Transcranial magnetic stimulation indices of cortical excitability enhance the prediction of response to pharmacotherapy in late-life depression

Abstract

Background.

Older adults with late-life depression (LLD) often experience incomplete or lack of response to first-line pharmacotherapy. The treatment of LLD could be improved using objective biological measures to predict response. Transcranial magnetic stimulation (TMS) can be used to measure cortical excitability, inhibition, and plasticity, which have been implicated in LLD pathophysiology, and associated with brain stimulation treatment outcomes in younger adults with depression. TMS measures have not yet been investigated as predictors of treatment outcomes in LLD, or pharmacotherapy outcomes in adults of any age with depression.

Methods.

We assessed whether pre-treatment single-pulse and paired-pulse TMS measures, combined with clinical and demographic measures, predict venlafaxine treatment response in 76 outpatients with LLD. We compared the predictive performance of machine learning models including or excluding TMS predictors.

Results.

Two single-pulse TMS measures predicted venlafaxine response: cortical excitability (neuronal membrane excitability), and the variability of cortical excitability (dynamic fluctuations in excitability levels). In cross-validation, models using a combination of these TMS predictors, clinical markers of treatment resistance, and age, classified patients with 73±11% balanced accuracy (average correct classification rate of responders and non-responders; permutation testing, p<0.005); these models significantly outperformed (corrected t-test, p=0.025) models using clinical and demographic predictors alone (60±10% balanced accuracy).

Conclusions.

These preliminary findings suggest that single-pulse TMS measures of cortical excitability may be useful predictors of response to pharmacotherapy in LLD. Future studies are needed to confirm these findings and determine whether combining TMS predictors with other biomarkers further improves the accuracy of predicting LLD treatment outcome.

INTRODUCTION

Whereas objective biological data are commonly used to guide treatment decisions in general medicine, patient-reported outcomes and trial-and-error guide treatment decisions for depression. Over half of older adults with depression do not respond to first-line antidepressant pharmacotherapy(1, 2). To date, demographic and clinical variables have shown only modest utility as predictors of antidepressant outcome(3). Brain-based measures may improve the prediction of treatment outcome beyond what can be accomplished using self-report or rater-assessed clinical data. For example, in large studies of primarily younger and middle-aged adults with depression, the inclusion of resting-state functional magnetic resonance imaging (fMRI) or electroencephalography (EEG) measures of brain function improved the prediction of antidepressant outcomes compared to demographic and clinical measures alone(4–6). However, fewer studies have investigated biological predictors of treatment outcome in older adults with depression.

Aging likely influences the pathophysiology of depression(7, 8), therefore some biological predictors of treatment response are likely unique in late-life depression (LLD). Preliminary studies have examined the relationship between baseline fMRI or EEG measures and antidepressant treatment outcomes in LLD(9–14). For example, Karim et al. found that fMRI indices improved the prediction of pharmacotherapy outcomes as compared to clinical measures in LLD(15). Previous evidence implicating executive function as a predictor of treatment response in LLD(16) might also suggest a role for cortical functioning(17) in the prediction of treatment outcomes. While fMRI and EEG provide valuable indices of brain activity and functional connectivity, these modalities alone cannot index specific forms of inhibitory and excitatory neurotransmission in the cortex.

Complementing fMRI and EEG measures of brain function, single-pulse and paired-pulse transcranial magnetic stimulation (TMS) can be used to measure specific excitatory and inhibitory mechanisms in cortical circuits in a cost-effective way. TMS uses electromagnetic induction to non-invasively stimulate the cortex. Cortical excitability, inhibition, facilitation, and plasticity can be measured by combining TMS of the motor cortex with electromyography measurements of motor output(18). These cortical mechanisms have been shown to change with advancing age(19, 20), and have been implicated in the pathophysiology of depression(21, 22) and LLD(8, 23). For example, TMS studies have found abnormally low cortical inhibition in patients with depression across the lifespan(8, 24–26), and an association between cortical plasticity and executive function in LLD(23). Furthermore, TMS studies in younger and mid-life adults with depression have found that less cortical inhibition(27, 28), greater plasticity(29), and greater excitability(29, 30) at baseline may be associated with better outcomes after brain stimulation treatment. Additionally, one study identified a relationship between cortical inhibition and response to 6 weeks of fluoxetine in sixteen youth with depression(31). However, neurophysiological predictors of pharmacotherapy and brain stimulation treatment likely differ(4, 32), and preliminary findings in children/adolescents cannot be directly extrapolated to adults. The relationship between TMS measures of cortical functioning and pharmacotherapy outcomes has not yet been studied in adults with depression of any age. Moreover, TMS measures have not been evaluated as predictors of response for any treatment avenues in LLD.

Therefore, we undertook a novel pilot study to assess whether pre-treatment single-pulse and paired-pulse TMS measures are neurophysiological predictors of depressive symptom improvement following 12 weeks of antidepressant (venlafaxine) pharmacotherapy in patients with LLD. We explored: (1) which TMS measures may be useful as predictors, and (2) whether combining TMS predictors with demographic and clinical predictors could be used to predict venlafaxine response more effectively than demographic and clinical predictors alone.

METHODS & MATERIALS

Participants

All participants were enrolled in one of two ‘Incomplete Response in Late-Life Depression: Getting to Remission’ (IRL-GRey) clinical trials at the Centre for Addiction and Mental Health (Toronto; ClinicalTrials.gov identifiers: NCT00892047 and NCT02263248)(33), which were designed to address incomplete response to conventional treatment in older adults with mild to severe depression. Baseline measurements were collected before participants initiated ~12 weeks of open-label treatment with venlafaxine (a serotonin-norepinephrine reuptake inhibitor, SNRI). A subset of the older adults with LLD studied here participated in a previously published investigation of longitudinal changes in cortical inhibition and plasticity during venlafaxine treatment(34). Inclusion and exclusion criteria are detailed in the Supplement. In accordance with the Declaration of Helsinki, the study was approved by the CAMH Research Ethics Board and all participants provided written, informed consent.

Venlafaxine Treatment and Clinical Outcomes

All participants took venlafaxine XR (extended-release) for approximately 12 weeks. Venlafaxine was initiated at 37.5 mg/day and was increased to a maximum dosage of 300 mg/day as needed. Low dosages of some sedatives or anxiolytics were allowed as needed, including benzodiazepines (up to the equivalent of lorazepam 2 mg/day), zopiclone (up to 15 mg/day), or trazodone (up to 50 mg/day). Details of the titration schedule and permitted medications have been reported previously(35). For this study, the primary clinical outcome was treatment response, defined as ≥ 50% MADRS score reduction, consistent with the observed bimodal distribution of symptom improvement (Figure 1A). In secondary analyses, treatment outcome was classified based on remission (MADRS score ≤10) or early partial response (>26.5% MADRS score reduction at 2 weeks(35); Supplement).

Figure 1. Depressive symptom changes in responder and non-responder subgroups.

A. Histogram of percentage change in MADRS scores pre- to post-treatment in participants with LLD. Negative values reflect a decrease (i.e., an improvement) in depressive symptoms. The vertical dashed line indicates the threshold used to define response subgroups. B. Individual participant MADRS scores are plotted prior to and immediately following 12 weeks of venlafaxine treatment.

Pre-Treatment Demographic, Clinical, and Cognitive Measures

A set of demographic, clinical, and cognitive measures were considered based on published evidence of their association with pharmacotherapy outcome in LLD or TMS measures. After excluding measures with >25% of data missing or minimal variability across participants, 15 measures were selected: age(36, 37), sex(35), years of education(36), depressive symptom severity(35–37), self-reported sleep loss(36), pessimistic thoughts (e.g. guilt)(36), current depressive episode duration (natural log-transformed)(35–38), age of onset of first lifetime depressive episode(36, 38, 39), antidepressant treatment resistance prior to initiation of venlafaxine(35, 36, 40), anxiety severity(36, 37), medical comorbidity(37, 41, 42), concurrent low-dose sedative/anxiolytic use(43), diagnosis of a comorbid anxiety disorder(37), executive function(37, 44, 45), and delayed memory recall(36). Clinical and cognitive measurements are detailed in the Supplement.

Pre-Treatment Assessment of Cortical Functioning using TMS

TMS measures were collected as described in detail previously(23). Briefly, monophasic TMS pulses were delivered to the left motor cortex during electromyography recordings of the right abductor pollicis brevis muscle. A figure-of-eight TMS coil (Magstim) was positioned with the coil handle approximately 45° laterally from the midline to elicit posterior-anterior current and maximize the motor evoked potential (MEP) amplitude. Electromyography signals were amplified, band pass filtered (2–2.5 kHz) and digitized at 5 kHz. The electromyography signal was continuously monitored to ensure muscle relaxation. We assessed 8 TMS measures of cortical functioning: (1) resting motor threshold (RMT), (2) 1mV stimulator intensity (SI_1mV), (3) coefficient of variation (CV) of MEP amplitudes, (4) cortical silent period, (5) short-interval intracortical inhibition (SICI), (6) intracortical facilitation (ICF), and (7) average and (8) maximum paired associative stimulation (PAS)-induced plasticity. See Table 1 for a summary of TMS paradigms.

Table 1.

Description of the single-pulse and paired-pulse TMS paradigms and their associated cortical mechanisms. The TMS paradigms were performed in the order listed in the table.

TMS Paradigm	Description of Paradigm	Primary Underlying Cortical Mechanism
Measures of Cortical Excitability
(1) Resting motor threshold (RMT)	Minimum stimulator intensity that elicited a ≥50μV MEP in 5 of 10 consecutive trials at rest.	Neuronal membrane excitability
(2) 1mV stimulator intensity (SI1mv)	Suprathreshold stimulator intensity that elicited an average MEP amplitude of approximately 1mV at rest.	Neuronal membrane excitability
(3) Coefficient of variation (CV) of MEP amplitudes	Trial-to-trial CV of the MEP amplitudes elicited by single-pulse TMS (12 trials delivered at SI1mv).	Variability of cortical excitability
Measures of Intracortical Inhibition and Facilitation
(4) Short-interval intracortical inhibition (SICI)	Paired-pulse TMS: Conditioning pulse (80% of RMT) preceded test pulse (SI1mv) by 2ms (12 trials per condition).	GABAA receptor-mediated inhibitory neurotransmission
(5) Intracortical facilitation (ICF)	Paired-pulse TMS: Conditioning pulse (80% of RMT) preceded test pulse (SI1mv) by10ms (12 trials per condition).	NMDA receptor-mediated excitatory neurotransmission
(6) Cortical silent period	Single-pulse TMS (140% of RMT) during tonic muscle contraction (10 trials).	GABAB receptor-mediated inhibitory neurotransmission
Measures of Cortical Plasticity
(7) Average paired associative stimulation (PAS)-induced plasticity	Pre-PAS: Single-pulse TMS (SI1mV) PAS: Right median nerve stimulation preceded TMS by 25ms (180 pairings). Post-PAS: Single-pulse TMS (SI1mv) immediately after, and 15, 30 and 60 mins after PAS (20 trials per timepoint). Average potentiation calculated.	Long-term potentiation-like cortical plasticity
(8) Maximum PAS-induced plasticity	Same PAS procedures. Maximum potentiation calculated.	Long-term potentiation-like cortical plasticity

Cortical Excitability

RMT and SI_1mV were first established as indices of cortical excitability. Following RMT determination using the relative frequency method(46), stimulation was increased to 120% of RMT, and adjusted until the average peak-to-peak MEP amplitude across 15–20 consecutive pulses was approximately 1mV (0.5–1.5 mV; SI_1mV). Both RMT and SI_1mV were established using single pulses of TMS with a 5s inter-trial interval. A lower stimulation intensity indicates greater cortical excitability(47). Additionally, the variability (CV) of MEP amplitudes in response to suprathreshold single-pulse TMS (5s inter-trial interval) was used to index intrinsic fluctuations in the level of excitability. Trial-to-trial variability of cortically evoked MEPs has been attributed to spontaneous fluctuations in the excitability of cortical neurons, and is insensitive to stimulus intensity, spatial constancy of coil placement, and mental alertness(48–51).

Intracortical Inhibition and Facilitation

Cortical inhibition and facilitation were then assessed using SICI and ICF paradigms. The conditioning pulse reliably attenuates (SICI) or potentiates (ICF) the mean MEP amplitude compared to single-pulse TMS(47, 52), which probes inhibitory or excitatory neurotransmission in the cortex, respectively. Cortical inhibition was subsequently assessed using the cortical silent period paradigm, in which the abductor pollicis brevis muscle was contracted at 20% of maximum force using a pinch gauge during TMS. The duration of the ‘silent period’ was calculated from MEP onset to the return of electromyography activity.

Cortical Plasticity

Lastly, long-term potentiation-like cortical plasticity was assessed using the PAS paradigm(53). During PAS, participants focused their attention on the stimulated hand, and kept a running count of the stimulations, which was recorded periodically to assess attention levels. (i) Average and (ii) maximum PAS-induced plasticity were calculated as (i) the average MEP amplitude across all post-PAS timepoints normalized to the mean pre-PAS MEP amplitude, and (ii) the post-PAS timepoint with the highest mean MEP amplitude normalized to the mean pre-PAS MEP amplitude.

Dimensionality Reduction

Single-pulse and paired-pulse TMS measures, i.e., features, have not yet been investigated as predictors of LLD treatment outcome, therefore we used a data-driven method to (i) identify candidate TMS predictors of treatment response, and (ii) further reduce the number of features used for classification, decreasing the likelihood of overfitting due to high dimensionality. Each feature was scaled to achieve a Gaussian distribution with zero mean and unit variance. Missing observations were imputed using the mean of that measure. The distribution of each feature was checked using a histogram and Q-Q plot.

Feature selection was carried out using a random subset of 70% of the data, thus 30% of the data were held out. We used a genetic algorithm(54) to select candidate predictors, implemented using Distributed Evolutionary Algorithms in Python (DEAP)(55). A genetic algorithm is a stochastic search tool based on the principles of natural evolution that has been successfully used for feature selection in a wide range of classification problems in bioinformatics(56–59). Importantly, genetic algorithms can consider the interactions between features during feature selection, and hence serve to evaluate which measures are jointly informative for predicting treatment outcome.

We ran the genetic algorithm 20 times, each time using a random subsample of the data that was used for feature selection. From the 23 initial features (15 demographic and clinical, and 8 TMS features), each run of the genetic algorithm generated feature subsets as potential predictors. We then calculated the frequency with which each feature was included in the top 100 most predictive feature subsets across all 20 genetic algorithm runs. With this approach, the genetic algorithm is used to both identify candidate feature subsets and rank the relative importance of each feature (based on the frequency with which each measure is selected). Similar strategies have been shown to be stable and reproducible for feature selection in small samples in other fields(56). See Supplement and Figure S1 for details of the genetic algorithm.

Classification

We used a support vector machine (SVM) classifier(60) with a Gaussian kernel (Figure S2) to predict treatment response based on the top 4 pre-treatment predictors that were consistently selected by the genetic algorithm. We also evaluated classifiers using the top 2 to 6 predictors (Table S1). An SVM is a supervised machine learning classifier that identifies a hyperplane that optimally differentiates between treatment response subgroups in a multidimensional space (# of features = # of dimensions). SVM classifiers have been successfully used to predict antidepressant treatment outcomes using biological data(6). We also explored the performance of k-nearest neighbours (k-NN), logistic regression, and linear SVM classifiers (Supplement). All classifiers were implemented using the scikit-learn Python package(61, 62).

Monte Carlo cross-validation was performed using 200 random 80-20 train-test splits of the full dataset. In other words, each model was trained on 80% of the data, and evaluated on the held out 20% of the data, which was repeated 200 times with different train-test splits of the data. To evaluate whether TMS data could improve predictive performance compared to clinical and demographic data, cross-validation was performed for classifiers using (i) demographic, clinical, and TMS predictors, and (ii) demographic and clinical predictors only. Five classifier performance metrics were calculated during cross-validation: balanced accuracy, accuracy, sensitivity, specificity, and precision (Table 3).

Table 3.

Mean ± SD cross-validation performance metrics (using the full dataset) are displayed for SVM models including TMS, demographic, and clinical predictors, or only demographic and clinical predictors. Accuracy is the proportion of all predictions that were correct. Balanced accuracy is similar to accuracy, but accounts for imbalance in the class sizes. Balanced accuracy is the average of sensitivity and specificity, which is equivalent to the ROC curve area under the curve here. Sensitivity (also known as true positive rate) is the proportion of responders who were correctly classified by the model. Specificity (also known as true negative rate) is the proportion of non-responders who were correctly classified by the model. Precision (also known as positive predictive value) is the proportion of predicted responders who were actually responders.

Performance Metric	TMS, Demographic & Clinical Data	Demographic & Clinical Data
Accuracy	0.75 ± 0.10	0.61 ± 0.11
Balanced Accuracy	0.73 ± 0.10	0.61 ± 0.11
Sensitivity	0.61 ± 0.18	0.50 ± 0.23
Specificity	0.84 ± 0.12	0.71 ± 0.16
Precision	0.73 ± 0.20	0.53 ± 0.21

Statistics

Non-parametric permutation testing was used to assess the statistical significance of each classifier’s cross-validation results (using the full dataset). Treatment outcome labels were randomly shuffled (permuted) 200 times, and for each permutation, cross-validation was repeated to assess classifier performance using the permuted data. The resulting p-value reflects the proportion of mean cross-validation balanced accuracy scores obtained using permuted data that were greater than the mean cross-validation balanced accuracy score obtained without permutation. To validate that the genetic algorithm-SVM approach used here does not overfit the current dataset, we also performed feature selection and classification after permuting the outcome labels. We assessed SVM performance using the top 4 predictors selected by the genetic algorithm using the randomly permuted dataset. The performance of the model using the randomly permuted dataset should not have significant predictive ability. Additionally, a one-tailed corrected resampled t-test(63, 64) was used to assess whether including TMS predictors significantly improved classifier cross-validation performance, as compared to classifiers including only clinical and demographic predictors.

RESULTS

Participants

TMS measurements were collected prior to venlafaxine treatment in 84 participants. Eight participants did not have follow-up clinical measures to assess outcome, thus the final sample consisted of 76 participants with LLD (49 (64.5%) female; mean ± SD age: 67 ± 7 years). After 12 weeks of treatment, 30 participants were classified as ‘responders’ (mean ± SD MADRS score % reduction: 80.8 ± 16.1) and 46 participants as ‘non-responders’ (MADRS score % reduction: 14.0 ± 24.7). Figure 1 presents the distribution of changes in depressive symptom severity, and Table 2 summarizes the demographic, clinical, and neurophysiological characteristics of response subgroups.

Table 2.

Demographic, clinical, cognitive, and neurophysiological characteristics of participants by response subgroup. Mean ± standard deviation is reported throughout unless stated otherwise. Response subgroup comparisons were performed using an independent two-tailed t-test (continuous variables), Mann-Whitney U test (non-normally distributed continuous variables), or Fisher’s exact test (categorical variables). MADRS = Montgomery-Asberg Depression Rating Scale, % MSO = % of maximum stimulator output, MEP = motor evoked potential, SI_1mV = 1mV stimulator intensity, CV = coefficient of variation, PAS = paired associative stimulation, SICI = short-interval intracortical inhibition, ICF = intracortical facilitation.

	Responders (n = 30)	Non-Responders (n = 46)	Statistics
Demographics
Age (years) [range]	66.2 ± 5.5 [57 – 86]	66.9 ± 7.3 [53 – 92]	t74 = 0.43, p = 0.67
% Female	70.0 (n = 21)	60.9 (n = 28)	p = 0.47
% Caucasian	93.3 (n = 28)	95.7 (n = 44)	p = 0.73
Years of education	14.6 ± 2.2	14.6 ± 3.3	t73 = 0.01, p = 0.99
Clinical Measures
Baseline MADRS score	24.2 ± 4.7	26.7 ± 5.7	t74 = 2.03, p = 0.05
Age of onset (years) [range]	40.2 ± 17.3 [11 – 69]	42.9 ± 18.3 [12 – 91]	t74 = 0.63, p = 0.53
Duration of current depressive episode, median [range]	46 [2 – 2100] weeks	78 [8 – 1248] weeks	t74 = 1.45, p = 0.15
% Non-response to ≥1 prior adequate antidepressant trial	60.0 (n = 18)	82.6 (n = 38)	p = 0.04
% Concurrent anxiolytic / sedative use	20.0 (n = 6)	32.6 (n = 15)	p = 0.30
% Comorbid anxiety disorder	30.0 (n = 9)	34.8 (n = 16)	p = 0.80
Brief Symptom Inventory – Anxiety	1.67 ± 0.90	1.59 ± 0.98	t74 = 0.39, p = 0.69
Medical comorbidity (CIRS-G total score)	6.4 ± 4.0	6.5 ± 3.2	t74 = 0.09, p = 0.92
Reduced sleep (MADRS item 4 score)	2.63 ± 1.47	2.83 ± 1.80	t74 = 0.49, p = 0.63
Pessimistic thoughts (MADRS item 9 score)	2.63 ± 0.96	2.78 ± 1.07	t74 = 0.62, p = 0.54
Executive function score	4.4 ± 1.6	4.4 ± 2.3	t60 = 0.11, p = 0.91
Delayed recall score (% retained)	71.1 ± 29.4	73.6 ± 21.5	t62 = 0.40, p = 0.69
Week 12 venlafaxine dosage (mg/day)	221.3 ± 79.0	266.6 ± 63.8	W = 894, p = 0.01
Neurophysiological Measures
Cortical Excitability
Resting motor threshold	49.1 ± 9.2 % MSO	48.5 ± 9.8 % MSO	t74 = 0.26, p = 0.79
SI1mV	62.7 ± 12.8 % MSO	63.1 ± 16.7 % MSO	t74 = 0.12, p = 0.90
CV of MEP amplitudes	59.0 ± 19.5 %	70.0 ± 24.7 %	t73 = 2.03, p = 0.05
Cortical Inhibition & Facilitation
SICI Ratio	0.58 ±0.38	0.68 ± 0.53	t73 = 0.60, p = 0.55
ICF Ratio	1.83 ± 0.77	1.84 ± 0.87	t73 = 0.25, p = 0.80
Cortical silent period	127.4 ± 34.5 ms	123.2 ± 34.5 ms	t72 = 0.50, p = 0.62
Cortical Plasticity
Average PAS Ratio	1.09 ± 0.44	1.20 ± 0.57	t58 = 0.64, p = 0.52
Maximum PAS Ratio	1.49 ± 0.67	1.64 ± 0.91	t58 = 0.52, p = 0.61

TMS and Clinical Predictors

The top 4 predictors selected by the genetic algorithm included two single-pulse TMS measures, one demographic measure, and one clinical measure: (1) variability of cortical excitability (the CV of MEP amplitudes; 99.6% selection frequency); (2) cortical excitability (the stimulator output required to elicit a mean MEP amplitude of ~1mV; 89.4% selection frequency); (3) age (86.7% selection frequency); and (4) treatment resistance (ATHF score; 88.3% selection frequency). These predictors were consistently selected as top predictors when different initial feature selection data subsets were used. The identified TMS predictors also showed acceptable reliability in the current sample of LLD participants (Supplement). There were no missing data for the predictors, except for one missing observation for the variability of cortical excitability. Moreover, both TMS predictors were found to be modulated by venlafaxine treatment (Supplement).

Prediction of Treatment Outcome Using Pre-Treatment TMS and Clinical Data

The mean SVM balanced accuracy score obtained with the top 4 TMS, clinical, and demographic predictors was significantly greater than the balanced accuracy scores obtained across 200 permutation tests (p < 0.005). See Table 3 for a summary of SVM performance metrics. Additionally, random permutation of the outcome labels prior to feature selection led to an average SVM balanced accuracy score of only 0.53 ± 0.09 (p = 0.21 in permutation tests), suggesting that the genetic algorithm-SVM approach was not overfitting the data. Classifier performance using TMS, clinical and demographic predictors was also significant using k-NN, logistic regression, and linear SVM classifiers (Supplement).

By contrast, when TMS predictors were excluded, and the top 4 demographic, clinical, and cognitive predictors optimized with the genetic algorithm were used to classify treatment outcome (age, education, ATHF score, and executive function), the SVM was only modestly predictive of treatment outcome (p = 0.035 by permutation testing, see Table 3). The inclusion of TMS predictors significantly improved the prediction of treatment response compared to demographic and clinical predictors alone (corrected t-test, t₁₉₉ = 1.96, p = 0.025). See Figure 2 for an illustration of SVM model performance including vs. excluding TMS predictors.

Figure 2. Illustration of model performances with different predictors.

Histograms of balanced accuracy scores from 200 random train-test splits of the full dataset, and the corresponding mean ± SD receiver operating characteristic (ROC) curves. (a) SVM classifiers based on TMS, demographic, and clinical predictors (4 predictors: variability of cortical excitability, cortical excitability, age, and history of treatment resistance). (b) SVM classifiers based on demographic and clinical predictors alone (4 predictors: age, education, history of treatment resistance, and executive function). (c) SVM classifiers based on TMS, demographic, and clinical predictors with randomly shuffled treatment response labels (4 predictors: SICI, reduced sleep (MADRS item 4), medical comorbidity (CIRS-G total score), and sedative/anxiolytic use).

Classifier performance was consistently improved by the inclusion of TMS predictors, as observed across different types of models (Supplement, Figure S3), different model parameters (Figures S2, S3), and different numbers of predictors included in the model (Table S1). A sensitivity analysis excluding participants <60 years of age also showed significant improvement in classification accuracy when TMS predictors were included (Supplement). The identified predictors were specifically predictive of response or remission after 12 weeks of treatment; they were not predictive of early partial response to treatment (Supplement). Exploration of the linear associations between predictors showed no strong correlations between predictors (Supplement, Figure S4).

DISCUSSION

Being able to use biological data to guide treatment decisions (“precision psychiatry”) is a major goal in contemporary psychiatry. To our knowledge, this is the first prospective study to investigate single-pulse and paired-pulse TMS measures as predictors of response to antidepressant pharmacotherapy in adults. We found that pre-treatment TMS measures of cortical excitability may serve as useful predictors of venlafaxine treatment response in older adults with major depression. Combining TMS predictors with demographic and clinical predictors significantly improved the accuracy with which machine learning models predicted treatment response compared to demographic and clinical predictors alone (mean ± SD balanced accuracy: 73 ± 11% vs. 60 ± 10%). This preliminary result demonstrates the potential additive predictive value of TMS measures relative to traditional demographic and clinical predictors. Of the TMS measures we investigated, cortical excitability and the variability of cortical excitability were identified as valuable predictors. These neuronal membrane potential-related TMS predictors were modulated by venlafaxine treatment (Supplement), suggesting that these indices of cortical excitability are examples of neural targets that are both engaged by venlafaxine treatment and predictive of response to venlafaxine. By contrast, TMS measures of cortical inhibition, facilitation, and plasticity were not identified as top predictors; this is consistent with earlier work showing that these GABA and glutamate-related measures are unchanged during venlafaxine treatment(34). For many older adults with depression who do not improve with first-line pharmacotherapy, a clinical decision support tool based on relatively easy-to obtain biomarkers such as these single-pulse TMS measures could substantially improve the efficacy and efficiency of treatment.

We used motor cortex TMS-electromyography as a proxy for general cortical functioning, given the similar cellular organization across cortical regions(65). Previous studies have investigated the relationship between motor cortex TMS measures and brain stimulation treatment outcomes in depression(27, 29, 66). However, motor cortex TMS predictors may be particularly relevant to pharmacotherapy outcomes, as compared to brain stimulation outcomes, given that repetitive TMS targets non-motor cortical regions, whereas antidepressant medications such as SNRIs have a more widespread influence in the cortex(67–69). Compared with more involved whole-brain measurement techniques, single-pulse motor cortex TMS predictors require relatively brief (< 20 mins) measurement and analysis procedures that place little burden on the patient. These single-pulse TMS predictors can be reliably measured by a non-clinician by (i) establishing the TMS intensity that elicits a ~1mV MEP, and (ii) delivering 12 – 30 single TMS pulses at this intensity. Following replication of our preliminary results in future work, these non-invasive neurophysiologic TMS predictors would be relatively cost-effective to obtain in a clinic, as compared to many neuroimaging modalities.

Our finding that neurophysiological data can be used to enhance the prediction of treatment outcome is in line with recent investigations of other brain-based predictors of pharmacotherapy outcome. In particular, machine learning models based on resting-state EEG and fMRI predictors significantly outperformed models based on clinical predictors alone(4, 5, 15). Moreover, a robust EEG predictor of sertraline treatment outcome (which differentially predicted outcome with brain stimulation treatment, and did not predict outcome with placebo) showed a moderately strong association with cortical excitability as assessed using single-pulse TMS(4), supporting the current findings that single-pulse TMS measures may be useful as predictors of antidepressant treatment outcome. Altogether, current evidence suggests that measures of brain function collected prior to treatment initiation may convey important information about a patient’s likelihood to improve with pharmacotherapy. Future studies should consider combining brain-based predictors (e.g., TMS, EEG and fMRI) and established clinical predictors to develop multi-modal signatures of brain function that can more accurately guide treatment decisions.

The SI_1mV TMS predictor is thought to reflect neuronal membrane excitability at suprathreshold stimulation intensities(47, 70). Like RMT, SI_1mV is mediated by voltage-gated Na⁺ channels(70) and K⁺ channels(71), but not by GABA(70). Congruent with our findings, previous studies have identified associations between similar metrics of excitability (RMT and MEP amplitude) and brain stimulation treatment outcomes in depression(27, 30), particularly when combined with other demographic, clinical, and neurophysiological predictors(29). It could be speculated that cortical excitability and the variability of excitability in particular were predictive of venlafaxine response because of their relationship with venlafaxine pharmacology. The observed modulation of cortical excitability by venlafaxine is in line with evidence from rodents that both norepinephrine and serotonin can influence resting membrane potential in the cortex(72–74). Of note however, while these TMS measures were identified as useful predictors of treatment outcome, they may not play a direct role in antidepressant outcomes; other biological factors, such as the presence of amyloid(75), could mediate the observed relationship.

TMS measures were predictive of treatment outcome when combined with previously identified demographic and clinical predictors. Age and a history of antidepressant treatment resistance were the two top demographic and clinical predictors of treatment response, in line with prior findings(35, 36, 76). Advancing age may interact with the identified TMS predictors(20), therefore further work is needed to clarify whether the TMS predictors we identified generalize across the lifespan or whether they apply specifically to LLD. Other previously identified demographic and clinical predictors, including years of education, baseline pessimistic thoughts (e.g. guilt), and executive function were also identified as predictors of treatment response in some of our models. However, the inclusion of these predictors did not improve model performance (Table S1). We also previously found that measures of executive functions related to LLD treatment outcomes (e.g. cognitive flexibility) were not strongly associated with the TMS measures studied here(23). On their own, demographic and clinical predictors were only modestly predictive of treatment outcome, in line with previous reports(3, 4, 6).

Some limitations of the current study should be considered. First, as the first investigation of TMS predictors of LLD treatment outcome, this novel pilot study was intended to explore the value of easy-to-obtain motor cortex TMS measures as predictors of pharmacotherapy outcome, rather than to develop a definitive predictive model. This preliminary work identified candidate neurophysiological predictors for future studies; replication of these results in a larger, independent dataset is needed. Second, the modest sample size, data-driven feature selection approach, and overlap in the data used for feature selection and classification likely led to some degree of overfitting to the current dataset. However, we took a number of steps to mitigate the risk of overfitting: (a) demonstrating consistent findings with multiple parsimonious classifiers, (b) using only a subset of the full sample for feature selection, (c) regularizing feature selection by choosing candidate predictors based on their frequency of selection across multiple genetic algorithm runs, (d) training and evaluating all models with a minimal number of features, and (e) using Monte Carlo cross-validation to assess model performance. We also evaluated the degree of overfitting by running the current pipeline with randomized treatment outcome labels; as expected, the resulting model performed at chance levels. Third, the specificity of the identified predictors to venlafaxine cannot be assessed in the absence of another antidepressant or treatment modality. Thus, the predictors we identified could be associated with treatment response in general and non-specific mechanisms of response. Fourth, although low-dose benzodiazepine use was taken into account as a potential predictor during feature selection, the possible interaction between low-dose benzodiazepine use in some patients and the TMS measures cannot be ruled out. Fifth, our findings were specific to the TMS paradigms investigated here, and future studies of predictive biomarkers should consider additional TMS measures that probe overlapping but distinct components of cortical functioning, such as long-interval intracortical inhibition. Similarly, our findings are specific to motor cortex measurements. Sixth, age-related cortical atrophy could have influenced the TMS measures. Future work expanding beyond the motor cortex should account for cortical atrophy using structural MRI. Finally, limited ethnic and racial diversity in our participants precluded the assessment of ethnicity or race as predictors. This limitation highlights the challenges with recruiting diverse samples in clinical research, which needs to be addressed in future work.

In conclusion, combining pre-treatment TMS predictors with established demographic and clinical predictors may improve the prediction of treatment outcome in LLD as compared to demographic and clinical predictors alone. In particular, cortical excitability and the variability of cortical excitability may be useful as single-pulse TMS predictors. This is encouraging given their relative simplicity from a clinical implementation standpoint. These preliminary results warrant a larger study wherein TMS measures, ideally in combination with other promising biomarkers, are applied to the prediction of LLD treatment outcomes.