088	0.0028	0.0148	0.001
F4	0.0222	0.0116	0.0105	0.0050	0.0160	0.0003
O1	0.0161	0.0108	0.0052	−0.0013	0.0118	0.02
O2	0.0158	0.0118	0.0041	−0.0017	0.0098	0.6
Absolute theta power
C3	−0.0359	−0.0402	0.0043	−0.0010	0.0096	1.0
C4	−0.0400	−0.0434	0.0034	−0.0017	0.0085	1.0
F3	−0.0407	−0.0437	0.0030	−0.0023	0.0082	1.0
F4	−0.0448	−0.0474	0.0026	−0.0020	0.0073	1.0
O1	−0.0247	−0.0251	0.0003	−0.0049	0.0055	1.0
O2	−0.0232	−0.0244	0.0012	−0.0041	0.0064	1.0
Absolute total power
C3	−0.0480	−0.0457	−0.0023	−0.0106	0.0061	1.0
C4	−0.0535	−0.0467	−0.0068	−0.0140	0.0004	0.3
F3	−0.0634	−0.0576	−0.0057	−0.0140	0.0025	0.4
F4	−0.0662	−0.0588	−0.0074	−0.0147	−0.0001	0.3
O1	−0.0405	−0.0357	−0.0048	−0.0133	0.0036	1.0
O2	−0.0389	−0.0359	−0.0030	−0.0102	0.0042	1.0

Of the remaining ten statistically significant features in comparing patients with DD to matched controls, all were in the wake stage and in either alpha band (all six channels) or delta band (channels C3, C4, F3, F4) (Table S5).

4. DISCUSSION

Theta rhythms are mainly generated in the hippocampus and a substantial body of research has demonstrated that the theta activity is involved in perceptual and cognitive processes, memory function, and spatial behavior (Begus and Bonawitz, 2020; Buzsaki, 2002; Hasselmo and Stern, 2014; Herweg et al., 2020; Kahana et al., 2001; Kaplan et al., 2012; Karakas, 2020; Klimesch, 1999; Nunez and Buno, 2021). The theta rhythms may also orchestrate the reactivation of memories during both wakefulness and sleep (Schreiner et al., 2018). It was also proposed that the theta rhythms in REM sleep allow the integration of novel memory traces that are previously consolidated during NREM sleep (Hutchison and Rathore, 2015).

Theta rhythms are a major operational mode of the hippocampus; functional impairment of its oscillatory mode is thought to be equally detrimental to mental health (Buzsaki, 2002). Indeed, numerous studies have suggested that altered theta activity is associated with neuropsychiatric disorders including depression and anxiety (Arns et al., 2015; Fernandez-Palleiro et al., 2020; Koller-Schlaud et al., 2020; McLoughlin et al., 2022; Newson and Thiagarajan, 2018; Shadli et al., 2021; Wutzl et al., 2021). While most studies concluded that patients with MDD had increased theta activity, reduced theta activity has also been reported (for recent reviews see (Fernandez-Palleiro et al., 2020; Newson and Thiagarajan, 2018)).

Although altered theta activity has emerged as a common theme in MDD, whether the theta activity changes during sleep in patients with MDD remained largely unexplored. Most of the existing electroencephalography (EEG) studies on MDD examined the mean power differences between MDD patients and controls, not differences in temporal trends of power. In this study, we compared temporal trends in EEG band power between 544 patients with MDD with recurrent episodes and 1662 age- and sex matched controls from overnight polysomnography studies. We found that patients with MDD had significantly larger upward trends of relative theta power in the N2 stage of NREM sleep in the frontal and central regions of the brain; this upward trend reached statistical significance on only one of the two occipital regions that we probed.

Our data showed that both absolute theta power and total absolute power trended downward during the N2 stage of NREM sleep in both patients with MDD and controls; however, the downward trend in absolute theta power was slower in patients with MDD than controls. Concomitantly, the downward trend in total absolute power was faster in patients with MDD than controls. As a result, the upward trend in relative theta power was faster in patients with MDD than in controls (Figure 3). Although relative theta power during the N2 stage of NREM sleep increased with time in both patients with MDD and controls, patients with MDD had lower relative theta power at the beginning of the N2 stage of NREM sleep than did controls. These two aspects together -- steeper upward trend but lower initial level -- in patients with MDD compared to controls mean that, by the end of the N2 stage of NREM sleep, the relative theta power in MDD patients was nearly as large as in controls; in a sense, the relative theta power during the N2 stage of NREM sleep in MDD cases caught up with that of controls over the course of a night’s sleep.

Figure 3.

Schematic summary of the major finding–as time increases, MDD patients display a greater divergence in relative theta power as compared to controls. Color coding: MDD patients (blue) and control subjects (red). (A) Average regression lines for the relative theta power (rate of change of the logarithm of the relative power per hour) in N2 sleep. (B) Average regression lines for the absolute theta power (rate of change of the logarithm of the absolute theta power) in N2 sleep. (C) Average regression lines for the absolute total power (rate of change of the logarithm of the absolute total power) in N2 sleep. In this schematic, to focus attention on slope differences, the span of the horizontal axis (time) is eight hours in each panel whereas the span of the vertical axis differs among panels; the intercepts for the two lines in each panel are set to be equal; the slopes depicted match estimates from theta rhythm from channel C4.

Exactly the same pattern with respect to relative theta power, absolute theta power, and total absolute during the N2 stage of NREM sleep occurred in patients with DD. We again found an upward trend in relative theta power, suggesting that the trend for relative theta power during NREM stage 2 sleep to increase across the sleep period might be common to depressive disorders, regardless of severity. This finding is surprising. Whether the steeper upward trend of the relative theta power seen in both MDD and DD patients is unique to depressive disorders or is common more broadly among mental disorders remains unclear.

Lastly, our analysis cannot specify the source of the theta rhythms - frontal-midline (Inanaga, 1998; Ishii et al., 1999; Mitchell et al., 2008) or hippocampus, both of which have been implicated in attention, cognitive, and mental activity, and memory.

Like many other studies, an aim of this study is to identify potential biomarker that could be used as a surrogate for treatment response. Several clinical trials of deep brain stimulation for treatment-resistant MDD have been reported (Alagapan et al., 2023; Delaloye and Holtzheimer, 2014; Dougherty et al., 2015; Holtzheimer et al., 2017). A recent landmark study demonstrated that deep brain stimulation resulted in upward trends in beta activity (measured by routine EE) in those who responded to 24 weeks of treatment (Alagapan et al., 2023). It would be interesting to see if the theta activity during the N2 stage of NREM sleep also changed in those patients.

Our analysis also indicated that the relative delta power had a slower downward trend among patients with MDD than among controls, though the difference did not reach significance after adjustment for multiple testing. This result is largely consistent with earlier finding in that slow-wave activity dissipation from the first NREM to the second NREM period was slower in patients with MDD (Armitage et al., 2000; Cheng et al., 2015; Goldschmied et al., 2019; Hoffmann et al., 2000).

Our study has a few marked strengths. First, our sample size is relatively large with 554 MDD patients and 1,662 controls with 653 separate DD patients and 1,959 controls. We focused on the severe cases of MDD in which patients have recurrent episodes; but we were able to replicate our main results in milder cases of DD. Secondly, the PSG studies were carried out in an AASM accredited sleep lab and thus are of high quality. Lastly, compared to routine EEG studies, our PSG recordings are long, thus permitting trend analysis.

Our study has several limitations. First, the data were not obtained from a well-designed prospective study. Thus, our patient population may not be representative of patients in the general population. Second, several limitations accrue through lack of information in our database. Though we have a record that a physician provided a diagnosis of MDD or DD, we do not know with certainty whether DSM criteria were used in every case. Further, we do not know the exact medication history for these patients, thus, cannot rule out the possibility that the differences we report might have resulted from medication effects. Several classes of antidepressants/antipsychotics suppress REM sleep with short REM duration and longer REM latency (Thase, 2006; Wichniak et al., 2017). Although our finding was for NREM sleep, we cannot completely rule out the possibility that antidepressants/ antipsychotics affected our finding. Third, we assumed a linear relationship between the log-transformed EEG power and time to estimate average trends (slopes). This assumption is likely be too simplistic and is unlikely to hold for all patients. We employed it as we felt that the linear slopes were a reasonable way to assess trends in an overnight study despite possible non-linear trajectories. Lastly, although we have suggested a steeper upward temporal trends in relative theta power as a potential biomarker, these estimated slopes can be small. Lastly, our data are not longitudinal, which limits our ability to demonstrate a clear association with clinically meaningful outcomes such as the presence or absence of depression, its progression, or response to treatment. Given the presumed heterogeneity in this population of patients, one should be cautious in making broad generalizations. It is important to have our findings validated in different cohorts followed longitudinally.

3. CONCLUSIONS

We analyzed the temporal trends in EEG power in 544 patients with major depressive disorders and 1,662 matched controls and a separate cohort of 653 patients with depressive disorder and 1,959 matched controls. We showed that relative power of the theta rhythms in stage 2 of the NREM sleep had a steeper upward trend in patients with major depressive disorders and in patients with depressive disorder compared to the trend in their respective age- and sex-matched matched controls. In a sense, the relative theta power during the N2 stage of NREM sleep in MDD and DD cases tried to catch up with that of controls over the course of a night’s sleep. The steeper trend is largely due to lower of the relative theta power at the beginning of the NREM sleep in MDD or DD patients compared to the controls. By the end of sleep, relative theta power in MDD or DD patients reached a level comparable to but still lower than that of controls subjects.

Highlights.

• Analyzed overnight time trends in relative power using polysomnography studies

• Included 544 patients with MDD and 1,662 age- and sex-matched controls

• Increase in relative theta power in NREM was steeper in MDD patients than controls

• Steeper trend reflects MDD theta power approaching that of controls by end of sleep

• Similar finding using 653 patients with DD and 1,959 age- and sex-matched controls

Availability of data

The data presented in this study are available on request from the corresponding author. The data are not publicly available due to ethical restrictions.