The nine-week protocol employed two phases. Patients first entered a one-week, single-blind placebo lead-in phase. Patients who failed to meet study entry criteria at the end of lead-in (e.g., because of intolerable side effects to placebo, or a strong placebo response (no longer meeting inclusion criteria)) were removed from the protocol and were referred for open-label treatment. Patients eligible to continue in the protocol entered a double-blind phase and were randomized to receive 8 weeks of either placebo or active medication (fluoxetine 20 mg po QD in the first protocol or venlafaxine 150 mg po QD in the second) dispensed in identical capsules. Patients receiving fluoxetine were given 20 mg/d and continued at that dose for the 8 weeks; patients receiving venlafaxine began at 37.5 mg/d, increased over a week to 150 mg/d, and then continued at that dose for the remaining 7 weeks. To preserve blinding, placebo “dose” was escalated in that trial.

All patients received brief sessions of supportive psychotherapy during the blinded phase of the study, in order to address safety concerns about dispensing placebo alone to patients with significant depression (15-25 minutes of unstructured counseling and assistance in problem solving by a research nurse at the follow-up visits). Follow-up visits for symptom/side effect monitoring and for the supportive therapy took place at two days and at weekly intervals thereafter after the start of the double-blind phase of the study. Symptoms were monitored with a focused clinical interview, clinician rating scales (e.g., HAM-D), and self-rating scales. At the end of the 8 week double-blind phase, the blinding was broken and HAM-D scores were used to categorize patients as responders or non-responders.

QEEG recordings were obtained (a) at pretreatment baseline prior to randomization, (b) at 48 hours (after two doses of drug or placebo), and (c) after 1 week on medication or placebo, as shown in Fig. (1). Recordings were made with the QND System (Neurodata, Inc., Pasadena, CA), using procedures employed in our previous reports and summarized here. Patients were instructed to rest in the eyes-closed, maximally alert state, in a quiet room with subdued lighting. The technicians monitored the QEEG data in real-time during the recording, and re-alerted the patients every 30-45 seconds as needed to avoid drowsiness. Electrodes were placed with an electrode cap (ElectroCap, Eaton, OH) using 35 recording electrodes distributed across the head according to the International 10-20 System arrangement (Fig. 2). Data were collected using a Pz referential montage and were digitized at 256 samples/channel/sec by the QND system (bandpass filtered 0.3 - 70 Hz).

Fig. (1) Experimental Protocol Timeline. Subjects were assessed and enrolled at “intake”, had the pretreatment baseline EEG recorded at that time, and then participated in a one week, single-blind placebo lead-in phase. Randomization to treatment modality (active medication or placebo) took place at the time marked “start of treatment”, Another EEG was recorded after 48 hours of treatment and again at 1 week. Clinical assessment to determine outcome (responder vs non-responder) took place after 8 weeks of treatment (9 weeks in study altogether). Subjects were monitored weekly during the double-blind treatment (arrows not shown) for clinical changes and adverse reactions [10Cook IA, Leuchter AF, Morgan M, et al. Early changes in prefrontal activity characterize clinical responders to antidepressants Neuropsychopharamcology 2002; 27: 120-31.].

Fig. (2) Electrode montage. The 35 scalp electrodes from the extended International 10-20 system. Line segments denote bipolar channels used in the reattributional montage. Electrodes included in the calculation of average prefrontal cordance are Fp1, Fp2, and Fpz [14Cook IA, O’Hara R, Uijtdehaage SHJ, Mandelkern M, Leuchter AF. Assessing the accuracy of topographic EEG mapping for determining local brain function Electroencephalogr Clin Neurophysiol 1998; 107: 408-14.].

Fig. (3) QEEG CART tree of fluoxetine and venlafaxine. Δ2_DA_FP1 = the change from baseline to day 2 for δ absolute score at FP1. Δ7_TZ_AF2 = the change from baseline to day 7 for θ Z score at AF2.

Each EEG recording was reviewed by a technician who was blinded to the identity of the patient, treatment condition, and clinical status; the first 20-32 seconds of artifact-free data were selected to be processed. A second technician reviewed these selections for accuracy. These selections were then processed using a fast Fourier transform to obtain absolute and relative power values in four frequency bands (0.5-4 Hz, 4-8 Hz, 8-12 Hz, and 12-20 Hz). “Absolute power” describes the amount of power in a frequency band at a given electrode (measured in µV²), and “relative power” is the percentage (%) of power contained in a frequency band, relative to the total power across the entire spectrum (0.5-20 Hz) computed separately for each electrode. The QEEG data were reformatted offline to compute linked-ear-reference absolute and relative power values; these values were used in the regional measure analyses described below. QEEG data were also reformatted to bipolar channel pairs and processed further to yield cordance values (below).

Cordance was calculated by combining conventional QEEG absolute and relative power measures in a common metric, and was computed in three steps using methods we have detailed previously [13Leuchter AF, Cook IA, Uijtdehaage SHJ, et al. Brain structure and function, and the outcomes of treatment for depression J Clin Psychiatry 1997; 58(suppl 16): 22-31., 14Cook IA, O’Hara R, Uijtdehaage SHJ, Mandelkern M, Leuchter AF. Assessing the accuracy of topographic EEG mapping for determining local brain function Electroencephalogr Clin Neurophysiol 1998; 107: 408-14.] and described briefly here. First, EEG power values were computed using a re-attributional electrode montage (Fig. 2) because that montage afforded the highest correlation between EEG measures and PET measures of regional cerebral blood flow. Second, these values were normalized across all electrode sites using a z-transformation, yielding Anorm(s,f) & Rnorm(s,f) for all sites s and frequency bands f. Third, cordance values (Z) were formed as the sum of Anorm and Rnorm.

Z(s,f) = Anorm(s,f) + Rnorm(s,f)

Software utilized for statistical analysis included SAS version 8 (SAS Institute, Cary, NC), and for CART analysis, Salford Systems’ CART version 5.0 (Salford Systems, San Diego CA).

Patients were categorized both by treatment (fluoxetine, venlafaxine or placebo) and by outcome (response/non-response). A final 17-item HAM-D score of ≤10 at the end of the 8 week double-blind phase was used to define the responders [10Cook IA, Leuchter AF, Morgan M, et al. Early changes in prefrontal activity characterize clinical responders to antidepressants Neuropsychopharamcology 2002; 27: 120-31., 15Cook IA, Leuchter AF, Uijtdehaage SHJ, Abrams M, Anderson-Hanley C, Rosenberg-Thompson S. Neurophysiologic predictors of treatment response to fluoxetine in major depression Psychiatr Res 1999; 85: 263-73.]. As noted earlier, patients from the fluoxetine and venlafaxine trials did not differ significantly in response rates or demographic parameters and thus were pooled for these analyses. Outcomes were evaluated using the entire HAM-D scores, and also using two frequently-employed unidimensional subscales derived from the full HAM-D: the Bech Melancholia subscale [16Bech P, Gram LF, Dein F, Jacobsen O, Vitger J, Bolwig TG. Quantitative rating of depressive states Acta Psychiatr Scand 1975; 51: 161-70.] (Bech et al., 1975) and Maier-Philipp Severity subscale [17Maier W, Philipp M, Gerken A. Dimensions of the Hamilton Depression Scale. Factor analysis studies Eur Arch Psychiatr Neurol Sci 1985; 234: 417-22.]. These were considered because, in a meta-analysis by Faries et al. [18Faries D, Herrara J, Rayamajhi J, DeBrota D, Demitrack M, Potter WZ. The responsiveness of the hamilton depression rating scale J Psychiatric Res 2000; 24: 3-10.], they had shown greater size effects and applicability to smaller sample sizes than the full 17-item HAM-D. Both subscales include items 1 (depressed mood), 2 (feelings of guilt), 7 (work and activities), 8 (psychomotor retardation) and 10 (anxiety/psychic). The Bech subscale also includes item 13 (somatic symptoms/general), while the Maier-Philipp subscale includes item 9 (agitation).

Our primary analyses with QEEG measures employed data from each of the 35 individual electrodes. Based on prior significant findings for the prefrontal region, we also examined the prefrontal region by averaging the individual values from the three prefrontal channels (FP1, FP2, FPZ) as had been done in those earlier reports [9Cook IA, Leuchter AF, Morgan ML, Stubbeman W, Siegman B, Abrams M. Changes in prefrontal activity characterize clinical response in SSRI nonresponders: a pilot study J Psychiatr Res 2005; 39: 461-6., 10Cook IA, Leuchter AF, Morgan M, et al. Early changes in prefrontal activity characterize clinical responders to antidepressants Neuropsychopharamcology 2002; 27: 120-31.].

Data were analyzed using classification and regression tree analysis (CART) [11Brieman L, Friedman J, Olshen R, Stone C. Classification and Regression Trees. Monterey, CA: Wadsworth & Brooks 1984.]. This is a novel application of CART analysis to clinical and physiologic data in mood disorders, and so merits more extensive description. CART is a nonparametric statistical methodology that creates binary decision trees: the methodology recursively partitions the parameter space to classify patients by category with the fewest mistakes (termed purity). The binary decision tree contains nodes connected by branches. The root node contains all patients. A binary decision tree is created by partitioning a node into two daughter nodes. The objective of the analysis is to find partitions of the data such that the end nodes (terminal nodes) are as homogeneous as possible. The quantitative measure of node homogeneity is termed the impurity function. The simplest idealization of the impurity function is the number of patients who meet an objective criteria divided by the total number of patients in the node. Ratios close to 0 or 1 are considered more pure.

To partition a node, CART examines all possible splits of the explanatory variables. In general, the number of possible splits for ordinal or continuous variables is 1 less the number of distinctly observed values. A potential split is judged by its reduction of the impurity function for both daughter nodes it creates. The partitioning iteratively continues by splitting both of the daughter nodes into 2 daughter nodes and continues until the tree is saturated, that is, until no further partitions can be performed.

Although the tree grown may fit a given dataset well, it may be overfit to that sample and perform poorly with a different sample. The optimal number of terminal nodes can be determined by cost-complexity pruning [11Brieman L, Friedman J, Olshen R, Stone C. Classification and Regression Trees. Monterey, CA: Wadsworth & Brooks 1984.]. Cost complexity is defined as R_α (T) = R(T) + α |T|, where α is the cost-complexity parameter, R(T) is the misclassification error and |T| is the number of terminal nodes in tree T. Brieman has shown that for any value of α, there is an optimal tree that minimizes the cost-complexity. To find the optimal tree, a cross-validation approach can be used. The data set is split into V subsets and a tree is constructed from V-1 subsets and validated on the sample that was left out. The procedure is repeated V times producing V trees. The average of the prediction errors across all trees of a certain complexity is used as the final prediction error. The optimal tree is selected as the minimum size tree among those trees whose cross-validation error is within 1 standard deviation of the minimum cross-validation error.

An advantage of CART is its handling of missing data through surrogate variables.  For every split, CART examines the primary splitter as well as other variable splits that minimize impurity.  A surrogate split attempts to mimic the result of the primary split.  Another advantage of CART is its handling of variable interactions.  The binary tree structure shows the effects of variable interactions at the optimal splits. Logistic regression is a special case of CART. However with a large number of parameters it is computationally infeasible to examine every possible subset along with their interactions with logistic regression. CART’s utility is that its algorithm does exactly that and determines the combination of variables that best increase its predictive power.

Sensitivity, specificity, and positive- and negative-predictive values were calculated using entire trees. We also calculated the likelihood ratio of a negative test which is defined as (1-specificity)/ sensitivity, the likelihood ratio of a positive test which is defined as sensitivity/ (1-specificity), and Youden’s index γ which is defined as sensitivity + specificity -1, all of which are summary measures of both sensitivity and specificity [19Hilden J, Glasziou P. Regret graphs, diagnostic uncertainty and Youden's index Stat Med 1996; 15: 969-86.]. Models were compared using Youden’s index where higher values of γ are desirable. Single node trees were assessed using Fisher’s exact test.

Treatment response served as the primary outcome variable used for the analyses. Potential predictors of treatment response were HAM-D individual items, Bech and Maier-Philipps subscales, and total score at baseline, 2 and 7 days, QEEG scores at baseline, 2 and 7 days. Models were compared using Youden’s index γ.