1 INTRODUCTION

Major depressive disorder (MDD) is a psychiatric condition concomitant of dysregulated mood and one of the leading causes of disability worldwide, affecting approximately 17% of the population (Kessler et al., 2003; Picco et al., 2017; Seney et al., 2018). Prevalence of MDD has been estimated to be twice as high in women, and females typically have more severe and a greater risk of recurrent episodes than men (Perugi et al., 1990). There is evidence regarding the role of sex and reproductive hormones in the pathophysiology of depression, as well as differences in the brain at the structural, cellular, and network levels that can potentially contribute to the susceptibility of sex to MDD (Tripp et al., 2012). Sex hormones have been shown to modulate functional neural networks, such as modulation of the central glutaminergic and serotonergic system by ovarian steroids. Considering the evidence regarding sex-dependent physiological and neural dynamics, it is possible to form hypotheses regarding potential differences between men and women, in terms of response to treatment across different conditions and treatments (Huang et al., 2008; Rubinow & Schmidt, 2019). Investigation of the potential differences between men and women has been the focus of the current research, with the condition and treatment of focus being MDD and repetitive transcranial magnetic stimulation (rTMS), respectively.

Repetitive TMS is a noninvasive, nonpharmacological brain stimulation treatment that takes advantage of a pulsed magnetic field near the scalp to induce neuronal depolarization in a targeted brain region. TMS for MDD starts with motor threshold (MT) determination, which calibrates the stimulator to an individual's cortical excitability. During MT determination, a clinician delivers single pulse TMS to the motor cortex, and records the amount of stimulator output necessary to induce movement in the contralateral hand in 50% of delivered pulses. Following calibration, a course of TMS is delivered to the prefrontal cortex at 120% of MT on a daily basis for up to 30 (or more) sessions, often followed by a taper phase (Philip et al., 2018). TMS treatment protocols have shown preliminary clinical efficacy for numerous neuropsychiatric and behavioral disorders; randomized clinical trials recently led to the US FDA approvals for TMS to treat OCD and for smoking cessation (Greenberg et al., 2021). The dorsolateral prefrontal cortex (DLPFC) of the brain is shown to be correlated with the cognitive-control and affective network thus playing a significant role in mood regulation (Downar & Daskalakis, 2013). Repetitive transcranial magnetic stimulation of the left DLPFC has been shown to induce an inverse correlation between resting connectivity of the DLPFC and the default mode network (DMN), a dysfunctional circuit characteristic of MDD (Williams et al., 2021). Superior clinical outcomes were associated with targets exhibiting the greatest DLPFC-to-sgACC negative connectivity, or anticorrelation, suggesting the antidepressant mechanism of TMS (Philip et al., 2018). TMS uses a pulsed magnetic field to induce neuronal depolarization in a targeted brain region, either “high” (≥5 Hz) or “low” (≤1 Hz), considered excitatory and inhibitory, respectively (Philip et al., 2018).

Sexual dimorphism has been observed in MDD patients not only in lifetime risk, clinical presentation, but also in their response to pharmacotherapy, perhaps attributed to sex-specific biological differences, including hormone levels and metabolic enzymes, which all together lead to different profiles of pharmacokinetics and pharmacodynamics (Hu et al., 2021). Interestingly, the neural networks that include the DLPFC have been previously shown to be affected by ovarian sex hormones that may suggest the possibility of sex-dependent rTMS treatment outcome (Huang et al., 2008; Rogers & Dhaher, 2017). Based on our recent work, it is apparent that sex, especially in the context of depression and rTMS treatment outcome, has not received appropriate research attention (Phillips et al., 2020). Therefore, this study was designed to investigate the sex differences in response to neuromodulation in major depressive disorder using state of the art deep learning methods applied on electroencephalography (EEG) data.

Among the numerous modalities of neuroimaging, EEG is by far the least expensive and complicated which makes it a popular neuroimaging method especially in clinical settings (Zhang et al., 2021b). With its high temporal resolution and noninvasive characteristics, the EEG measurements of the sum of synaptic potentials, changes in the amplitude, and latency of cortical reactivity can be reflected in the corticocortical interactions on a millisecond time-scale (de Aguiar Neto & Rosa, 2019; de la Salle et al., 2020). EEG can measure cordance, a computation of regional brain activity, that has a strong correlation with cerebral brain perfusion, and has been used in the past before and after TMS to assess cortical inhibition and excitation, connectivity, and pharmacology of TMS in patients with MDD (Leuchter et al., 2009; Tremblay et al., 2019). With the use of EEG data, one can compute its recordings of band power, alpha asymmetry, signal-based features, network-based features, or evoked potentials to characterize the brains’ functional response (Bares et al., 2015).

EEG, among many other types of data, can be used after artifacts are removed, in deep learning models for the purposes of feature extraction and thus classification of a variable of interest (Abbasi & Goldenholz, 2019; Ay et al., 2019; Craik et al., 2019). Working with the raw (uncleaned) data is not preferable in the modeling via deep learning. Because the learning process interferes with the noise, it may cause overfitting or biased learning. “Raw” term is generally stated as clean data after preprocessing (noise filtered). Deep learning automatically extracts valuable features from the preprocessed data. The artifacts originated from various physiological or nonphysiological sources such as eye movements, blinks, and cardiac or muscle activity, blood pressure, and magnetic field of electronic devices, mobiles wave, power-line, while continuous EEG recordings are collected. The artifact-corrected EEG can be estimated by subtracting the pure artifact activity, which is calculated as the product of modeled artifact patterns from the original artifact-contaminated EEG.

Convolutional neural networks (CNN), as a popular deep learning model, has been used for biosignal classification in the past (Chen et al., 2020; Nguyen et al., 2020; Ouhame et al., 2021; Zhang et al., 2021a). LSTM was developed for overcoming the challenge to address long-term information preservation and short-term input skipping in latent variable models. LSTM inherits a memory cell that has the same shape as a hidden state. The memory cell is controlled by several gates (i.e., input, output, forget). These gates represent a dedicated mechanism, which can decide when to remember and when to ignore inputs in the hidden state. The combination of CNN and LSTM is designed for sequence prediction problems with spatial inputs, like images or videos. The combination means that CNN-LSTM architecture involves using CNN layers for feature extraction on input data combined with LSTMs to support sequence prediction. Therefore, there is no need for any feature extraction process before training the deep learning architecture (Abdelhameed et al., 2018; Courtney & Sreenivas, 2020; Nagabushanam et al., 2020). Improved and recurrent neural networks (RNN) both perform optimally when classifying EEG data, but due to size and type of data set, LSTM has been used for optimal results here. Offering a high predictive value, machine learning approaches have been widely used in psychiatry research (Yao et al., 2020). Deep learning algorithms for the purposes of classifying psychiatric illnesses and predicting treatment outcomes have yielded successful results in some recent studies (Cosmo et al., 2021; de Bardeci et al., 2021; Etkin, 2019; Zhang et al., 2021a). Recently, taking advantage of signal processing methods in the context of mood disorders has been extensively studied, however, implementing them is challenging due to disorder heterogeneity and limitations inherently associated with machine learning in classification methods. Despite several limitations, we aimed to demonstrate that using a combined CNN-LSTM architecture for advanced EEG signal processing may be a useful approach to distinguish the differential responses of male and female MDD patients to rTMS treatment. The relatively high specificities associated with multivariate autoregressive models and deep learning models indicate that these computational frameworks may provide a basis for an adjunctive therapy.

EEG data before and after rTMS treatment in a population diagnosed with MDD were used in the current study and five EEG-based deep learning models were created to classify male and female subjects.

2 METHODS

2.3 Guided selection of deep learning algorithm

Long short-term memory (LSTM) is derived from the recurrent neural network (RNN), which consists of recurrent structures that locally feed firing strength memories. Thus, RNNs are suitable for time-series based feature learning tasks, but suffer from gradient vanishing or exploding problems. For this reason, the LSTM network is utilized. In this network, membership functions join the nodes depending on the input variable, and spatial and temporal firing are sustained to assign one-dimensional membership functions. While learning parameters minimizes the error cost function, the learning structures decide when to generate a rule and activate it with firing strengths exceeding the threshold (Chen et al., 2020; Nguyen et al., 2020; Ouhame et al., 2021). The internal structure of the LSTM algorithm is demonstrated in Figure 1.

The sole use of the LSTM network is not feasible due to considerable memory consumption and computational burden that comes along with it. To overcome this problem, a 1D-CNN should be added to learn local features on each sensor dimension. 1D convolutional operation on each sensor dimension for the rth filter can be expressed in the form of Equation (1):

$$\begin{equation}C_l^r = {f^r}\left( {x_{l:l + s - 1}^i*{w^r} + {b^r}} \right),l \in \left\{ {1,2, \ldots ,L - s + 1} \right\} \ldots .,\end{equation}$$()

where s is the filter size, f(.) is the activation function, w is the weight vector, and b is the bias. Output of the 1D convolutional operation can be stated as $$C \in {R^{( {L - s + 1} ) \times 4}}$$. To obtain a more compact form, a pooling operation can be applied to the output of the 1D convolutional operation. The max-pooling operation takes the maximum value over that of consecutive features from one sensor dimension. The maximum value is determined by Equation (2):

$$\begin{equation}h_k^r = \max \left( {C_{\left( k \right)d + 1}^r, \ldots ,C_{\left( {k + 1} \right)d}^r} \right),k \in \left\{ {1,2, \ldots ,\left[ {\left( {L - s + 1} \right)/d} \right]} \right\}..,\end{equation}$$()

where d is the pooling size and r is the number of filters in 1D convolutional operation.

The feature matrix $$h \in {R^{[ {( {L - s + 1} )/d} ] \times r}}$$ is obtained by performing the 1D convolutional and max-pooling operations. Since the convolution window slides iteratively from the beginning to the end of the raw signal at each sensor dimension, the first dimension of the feature matrix holds the temporal relationships and the second dimension highlights the high-level representation explored through the CNN at each sequential step. Outputs of the CNN are high-level features having temporal characteristics. A flattening layer is used between the convolutional layers and the LSTM layer to reduce the feature maps to a single one-dimensional vector. Then, the local features are fed into the LSTM to solve the issue of temporal dependency. LSTM provides a solution by saving long-term memory through memory units that can update the previous hidden state. Output values from the previous CNN layer are fed to gate units. In general, an LSTM network is constituted from a forgetting gate, input gate, input candidate gate, and output gate. The forget gate produces a vector with values between zero and one, which will be then multiplied to the cell state of the former time step for erasing values that were not required and keeping those that are essential for the prediction. The input gate and the input candidate gate work together to render the new cell state, which will be renewed at the next time step. A combined CNN-LSTM network automatically learns representative features with a high sampling frequency (Abdelhameed et al., 2018; Courtney & Sreenivas, 2020; Nagabushanam et al., 2020; Zhang et al., 2021b).

2.4 CNN-LSTM feature extraction and classification

The algorithm implemented here was based on the publicly available Google Colab, which is a free cloud service that allows AI developers to apply DL-based algorithms. Our graphic processor unit (GPU) model was a Tesla K80 that supports all Python codes and DL libraries. In our deep learning models, multiple one-dimensional CNN layers were used to extract features and reduce the length of EEG segments. After that, the LSTM layers were applied to detect sequential relationships to increase the training speed and classification accuracy. While CNN is used to attenuate noise and consider the correlation among multivariable signals associated with each electrode, the LSTM model extracts temporal information and maps time series into separable domains to realize the classification. With the help of this combined model, both the spatial and temporal features can be revealed from the multivariate nonstationary time-series input data. The block diagram of the proposed model is given in Figure 2.

3 RESULTS

In this study, five different classification models were built. These models include pre-/post-TMS female (model 1), pre-/post-TMS male (model 2), pre-TMS female responder versus pre-TMS female nonresponders (model 3), pre-TMS male responder versus pre-TMS male nonresponder (model 4), pre-TMS responder versus nonresponder for all sexes (model 5). The results of the training process and accuracy plots for the five models that have been implemented in this study are given in Figure 3.

A confusion matrix is a table that is often used to describe the performance of a classifier and visualize the correctly classified labels for each model on a set of test data for which the true values are known. This table layout allows for visualization of the performance of an algorithm and the name stems from the fact that it makes it easy to see whether the system is confusing two classes or not. Each row of the matrix represents the instances in a predicted class, while each column represents the instances in an actual class thus the results were aggregated from the 10 testing results of the 10-k cross-validated models.

A confusion matrix is generated for each model implemented here, as given in Figure 4.

Seen in the confusion matrix for each model, the number of correctly classified labels is high, and its performance is also reflected in its high sensitivity, specificity, overall accuracy, Cohen's kappa value, and AUC values of the models in Table 4.

The results for each model are promising for employment in a clinical context; however, a well-defined graphical user interface (GUI) needs to be developed to make the process more user-friendly in addition to further validations with more data for better approximation and generalization. Each of the five models performed with high sensitivity, specificity, and overall accuracy thus resulting in a high-classification performance. Cohen's kappa is a robust statistic useful for the purposes of either interrater or intrarater reliability testing. Similar correlation coefficients range from −1 to +1, where 0 represents the degree of agreement that can be expected from random chance, and 1 represents perfect agreement between raters. The kappa values for the proposed models all have high areas under the ROC curve.

4 DISCUSSION

Although all the five implemented models in the current study were able to classify the pre- and post-rTMS responses with an accuracy of above 91%, model 4 performed significantly better in classifying responsivity in pre-rTMS data in males. However, this does not necessarily suggest that males respond better to rTMS, but only highlights the fact that model 4 outperformed the others in determining sex differences in response to rTMS. Several factors may have contributed to this result, one of them being that out of 25 participants of each sex, there were 14 female while only 7 male nonresponders. This suggests that the model may not have been able to classify the females with the same accuracy as the males. The small sample size—as small as 7 in the male responder group—poses an additional limitation on the significance of the results reported. Since in this study we did not apply statistical methods, it is not correct to express the findings in terms of statistical significance. Instead we used other metrics specific to machine learning algorithms such as the confusion matrix of the feature set, area under ROC curve, and kappa values which are listed in Figure 4 and Table 4.

According to a recent review of sex differences in response to neuromodulation treatments Phillips et al. (2020) have noted that differences in cortical thickness may require different protocol parameters such as higher intensities to achieve comparable results in females. Considering that the treatment in the current work included low- and high-frequency stimulation of right and left DLPFC, respectively, further investigation of the implications of other rTMS parameters regarding treatment response, also as a function of sex, is required to realize a more personalized treatment approach. It must also be noted that the study is done on an Iranian sample of depressed patients; therefore, cultural factors of MDD should have been accounted for to be able to make more firm conclusions based on the reported results. It has been noted in a recent epidemiological review that the prevalence of MDD is twice as high for women in Iran and some neighboring countries (Gharraee et al., 2019) thus the cultural and geographical context in relation to sex may have affected the result tendencies and thus must be considered in future investigations to understand sex differences in treatment response to rTMS.

It is worth noting that model 5 was able to predict response to rTMS with an accuracy of 94% regardless of sex, using the baseline's pre-rTMS EEG. Thus, the results of the current investigation provides that it is feasible to make preliminary predictions about the likelihood of the efficacy of rTMS for MDD patients, given all other variables are standardized.

5 CONCLUSIONS

To the best of our knowledge, this is the first investigation using deep learning methods to specifically investigate sex differences in response to rTMS based on EEG in the context of MDD. The proposed method is robust and automated due to the representation power of time-invariant features from raw EEG signals (Liao et al., 2021). In this methodology, intraslice features of brain images are processed and extracted and can be adopted for other physiological data and imaging modalities such as fMRI, CT, or PET (Cheng & Liu, 2017). The results verify that the proposed method could achieve comparable performance with other state-of-the-art methods to identify sex differences in response to rTMS in MDD patients. Therefore, it is safe to believe that our developed model can also be used for the purposes of accurate diagnosis of other neurological and psychiatric pathologies as well (Ay et al., 2019; Saeedi et al., 2021).

While advances in neuroimaging techniques have rapidly improved over the past few decades, there is still a lack of acceptable precision for the diagnosis and treatment of mental illness (Etkin, 2019). The paradigm shift to a personalized approach toward treatment in recent years has led to the use of techniques such as deep learning among other computational methods applied on neurobiological data, to enhance diagnosis accuracy and validity (de Bardeci et al., 2021). The use of neural networks may lead to a more effective and efficient way to identify previously unrecognized patterns in EEG traces by reducing the need to extract diagnostic information from highly complex EEG time-series and shifting to a more automated, objective, and reliable approach. Physician's management of psychiatric illness can thus be facilitated with the help of computational techniques in psychiatry (de Bardeci et al., 2021).

The preliminary results of the models can have extensive implications in providing precise treatment parameters or target brain regions for innovative neuromodulation treatments especially for vulnerable populations such as females with MDD. There is definitely the need for further studies or meta-analyses of studies on sex as well as cultural differences in diagnosis of mental illness and response to specific treatments. Considering the relative ease of use and accessibility of EEG equipment, it is a very promising candidate for investigations of predictors of response to treatment and diagnostic markers. With further replicable deep learning investigations on EEG data across diverse populations, mental health care providers can will be able to provide more efficacious treatment protocols, personalized to a patient's sex, medication profile, or transdiagnostic symptoms, among many other potential variables each of which requiring more research efforts and attention.