Changes in Functional Connectivity Predict Outcome of Repetitive Transcranial Magnetic Stimulation Treatment of Major Depressive Disorder

Abstract

Repetitive transcranial magnetic stimulation (rTMS) treatment of major depressive disorder (MDD) is associated with changes in brain functional connectivity (FC). These changes may be related to the mechanism of action of rTMS and explain the variability in clinical outcome. We examined changes in electroencephalographic FC during the first rTMS treatment in 109 subjects treated with 10 Hz stimulation to left dorsolateral prefrontal cortex. All subjects subsequently received 30 treatments and clinical response was defined as ≥40% improvement in the inventory of depressive symptomatology-30 SR score at treatment 30. Connectivity change was assessed with coherence, envelope correlation, and a novel measure, alpha spectral correlation (αSC). Machine learning was used to develop predictive models of outcome for each connectivity measure, which were compared with prediction based upon early clinical improvement. Significant connectivity changes were associated with clinical outcome (P < 0.001). Machine learning models based on αSC yielded the most accurate prediction (area under the curve, AUC = 0.83), and performance improved when combined with early clinical improvement measures (AUC = 0.91). The initial rTMS treatment session produced robust changes in FC, which were significant predictors of clinical outcome of a full course of treatment for MDD.

Keywords: depression; electroencephalogram (EEG); functional connectivity; machine learning; repetitive transcranial magnetic stimulation (rTMS).

Figure 1.

Illustration of the analysis pipeline. (A) Locations of seed electrode locations for connectivity analyses, with the nine left prefrontal seeds shown in yellow (Fpz, Fp1, AF3, AF7, Fz, F1, F3, F5, F7) and the seven right prefrontal seeds shown in blue (Fp2, AF4, AF6, F2, F4, F6, F8). All electrode locations (N = 61) were utilized as connectivity nodes. Connectivity pairings of selected left and right seeds with all other electrodes yielded a total of 783 connectivity features per subject. (B) Examples of the three connectivity metrics of coherence, envelope correlation, and αSC for one responder (upper row) and one non-responder (bottom row). Magnitude squared coherence takes amplitude and phase information into account. Envelope correlation is an amplitude-amplitude coupling measure. αSC is the similarity of the spectral waveform between two channels, which was found to be more similar for responders than non-responders. (C) Elastic net regularization was used to build a model for each neurophysiologic measure that distinguished between responders and non-responders. Models were subjected to training and testing cross-validation repeated 100 times to minimize overfitting and spurious classification due to random sampling effects. For each repetition, the full sample (N = 109) was divided into 70% training and 30% testing sets. For each run, training cross-validation consisted of splitting the training set 10-fold, training the model on 9/10-folds and using the 10th-fold to make predictions. This procedure was repeated 10 times, so that each fold served as both a training and testing set. Testing validation consisted of applying the model obtained from training to make predictions about the test data set. (D) For each neurophysiologic measure, the connections that most reliably predicted outcome in validation were identified. These consisted of the 10 features most consistently selected in all training models across 100 repetitions compared among the predictors. These features then were plotted topographically (right panel).

Figure 2.

Treatment-emergent connectivity differences. Top row: histograms of mean values for all features for responders (red) and non-responders (blue) before the first rTMS treatment. None of the features differed significantly between groups. Bottom row shows differences in treatment-induced connectivity changes, where group separation is more apparent. Coherence and envelope correlation (left and middle column) showed smaller group separation than αSC (right column) (respective effect sizes Cohen’s d = 0.84; 1.22; 1.52). For coherence and envelope correlation, non-responders had on average greater connectivity changes than responders. Inversely, change in αSC was on average greater for responders.

Figure 3.

AUCs and feature topography for all EN models. (A) AUCs for all three predictors for training (green) and testing performance (gray) averaged over 100 repetitions. Coherence (training and testing): 82.5 and 52.7; Envelope correlation: 82 and 57.8; αSC: 83.2 and 66.1, with highest AUCs for αSC. (B) The corresponding top 10 selected features per predictor across all 100 trained models. EN models for coherence and envelope correlation showed a diffuse coupling pattern, while αSC showed a more focal connectivity. (C) AUCs for a logistic regression model using solely the early clinical response to rTMS treatment: 74.4 and 72.7. (D) AUCs for a logistic regression model combining the top αSC features and the early clinical response: 85.4 and 77.8, which represented the overall best predictive model.