Non-linear EEG analyses predict non-response to rTMS treatment in major depressive disorder

Highlights

• There is no difference between MDD and HC in non-linear EEG measures.

• The change in LZC across time was significantly different for non-responders as compared to responders to rTMS treatment and HC.

• Non-linear EEG measures have added value over linear EEG measures in predicting treatment outcome.

Abstract

Objective

Several linear electroencephalographic (EEG) measures at baseline have been demonstrated to be associated with treatment outcome after antidepressant treatment. In this study we investigated the added value of non-linear EEG metrics in the alpha band in predicting treatment outcome to repetitive transcranial magnetic stimulation (rTMS).

Methods

Subjects were 90 patients with major depressive disorder (MDD) and a group of 17 healthy controls (HC). MDD patients were treated with rTMS and psychotherapy for on average 21 sessions. Three non-linear EEG metrics (Lempel–Ziv Complexity (LZC); False Nearest Neighbors and Largest Lyapunov Exponent) were applied to the alpha band (7–13 Hz) for two 1-min epochs EEG and the association with treatment outcome was investigated.

Results

No differences were found between a subgroup of unmedicated MDD patients and the HC. Non-responders showed a significant decrease in LZC from minute 1 to minute 2, whereas the responders and HC showed an increase in LZC.

Conclusions

There is no difference in EEG complexity between MDD and HC and the change in LZC across time demonstrated value in predicting outcome to rTMS.

Significance

This is the first study demonstrating utility of non-linear EEG metrics in predicting treatment outcome in MDD.

Introduction

The application of repetitive transcranial magnetic stimulation (rTMS) for major depressive disorder (MDD) has been investigated intensively over the last 15 years. Several meta-analyses have demonstrated that compared to placebo, the effects of rTMS applied to the left or right dorsolateral prefrontal cortex (DLPFC) have antidepressant effects (Berlim et al., 2013, Schutter, 2009, Schutter, 2010). With the establishment of the efficacy of rTMS, there has been increased interest in finding predictors of clinical response. The value of clinical features in predicting treatment outcome in MDD is very limited (Bagby et al., 2002, Simon and Perlis, 2010) and a shift towards biomarkers is evident, evidenced by new initiatives such as the NIMH research domain criteria (RDoC) and precision medicine.

Since the first description by Lemere in 1936 who described that the capability to ‘…produce “good” alpha waves…’ was associated with the ‘…affective capacity of the individual…” (Lemere, 1936), increased EEG alpha power is an often replicated finding in MDD (Olbrich and Arns, 2013). Recent studies further suggest that alpha power can be considered an endophenotype, mediating the pathway between the Brain Derived Neurotrophic Factor (BDNF) Val66Met polymorphism and depressive mood (Gatt et al., 2008, Zoon et al., 2013). Increased alpha power has also been associated with a favourable treatment outcome to antidepressant drug-treatments (Bruder et al., 2008, Bruder et al., 2001, Tenke et al., 2011, Ulrich et al., 1986), and a slow alpha peak frequency (APF) has been associated with less favourable treatment response to rTMS and antidepressants (Arns et al., 2012, Arns et al., 2010, Conca et al., 2000). Conceptually, the increased posterior alpha power can be regarded as a hyperstable vigilance regulation resulting in MDD behaviour, and it is known that treatments that destabilize vigilance (e.g., sleep deprivation) have strong antidepressant effects (Hegerl and Hensch, 2012). Given this supposed effect of destabilization, we hypothesised whether nonlinear measures would be more appropriate to assess the predictive power of these two dimensions of alpha (its amplitude fluctuations and frequency domain).

Nonlinear measures such as global and local scaling exponents and quantification of phase space dynamics have been successfully used to distinguish between healthy and pathological time series of human physiology and performance (e.g., Attention deficit disorders Gilden and Hancock, 2007; Ageing and disease Goldberger et al., 2002, Little et al., 2006; Developmental dyslexia Wijnants et al., 2012).

Fitting within a more general framework of health, wellbeing and complexity (Stam, 2010, Van Orden et al., 2009), analyses of neurophysiological time series and functional and structural networks of the brain also reveal dynamics such as long-range temporal correlations (LRTC) and topological structure associated with self-organised critical states of complex systems (Berthouze et al., 2010, Freeman et al., 2009, Orlandi et al., 2013, Palva et al., 2013, Rubinov and Sporns, 2010, Tagliazucchi et al., 2012, Yu et al., 2013). Nonlinear analyses of neurophysiological data have been used to characterize sleep, coma, anaesthesia, epilepsy, drug effects, schizophrenia, Alzheimer’s disease and dementia as deviations from such optimal critical states or “small-world” network topology (Stam, 2005 for a review).

Nonlinear measures have been used to study MDD, Linkenkaer-Hansen et al. (2005) found a breakdown of LRTCs in the theta band (3–7 Hz) range in MDD patients during eyes-closed wakeful rest. The theta band LRCT magnitude over the left temporocentral region was negatively associated with severity of depression (r = −0.79), whereas the LRCT magnitude in the alpha band over the occipitoparietal region was positively correlated with severity (r = 0.59). These results seem to be in line with the reported increased alpha in MDD mentioned earlier, however, Linkenkaer-Hansen et al. (2005) did not find any correlations between amplitude and severity in different frequency bands which warrants further study of the relationship between these measures in MDD. In addition, few studies exist where nonlinear analyses have been applied to investigate efficacy of intervention in general and to the best of the authors knowledge, no studies exist where they have been used to predict treatment outcome in MDD to rTMS or antidepressant medication. Two studies have investigated non-linear measures and their association to treatment outcome to Electroconvulsive Treatment (ECT), in which a smaller post-seizure fractal dimension (Gangadhar et al., 1999) or smaller Largest Lyapunov Exponent (LLE) were interpreted to indicate a more predictable pattern of EEG seizure activity (Krystal et al., 1997) that was associated with a more favourable treatment outcomes.

The goal of many intervention studies using nonlinear analyses is to evaluate the added value of these methods above more traditional linear methods (Granic and Hollenstein, 2003, Hayes et al., 2007, Lichtwarck-Aschoff et al., 2012). Especially in the context of physiological measurements, the potential for clinical applications of these techniques receives much attention (Bravi et al., 2011, Harbourne and Stergiou, 2009, Huikuri and Stein, 2012). Although the potential of nonlinear tools is widely recognised, Bravi et al. (2011) note in their evaluation of 70 different variability analyses that the challenges for the field are to develop a shared vocabulary and increase coherence between results of different studies and the techniques that were used. With these challenges in mind and given the results described earlier, in this study we compared three non-linear EEG metrics and their added value to linear metrics to predict treatment outcomes. For this study the dataset from Arns and colleagues (2012) was used which consisted of 90 MDD patients, all treated with rTMS and psychotherapy. In this previous study the slow alpha peak frequency (APF) was the strongest measure in the discriminant analysis, hence in the present study we investigated the non-linear dynamics of the alpha band, and evaluated whether such measures would add to the prediction of treatment outcome, using exactly the same EEG channels and frequency bands as reported in the previous study. The alpha band employed by Arns et al. (2012) was 7–13 Hz for posterior APF and 6–13 Hz for anterior APF. These bandwidths differ from traditionally used frequencies that have studied alpha power. However, it is well known that the APF can range from 4–6 Hz in very young children (Niedermeyer and Da Silva, 1999), to 10 Hz in adults and 7–9 Hz in chronic pain patients (Boord et al., 2008) and Alzheimer patients (Rodriguez et al., 1999). Given that specifically the slow APF values have yielded predictive power (for overview see Arns and Olbrich, 2013) and LRCTs indicate power laws in which slow frequencies dominate the signal (Babloyantz, 1991, Linkenkaer-Hansen et al., 2005, Linkenkaer-Hansen et al., 2001), we used the above frequency band. A second reason was to keep the results comparable to the previous report and contribute to improve coherence between different studies and their interpretation and evaluation of these nonlinear measures (cf. Bravi et al., 2011). Furthermore, in previous analyses as well as in the present study, the focus will be on predicting non-response to treatment. This excludes the effects of placebo response and differential effects of rTMS and psychotherapy. Characterizing non-response to treatment is not only clinically meaningful, but could potentially also result in the development of new treatments based on such biomarkers.