Pretreatment anterior cingulate activity predicts antidepressant treatment response in major depressive episodes
- First Online:
- Cite this article as:
- Rentzsch, J., Adli, M., Wiethoff, K. et al. Eur Arch Psychiatry Clin Neurosci (2014) 264: 213. doi:10.1007/s00406-013-0424-1
- 346 Downloads
Major depressive disorder leads to substantial individual and socioeconomic costs. Despite the ongoing efforts to improve the treatment for this condition, a trial-and-error approach until an individually effective treatment is established still dominates clinical practice. Searching for clinically useful treatment response predictors is one of the most promising strategies to change this quandary therapeutic situation. This study evaluated the predictive value of a biological and a clinical predictor, as well as a combination of both. Pretreatment EEGs of 31 patients with a major depressive episode were analyzed with neuroelectric brain imaging technique to assess cerebral oscillations related to treatment response. Early improvement of symptoms served as a clinical predictor. Treatment response was assessed after 4 weeks of antidepressant treatment. Responders showed more slow-frequency power in the right anterior cingulate cortex compared to non-responders. Slow-frequency power in this region was found to predict response with good sensitivity (82 %) and specificity (100 %), while early improvement showed lower accuracy (73 % sensitivity and 65 % specificity). Combining both parameters did not further improve predictive accuracy. Pretreatment activity within the anterior cingulate cortex is related to antidepressive treatment response. Our results support the search for biological treatment response predictors using electric brain activity. This technique is advantageous due to its low individual and socioeconomic burden. The benefits of combining both, a clinically and a biologically based predictor, should be further evaluated using larger sample sizes.
KeywordsDepression Treatment response prediction Delta Theta Anterior cingulum EEG
Major depressive disorder (MDD) is one of the leading contributors to the global burden of disease, and it is associated with extensive socioeconomic costs. Over the last decades, a considerable effort has been made to investigate new pharmacological, biological and psychotherapeutic strategies in the treatment for MDD. Although several algorithm-guided treatment regimens have been evaluated , individual pharmacological treatment for depression still follows a trial-and-error approach. Within a treatment regime, each step requires several weeks of treatment before the individual response to it can be determined. In the case of a non-response, a switch to the next step is recommended, which requires more time and prolongs treatment significantly. Hence, the urgent need to identify early response markers confirms the potential effectiveness of a chosen treatment so as to reduce the time needed to achieve substantial improvement in a particular patient.
One of the main clinical predictors for treatment response is symptom severity reduction during the early phase of therapy (e.g., clinical improvement after 2 weeks) [2, 3]. The analysis of treatment responses in 6,568 depressed patients from different single- or double-blinded clinical trials by Szegedi et al.  showed that a lack of early improvement is related to antidepressant non-response in the following weeks. In this study, only 4 % of patients without early improvement showed subsequent remission, whereas 90 % of patients who showed long-term remission had improved at week 2. However, only 25 % of improvers showed remission at week 4 and at later assessments. Comparable results were also found using a more naturalistic study design . These studies suggest that symptomatic improvement at week 2 is a sensitive predictor for treatment response. However, the low specificity of this marker limits its clinical use.
Other studies have evaluated biologically based response predictors, e.g., using genetic [5, 6], electroencephalographic [7, 8, 9]; for review see [10, 11] and metabolic brain imaging approaches (for review see [12, 13]). These studies suggest that searching for markers related to the affective neuronal network may be a promising strategy to identify valid treatment response predictors. Using different brain imaging techniques, it was shown that the anterior cingulate cortex (ACC), a key region within the affective neuronal network, may be a promising response predictor to pharmacological, psychological and biological treatment strategies (see for review ). However, brain imaging techniques such as PET, SPECT and fMRI are time-consuming, expensive and burdensome for patients, which limits their use in daily clinical practice.
Neuroelectric brain imaging techniques based on electroencephalography such as standardized low-resolution brain electromagnetic tomography (sLORETA) can be used to examine the functional activity within the affective network with the advantage of being clinically feasible and cost-effective . A first account of the usefulness of this technique in the search for an antidepressant response predictor was given by Pizzagalli et al. . This study found differences between nine responders and nine non-responders to an antidepressive treatment in pretreatment low-frequency activity within the rostral ACC using a whole cortical brain LORETA analysis. Korb et al.  and Mulert et al.  corroborate these results using the ACC as a region of interest. In Korb et al.’s  study, the ACC predicted response to medication with a sensitivity of 64 % and a specificity of 67 %.
The aim of this study was to evaluate the predictive value of a biological and a clinical treatment response predictor, as well as a combination of both using a naturalistic study design. Pretreatment resting brain activity served as a biological and early clinical improvement served as a clinical response predictor. We hypothesized that the biologically based predictor would be more specific and more sensitive than the clinical parameter. We also hypothesized that combining the biological and the clinical predictor would be superior to using each predictor separately.
The study was approved by the ethics committee of the Charité–Universitätsmedizin Berlin and was carried out in accordance with the guidelines set forth by the Declaration of Helsinki. Each subject gave written, informed consent before participating in the study.
Demographic and clinical characteristics
Patients, n = 31
Controls, n = 31
Responders, n = 11
Non-responder, n = 20
Age (range) (years)a
49.2 ± 14.0 (26–65)
45.8 ± 11.1 (21–65)
47.0 ± 11.9 (21–64)
Higher education (yes/no)b
18/11 (2 missing)
Partnership (single/not single)b
16/14 (1 missing)
Age of first manifestation
39.8 ± 11.7
38.1 ± 13.7
For all t tests, t < 0.9, p > 0.1
Number of major depressive episodes
3.7 ± 2.8
2.8 ± 2.8
Duration of episode before inclusion (weeks)
14.1 ± 22.7
18.7 ± 21.7
Duration of hospital stay from EEG to discharge (days)
44.5 ± 22.9
67.7 ± 25.2
t = −2.54, df = 29, p < 0.02
29.1 ± 7.0
24.7 ± 6.0
t = 1.85, df = 29, p = 0.08
HDRS week 4
7.4 ± 5.1
18.6 ± 5.5
t = −5.6, df = 29, p < 0.0001
Week 2 residual symptoms (as % of HDRS baseline symptoms)
51.7 ± 22.0
77.9 ± 25.3
t = −2.87, df = 29, p < 0.008
Clinical ratings were done using the Hamilton Depression Rating Scale (21-item version, HDRS) at baseline, and after 2 and 4 weeks of antidepressant treatment. Depressive patients were classified as responders when, at week 4, the HDRS score had improved at least 50 % compared to pretreatment score. Of the 31 patients examined, 11 were classified as responders and 20 as non-responders. At the time of EEG recordings, 27 patients had gone without antidepressive medication for at least several days: Five responders and 13 non-responders had been without medication for at least 5 days, and four responders and five non-responders had been without medication for at least 2 days. Prophylactic antidepressant medication prescribed prior to the patient’s inclusion in the study had been taken by one patient in each of the two groups on the day that EEGs were carried out. One patient of the responder group had been given 10 mg of olanzapine per day for 2 days for agitation, and one patient of the non-responder group had been given 15 mg of mirtazapine for sleep disturbance. Comedication at the time of EEG recording was as follows: (responders/non-responders) benzodiazepines (4/4), choralhydrate (1/3) and non-benzodiazepine sleep medication (1/1). Four weeks of antidepressive or augmentative medication was started after EEG and baseline ratings in all patients. Medication regimes were as follows: (responders/non-responders) amitriptyline or nortriptyline (2/2), reboxetine (1/4), venlafaxine (4/2), SSRIs (0/10), mirtazapine (0/1), lithium augmentation to either SSRI or mirtazapine (2/1) and olanzapine augmentation to venlafaxine (1/0). All patients also received ergotherapy and group psychoeducation by a physician or psychologist as part of a weekly, unspecific, individual therapy. At least three medical visits per week took place, as is standard in the department.
The control group consisted of 31 age- and sex-matched healthy subjects recruited by newspaper advertisements. Controls had no psychiatric history according to MINI  and had no first- or second-degree family members with an affective disorder. Neither control subjects nor patients in the study had any neurological or medical disorders affecting the CNS.
EEG recording and analysis
EEG recording was done before changing antidepressive treatment. Recordings were performed in a sound-attenuated and electrically shielded room adjacent to the recording apparatus (Neuroscan SynAmps model 5083, El Paso, TX). Subjects were seated with closed eyes in a slightly reclined chair with a headrest. Resting EEGs were recorded with 32 tin electrodes referred to Cz, using an electrode cap with additional electrodes arranged according to the International 10/20 System. The recordings were performed for at least 10 min on each patient. Fpz served as a ground. Eye movements were recorded across electrodes placed 1 cm lateral to the left eye. Electrode impedance was less than 10 kOhm. Data were collected at a sampling rate of 250 Hz (gain 5,000; analogous band pass filter 0.15–100 Hz). The EEG was analyzed off-line using “Brain vision analyzer” software (Version 2.01, Brain vision, Munich, Germany). Independent component analysis (ICA) was used for artifact identification and elimination. Periods with any phase artifacts (e.g., movements, technical artifacts) were excluded from further analysis. Tonic and recurring artifacts (e.g., electrocardiogram, ocular and tonic muscle artifacts) were corrected by elimination of the corresponding independent component before transformation back to the EEG. After ICA-based artifact correction, a resting EEG was referenced to average reference, filtered (0.5–60 Hz, 50 Hz notch) and segmented into 2-s epochs. For further analysis, all artifact-free 2-s epochs in the period between 3 and 6 min of the 10-min resting EEG were used. This time period was chosen to ensure that data were collected when patients were relaxed but not too tired.
Further analyses were done using a standardized, low-resolution brain electromagnetic (sLORETA) software package. sLORETA computes the standardized current density at each of 6,239 voxels in the cortical gray matter and the hippocampus of the digitized Montreal Neurological Institute (MNI) standard brain. EEG epochs were subjected to cross-spectrum analysis of the delta (1.0–4.0 Hz), theta (4.5–7.5 Hz) and alpha (8.0–12.0 Hz) EEG power bands. EEG cross-spectra data were converted to intracranial spectral power standardized current density, using the sLORETA transformation employed by the software package. sLORETA computed standardized current density as the linear, weighted sum of the scalp electrical potentials and then squared this value for each voxel to yield the power of standardized current density. sLORETA units are proportional to amperes per square meter .
The sLORETA software package was used to perform voxel-by-voxel, between-group comparisons (responders vs. non-responders and patients vs. controls) of the current density distribution, as well as the regression analysis between voxel current density and changes in depression symptom score for all three frequency bands. Statistical, nonparametric mapping (SnPM; corrected for multiple comparisons, randomization with 5,000 permutations) was performed for this purpose. For the details of the SnPM procedure, see Nichols et al. . To perform further analysis, sLORETA current density values were extracted for the significant clusters of the between-group comparisons. All tests were two-sided. To obtain current density cutoff values that best discriminated between responders and non-responders, we performed the receiver operating characteristic (ROC) test. The ROC graph plots the true-positive rate (sensitivity) against the false-positive rate (1—specificity). The area under the ROC plot (AUC–ROC) gives a single number ranging from 0.5, indicating no apparent distributional difference between the two groups of test values, to 1.0, indicating perfect separation of the test values of the two groups. An area of 0.8, for example, indicates that a randomly selected individual from the responder group has a larger current-density test value than a randomly chosen individual from the non-responder group for 80 % of the time. The ROC, in this case, does not mean that a positive result occurs with a probability of 0.8, nor does it mean that a positive result is associated with response 80 % of the time [21, 22].
Analysis in whole sLORETA-space
Comparing responders and non-responders
Comparison of depressive patients and healthy controls
For all further analyses, current density values of all voxels showing significant differences in the right pg/adACC (region of interest) between responders and non-responders were extracted for the delta power band.
Delta pg/adACC sLORETA current density as response predictor
Based on the receiver operating characteristic (ROC) analysis, a pg/adACC current density cutoff of 55.1 predicted response with 94 % overall accuracy (82 % sensitivity, 100 % specificity, 100 % positive predictive accuracy, 91 % negative predictive accuracy, 0 % false-positive rate and 0 % false discovery rate). In other words, all nine patients predicted to be responders were in fact responders at week 4. The same held true for non-responders: 20 of 22 non-responders at week 4 were correctly predicted to be non-responders (χ2 = 23.1, df = 1, p = 0.000002). The area under the ROC curve was 0.93 (see inserted small figure in Fig. 4) (SE 0.05, p = 0.0002, 95 % CI 0.83–1.0).
Week 2 improvement as response predictor
Based on the literature, early improvement was defined as reduction in symptom severity of at least 30 % at treatment week 2. Fifteen patients showed early improvement. Of these early improvers, eight (53 %) were responders and seven (47 %) were non-responders at week 4. Of the 16 non-improvers, 13 (81 %) were non-responders at week 4 and three (19 %) were responders at week 4 (χ2 = 4.05, df = 1, p = 0.044). Early improvement predicts response at week 4 with 68 % overall accuracy (73 % sensitivity, 65 % specificity, 53 % positive predictive accuracy and 81 % negative predictive accuracy). Figure 4 summarizes the results, showing a scatter plot of week 4 residual depressive symptoms (ordinate) and pg/adACC sLORETA current density (abscissa). The pg/adACC sLORETA cutoff for response prediction and the week 4 residual depressive symptoms’ cutoff for “real” week 4 response are marked by dashed lines, which divide the figure into four different boxes: Both the gray and the white upper boxes encompass the week 4 non-responders, while the lower boxes encompass the week 4 responders. Both boxes on the left encompass the pg/adACC-predicted non-responders, while the boxes on the right comprise patients who were predicted to be responders at week 4. The upper gray box encompasses all week 4 non-responders predicted to be non-responders (i.e., the true negatives), while the lower gray box encompasses all week 4 responders that were predicted to be responders (i.e., the true positives). Note that early improvers and non-improvers are marked with black and gray circles, respectively. Pretreatment pg/adACC current density misclassified two and early improvement misclassified three of the four 11-week responders as non-responders.
We investigated pretreatment resting EEG activity using standardized, low-resolution brain electromagnetic tomography (sLORETA) as a biological predictor of treatment response to a 4-week antidepressant medication in 31 inpatients with an acute major depressive episode.
Treatment responders showed higher delta power compared to both non-responders and healthy controls in the right anterior cingulate cortex (ACC), i.e., in the perigenual/anterior dorsal ACC (pg/adACC), as well as in a region located more rostrally in the perigenual ACC. The correlation between delta power in the right pg/adACC and symptom reduction underscored the robustness of this result. The second main finding of our study is that pretreatment delta pg/adACC power predicts response to treatment with higher accuracy than the clinical predictor, early improvement. While early improvement also did prove to be a sensitive predictor, this parameter had a lower specificity, a finding reflected by other studies . Combining both predictors did not further enhance accuracy.
The finding of differences between responders and non-responders in the pg/adACC highlights the importance of the ACC in the pathophysiology of depression [23, 24, 25]. Our results are in line with the studies employing PET, SPECT, fMRI or neuroelectric brain imaging methods, demonstrating that pretreatment activity of the ACC at rest as well as during stimuli processing is related to response to different pharmacological and biological antidepressant therapies [23, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35]. A recent meta-analysis conducted by Pizzagalli et al.  on 23 studies showed that pretreatment ACC activity can be regarded as a predictor of treatment response with a weighted effect size of 0.9 (95 % CI 0.44–1.39).
Pizzagalli et al.  was the first to demonstrate a more pronounced, low-frequency power in the ACC of patients showing better responses to treatment when compared to those showing a worse response. This was confirmed by the studies of Korb et al.  and Mulert et al. [18, 36] using a region of interest approach. It should be mentioned that, in both studies, the ACC regions were more rostral than in our study, which may be explained (in part) by the electrode configuration , the low spatial resolution of EEG-based techniques and the LORETA algorithm used (“old” LORETA vs. standardized LORETA ).
However, in contrast to these studies, our whole sLORETA-space analysis showed a relation between treatment responses in the delta band, but not in the theta band (Korb reported only theta band activity). This difference could not have been caused by different frequency-band definitions, as reanalyzing our data using the frequency-band definitions of these studies did not change our results (data not shown). In a post hoc reanalysis of our data, we extracted the mean theta power of our significant pg/adACC cluster as region of interest and compared it between responders and non-responders. In accordance with Pizzagalli et al. , Korb et al.  and Mulert et al. , responders showed significantly higher theta power in the pg/adACC than non-responders (sLORETA units 31.6 ± 25.9 for responders and 13.9 ± 6.4 for non-responders, t = 2.86, df = 28, p = 0.008; one male non-responder had to be excluded because of being a strong outlier). Theta- and delta-band power were highly correlated to each other (r = 0.84, p < 0.0001).
The link between the ACC and the response to an antidepressive treatment was demonstrated for different treatment strategies. It is suggested that increased activity in the ACC and related areas may be part of a compensatory neural mechanism that may increase the likelihood to respond to a given treatment. Interestingly, there is some evidence that low-frequency power in the ACC is related to lower levels of maladaptive rumination, which in turn may facilitate antidepressive treatment response. “Mind-wandering,” in depression often associated with rumination, is closely correlated with reduced self-esteem and negative self-reflection , causes performance difficulties in both healthy  and depressed individuals [40, 41] and represents a risk factor for depressive relapse . At the neurobiological level, “mind-wandering” and rumination are related to the default mode network (DMN) in healthy subjects [42, 43] and depressed patients  and are associated with both hyperactivity [45, 46, 47] and resting-state DMN hyperconnectivity in the anterior medial cortex [48, 49], especially the perigenual ACC is involved in self-processing and DMN activity . Meditation, which is the opposite of rumination, is related to enhanced slow-frequency power at frontal sides [51, 52] and to reduced functional connectivity between DMN regions involved in self-referential processing . Thus, our findings of higher pg/adACC slow-frequency power in the responder group may constitute an indicator for preserved DMN function and lower maladaptive rumination in this group. This may in turn facilitate treatment response, as slow-frequency power is negatively linked to DMN connectivity  and medial frontal DMN activity . Besides this interpretation, the association between rumination and treatment response should be further investigated, as it gives the possibility to identify subgroups that may be in need for an adjunct cognitive psychotherapy to an antidepressant treatment to reach response [56, 57].
However, some facts may complicate our interpretation: Firstly, we found no delta ACC power differences between non-responders and healthy controls, and secondly, Andersen et al.  showed a relationship between rumination about a personal conflict and low-frequency power at the parieto-occipital but not at the frontal electrodes. However, this study was limited to healthy subjects. Furthermore, the practical implications of our findings remain still unclear at this point. What does a pretreatment “non-response” prediction mean? It is difficult to ascertain from pretreatment-based data whether an individual patient would require an early escalation of therapy or would simply need more than 4 weeks to respond. The treatment response rate in our study was lower than the response rates reported in the literature. Response rates have been shown to be influenced by varying factors, e.g., somatic , genetic , sociodemographic  and clinical factors . The non-responder group had longer current episode duration, but this was not statistically significant. No other differences were found regarding clinical and demographic parameters between responders and non-responders. One of the main factors may be the time of response-to-treatment measurement. We measured response to treatment at treatment week 4, but did not record depression severity at a later stage. Thus, we cannot rule out that some patients of the non-responder group might have been late responders who would have showed treatment response at week 8 of treatment as seen in other studies (e.g., 54.9 % at week 8 of treatment ). However, the range of the proportion of responders reported in the literature is very wide (e.g., range 31.6–70.4 % ).
Besides the small sample size, we need to consider several other limitations that possibly reflect the reality of clinical practice. Importantly, our subjects were not free of medication at the time of the EEG recordings, and there was no washout phase of antidepressants for patients as is the case in other EEG studies (e.g., 17). Pre-EEG antidepressants were discontinued for some patients at lest 6 days (percentage of responder/non-responder: 45.5/65 %) and for other patients at least 3 days (36.4/25 %). Discontinuation of medication may well have had some effects on EEG activity . Further, more responders than non-responders received benzodiazepines (36.4/20.0 %) the night before EEG. However, benzodiazepines are known to enhance beta and reduce theta/alpha but may did not change delta absolute power . Moreover, pharmacological treatment strategies differed within the patients and between the groups, e.g., more patients in the responder group received venlafaxine. However, whether the effect of venlafaxine as compared to other antidepressants produces clinically important differences in 50 % reduction of depression symptoms remains unclear . In addition, although ocular blink artifacts were removed using independent component analysis, contamination of frontal brain activity by ocular artifacts cannot be completely excluded due to the limitations of EEG signal recording technology.
We conclude that low-frequency ACC power served as a biological treatment predictor with good predictive accuracy. As the measurement of electroencephalic activity is noninvasive and does not produce high distress in patients, it may be a suitable method for identifying a predictive biomarker of treatment response in daily clinical practice. Future studies with higher sample size should also combine biological and clinical predictors to test whether this approach may yield a more superior predictor.
J. R. received support from the NARSAD young investigator award 2009, National Alliance for Research on Schizophrenia and Depression.
Conflict of interest