Resting-state EEG distinguishes depression in Parkinson’s disease

INTRODUCTION

Depression is a prominent non-motor symptom of Parkinson’s disease (PD) [1]. PD-related depression affects ∼20%–40% of PD patients, several times the expected prevalence within this population [2]. Importantly, this aspect of PD is often missed by physicians, contributing to morbidity and decreased quality of life [3–6]. Despite its significance and impact [7], it is unclear which brain circuits contribute to PD-related depression [8]. Determining which brain circuits are involved could lead to the development of new diagnostic tools to identify PD-related depression, as well as targeted treatments such as neuromodulation [9]. A fast and accurate neurophysiologically-based diagnostic tool may also facilitate neuromodulation. In addition, a better understanding of depression in PD may help us illuminate fundamental mechanisms of both diseases. PD and depression involve several overlapping circuits and associated neurotransmitters, including dopamine and serotonin [10]. These projection systems affect cortical physiology [11,12]. Cortical regions can be profoundly dysfunctional in PD [13] and in depression [14].

One technique that is particularly well-suited to capture cortical neurophysiology is electroeencephalography (EEG), which uses scalp electrodes to record activity from the cortex via an array of scalp electrodes. An early EEG study comparing depressed and non-depressed PD patients found widespread differences in alpha bands (8-13 Hz) in posterior and frontal sites [7]. Quantititave EEG (qEEG) studies have found spectral differences that distinguished PD vs depresson [15]. Furthermore, prefrontal cortical regions are responsive to targeted interventions, such as transcranial magnetic stimulation [16]. Here, we tested the hypothesis that spectral features of EEG can distinguish PD patients with depression.

We tested this hypothesis by collecting resting-state scalp EEG in PD patients with and without depression. We compared these data with control patients with depression but without PD. We report three main results. First, PD patients with depression had globally attenuated alpha (8–13 Hz) rhythms, as well as attenuated central beta (13–30 Hz) and gamma (30–80 Hz) rhythms relative to PD patients without depression. Second, PD patients with depression had strong global differences in gamma rhythms relative to non-PD patients with depression. Third, we used a linear predictive coding of EEG Algorithm for PD (LEAPD) formulated by Anjum et al. [17,20], which provides binary classification based on resting-state EEG power spectra. LEAPD-based classification accurately identified PD patients with depression relative to PD patients and non-PD depressed patients. Collectively, these data implicate cortical rhythms in PD-related depression, which could lead to novel targeted therapies or new diagnostic biomarkers for this important non-motor aspect of PD.

METHODS

EEG recording and analysis

Resting-state EEG was collected from patients while they sat in a quiet room with their eyes open for two minutes. Scalp EEG signals were collected from 64 channels of an EEG actiCAP (Brain Products GmbH) using a high-pass filter with a 0.1-Hz cutoff and a sampling frequency of 500 Hz. Electrode Pz was used as a reference, and electrode FPz was used as the ground. We used recording methods described previously in detail using a custom EEG cap with Iz, I1, and I2 leads in place of FT9, PO3, and PO4 leads; these leads were not analyzed [17–19]. We also removed FP1, FP2, FT10, TP9, and TP10 channels as these channels are often contamined by artifact, resulting in 56 channels for pre- and post-processing. EEG activity at the reference electrode Pz was recovered by computing the average reference. Bad channels and bad epochs were identified using the FASTER algorithm and the pop_rejchan function from EEGLAB and were then interpolated and rejected, respectively. Eye blinks were removed using independent component analysis (ICA). All channels were low-pass filtered at 100 Hz. Power was calculated using the pwelch function and was normalized to the mean power between 0–100 Hz for each channel. Scalp topography was plotted using topoplot from EEGLAB in delta (1-4 Hz), theta (4-8 Hz), alpha (8-13 Hz), beta (13-30 Hz), and gamma (30-80 Hz; Figure 1) bands.

• Download figure
• Open in new tab

Figure 1: Scalp topography of relative EEG power in PD patients with depression.

A) Relative power in PD patients with depression (PDDEP) compared to PD patients without depression for delta (1-4 Hz), theta (4-8 Hz), alph (8-13 Hz), beta (13-30 Hz), and gamma (30-80 Hz). B) Relative power in PDDEP compared to non-PD patients with depression (DEP). Electrodes are indicated by black dots; electrodes with significant differences between groups via ranksum testing are shown with white diamonds. Data from 18 PD, 18 PDDEP and 12 DEP.

RESULTS

PD patients with and without depression had similar age (p = 0.23), motor function as measured by UPDRS (p = 0.22), and cognitive profiles as measured by the MOCA (p = 0.94 value; Table S1). We collected resting-state EEG data and compared scalp topography of relative power for PD patients vs PD patients with depression (PDDEP) at delta (1-4 Hz), theta (4-8 Hz), alpha (8–13 Hz), beta (13–30 Hz), and gamma bands (30–80 Hz; Figure 1A). We also compared scalp topography for PDDEP vs non-PD patients with depression (DEP; Figure 1B). These data illustrate that there can be band-specific differences that distinguish depression in PD.

Our machine learning approach, LEAPD, compress power spectra into a series of autoregressive coefficients that holistically captures the shape of each power spectra with a few numbers [17,20]. Here, we used LEAPD to classify PD vs PDDEP and PDDEP vs DEP from single channels, as well as combinations of two channels (Figure 2).

• Download figure
• Open in new tab

Figure 2: LEAPD Classification approach:

Flow chart of classification.

We first used LEAPD to discriminate 18 PD from 18 PDDEP patients across all EEG electrodes (Figure 3A). Single-channel accuracy for channel CP3 was 86% and for TP8 was 86% (Figure 3A). Combining both CP3 and TP8 resulted in an overall LOOCV classification accuracy of 97%. These channels had distinct LEAPD indices between PD and PDDEP (CP3: p = 0. 00009, Cohen’s d = 1.8; TP8: p = 0. 0.00004, Cohen’s d = 1.8; CP3+TP8: p < 0.001; Cohen’s d = 3.25; Figure 3B). Receiver-operator curves (ROCs) for these channels in predicting PD vs PDDEP are shown in Figure 3C.

• Download figure
• Open in new tab

Figure 3: Machine-learning classification of LEAPD.

A) We constructed LEAPD indices from LPC coefficients from electrodes CP3 and TP8 for PD patients without depression (PD) vs PD patients with depression (PDDEP). B) Receiv r-operating curves (ROC) for single-channel performance of CP3, TP8, and CP3+TP8 combined, and C) channel performance across single electrodes. Data from 18 PD and 18 PDDEP patients. D) We also generated LEAPD indices for PDDEP (green) compared to depressed patients without PD (DEP; dark green) at CP4, CPz, and CP4+CPz combined. E) ROC curves and F) single channel performance across single electrodes. Data from 12 PDDEP and 12 DEP patients.

In addition, we found that LEAPD was highly accurate in differentiating 12 PDDEP patients (selected at random from 18 total) from 12 DEP patients, with 96% single-channel signal accuracy for electrode CPz and 92% for electrode CP4. Combining both channels resulted in 100% classification accuracy (Figure 3F). For these electodes, LEAPD distguished PDDEP vs DEP (Figure 3D; CPz: p = 0.00004, Cohen’s d: = 4.3; CP4: p = 0.0007, Cohen’s d: 2.0; CP4+CPz: p = 0.0004; Cohen’s d = 4.3; Figure 3E).

Additionally, we performed a truncation analysis of CP3, TP8, and CP3+TP8 combined for PD vs PDDEP and of CPz, CP4, and CPz+CP4 combined for DEP vs PDDEP. Recorded EEG data were truncated from full-length samples to samples that were a fraction of the original length. LEAPD analysis was then performed on the shortened signal using the same hyperparameters as those of the original signal. Truncation fractions of 0.05, 0.33 and 0.67 were tested. Performance of the channels at each truncation fraction is shown in Table S2. Although truncation did reduce the accuracy of the channels, each channel still retained significant discriminatory ability at shorter signal lengths. The performance degraded gracefully with truncation, indicating that the signals chosen are likely measuring a fundamental difference in EEG behavior between classes, rather than an artifact of overfitting. It is notable that accuracy of greater than 85% was achieved from two minutes of resting-state EEG signals. Performance on truncated data is shown in Figure 4 for channels of interest for PD vs PDDEP (Figure 4A) and PDDEP vs DEP (Figure 4B). Collectively, these data suggest that spectral features of scalp EEG can distinguish depression in PD.

• Download figure
• Open in new tab

Figure 4: Truncation analysis of LEAPD-based classification.

A) Data from PD vs PDDEP for 5, 33, 67, and 100% of data for PDDEP vs PD for CP3, TP8, and CP3+TP8 and B) PDDEP vs DEP for CP4, CPz, and CP4 + CPz.

DISCUSSION

We explored the cortical basis of depression in PD using resting-state scalp EEG. We found that PD patients with depression had central differences in beta and gamma rhythms. We used LEAPD, a spectral machine-learning approach, to detect differences in EEG signals from two minutes of resting-state data from a single electrode, achieving accuracies of 97% for PD patients with and without depression and 100% for PD vs non-PD patients with depression.These data indicate that PD patients with depression can be accurately differentiated from PD patients without depression and from depressed non-PD patients using machine learning.

Depression is a complex disorder [22] involving many brain networks; however, one consistent finding is abnormal cortical function [14,23]. Scalp EEG studies have found dysfunctional alpha rhythms in depressed patients [24,25], a finding that we report here comparing PD patients with and without depression. Beta rhythms can be profoundly abnormal in PD [26] and our data here indicate that depression decreases resting-state beta, alpha, and gamma rhythms in PD. We find that many cortical regions are implicated in PD-related depression, including prefrontal and parietal regions that have been found in prior studies of depression [14,27]

These data suggest that EEG, which is relatively inexpensive and ubiquitously available, can identify PD patients with depression. This is important because depression can be missed in PD [3–5], and electrophysiological diagnostic tools may aid in this effort. We report that our spectral approach can rapidly, robustly, and accurately identify EEG signals from PD patients with depression. Our results are in line with previous efforts to use LEAPD to identify local field potentials from animal models of PD and EEG data recorded from PD patients and controls [17,20]. LEAPD-based techniques might have additional utility in settings where neurophysiology is common, such as during deep-brain stimulation surgeries, and they may be helpful for closed-loop control applications. Apart from being robust and accurate, LEAPD is amenable to fast implementation and can serve as a trigger mechanism for brain stimulation.

Our work is supported by prior qEEG studies describing that a single parameter can differentiate depression and dementia in PD [15]. An early study which averages across all EEG electrodes reported distinct scalp topography of depressed PD patients, focusing on alpha rhythms [7]. Our study is supportive of these differences, and we are able to localize these results to the left frontal electrodes. In addition, we find broader differences over central electrodes in beta and gamma bands, which may have been averaged out in prior work that averaged EEG signals from multiple electrodes. Finally, we used advanced machine-learning to distinguish PD patients with depression from both PD patients and non-PD patients with depression. Recent work has reported frontal differences in sleep in PD patients with depression[28], as well as differences between midline event-related potentials between PD patients with and without depression [29]. Our study extends these findings and helps define the spectral topography of resting-state EEG in PD patients with depression, and demonstrates the potential of machine-learning for identifying PD patients with depression.

In this manuscript, we illustrate these effects from relatively high-performing channels: CP3/TP8 in PD vs PDDEP and CP4/CPz in PDDEP vs DEP. We chose these exemplars to illustrate high-performing channel combinations from each comparison. However, we note that channels also had high performance, and could be used for classification and identificiation of depression in PD.

Our study has several limitations. First, our sample size was limited, although in line with prior EEG studies in PD patients with depression [28,29]. Second, all of our patients were medicated, and it is possible that medications could influence these EEG signals [30]. Third, our method of diagnosing depression and quantifying symptom burden in PD patients was distinct from the method used with non-PD patients, limiting comparisons between these groups. Finally, our LEAPD approach did not include an out-of-sample prospective test, though the truncation analysis does remove concerns of overfitting. Despite these shortfalls, our findings describe spectral changes in PD patients with depression compared to PD patients without depression and non-PD patients with depression. We report that LEAPD-based machine learning approaches can identify EEG signals from PD patients with depression. These data could help illuminate the cortical neurophysiology of PD-related depression and could help lead to new biomarkers or diagnostic tools.

AUTHOR DECLARATION

None of the authors have any potential conflicts of interest to disclose.

Supplementary Tables