1. Diagnostics of Depressive Disorders
Depressive symptoms can have many etiologies, and their potential neurophysiological mechanisms have not been well understood yet. The current consensus is to consider depressive states as a consequence of poorly differentiated functional, neurotransmitter, and metabolic changes in the brain
[64][1]. Since there is no specific localized pathological process in the nervous tissue in depressive disorders, neuroimaging methods for the diagnosis of depression have been rarely used in practice. Biomarker studies of blood and cerebrospinal fluid have been also rarely used. Diagnosis of major depressive disorder, minor depressive disorder, or bipolar depression is usually based on the results of clinical interviews and standardized questionnaires. The potential use of brain electrical activity data to diagnose depression and predict therapeutic responses to antidepressants is currently the subject of active research.
1.1. Use of Standardized Questionnaires
In addition to the clinical interview, standardized questionnaires are used to assess the severity of depressive symptoms. In Russia, among others, the following are used for this: the Beck Depression Inventory (BDI)
[66][2], the Hamilton Scale (HAM-D)
[67][3], and the Hospital Anxiety and Depression Scale (HADS)
[68][4]. The diagnosis of a depressive disorder is not made solely based on questionnaires. Their results serve as a guide for a psychiatrist in determining the severity of symptoms and finalizing the diagnosis. The final diagnosis is based on the combination of identified symptoms that fit the diagnostic criteria of the International Classification of Diseases. However, the periodic completion of questionnaires by the patient can be used as a relatively objective way to track the dynamics of depressive symptoms to adjust the course of treatment.
1.2. Analysis of the Functional Activity of the Brain (EEG)
Methods for analyzing brain activity can be used both to diagnose depressive states (in particular, to assess the severity of depression) and to predict the response to pharmacological treatment. For example, a decrease in interhemispheric coherence of the frontal leads, observed regardless of the etiology of the depressive state, has been identified as a potential predictor of resistance to pharmacotherapy
[69][5]. Identification of predictors of drug resistance and response to drugs is one of the directions of research on the use of EEG methods in the diagnosis of depression
[70][6].
1.3. Application of Machine Learning
Attempts have been made to use neuroimaging data (MRI, functional MRI, and diffusion MRI)
[71][7] and brain electrical activity (EEG) data
[72][8] to diagnose depressive disorders. The recorded data were analyzed with classical machine learning algorithms (e.g., the support vector machine) and deep learning methods
[73][9].
The main types of data sources and methods used in the assessment of various mental disorders, including depression, were described in the review by Chung and Teo
[74][10]. The authors analyzed several works on the assessment of anxiety and depression based on written texts, voice recordings, data from MRI scans, survey methods, and their combinations. The
presea
perrch noted that different algorithms were better suited for different types of data (they examined gradient boosting, random forest, and neural networks), but the average accuracy of the best-performing algorithms did not exceed 80%. The authors attribute the low results both to the small amount of data and to incomplete or imbalanced training datasets. Chen et al.
[75][11] also noted the need to use additional data channels, including voice, activity, sleep, questionnaire, and instrumental methods (e.g., MRI) to potentially improve the accuracy of diagnosis. Additional data channels can also represent different aspects of everyday activity to add behavioral patterns into the data. The authors emphasized the promise of multimodal datasets, since various mental disorders are often accompanied by external signs progressing with time (
Table 1).
Table 1. Multimodal data features and their uses in mental health [75]. Multimodal data features and their uses in mental health [11].
The use of multimodal data potentially makes it possible not only to diagnose a wide range of mental disorders, but also to monitor the dynamics of the development of the disease, since the observed signs become more pronounced at later stages. However, the collection of multimodal data is often associated with a number of difficulties, in particular with the need to observe the patient for a long time, as well as to confirm the observed signs by cross-checking on historical data performed by an experienced doctor. Big data collected from mobile devices reflect the patient’s social interactions and fine motor skills features such as taps duration, typing speed and rhythm, reaction speed, etc. It is worth noting that in order to automatically collect such an amount of data, the mobile application installed on the patient’s device should have access to the correspondence and system software, which can become a security issue. However, such solutions can be easily implemented on the architecture of social interaction platforms such as social networks. The authors proposed the implementation of platform solutions for big data as the main processing unit. The units then are used as a basis for more complex models. In addition, the authors proposed a machine learning model that can be used to evaluate the effectiveness of treatment and to monitor the patient’s condition using biofeedback.
One of the features of mental disorders, when compared to other disorders, is that their assessment relies heavily on the subjective experience of a patient subjectively described by him or her. The quality of the assessment is highly dependent on the clinical experience of the medical professionals conducting the assessment. To systematize the data collection process of mental illness clinical signs, a general methodology for filling the database is needed. The same methodology, on the other hand, forces doctors to use standardized data collection forms, which positively affects the entire dataset. The fact is that the different experiences of doctors and the peculiarities of the method of presenting the picture of the disease cause certain markup anomalies in the database
[76][12]. Often, this feature does not allow one to reasonably assert that the sample data are not biased. In addition, the amount of relevant data differs from channel to channel, which often requires the deployment of different cleaning and pre-processing algorithms.
In their review of machine learning methods for neuroimaging, Quaak et al.
[77][13] pointed out the need for a more conservative approach to testing the quality of trained models, since their results on the test set often do not reflect the model’s real performance. In addition, the authors noted the popularity of the EEG as a source of data for the diagnosis of mental disorders. MRI/EEG databases for diagnostics of depression, which are used to train machine learning algorithms, can be recorded in different modes: when a patient is performing certain tasks (game), responding to external stimuli (video, images, and music), and in a resting state. The most common approach is to receive data from a person in a resting state (eyes open or closed).
The information about the patient’s depression is often considered protected private information, which leads to very few datasets being openly available. The information on openly available datasets for diagnostics of depression by EEG is presented in Table 2.
Table 2.
Available open datasets for diagnostics of depression by EEG.
The common problem with openly available datasets is that they often include a limited number of participants (the average sample size of patients with depressive disorder is around 30 people). This problem calls into question the diagnostic value of studies conducted using the datasets. This problem can be mitigated either by combining existing datasets
[82][18] or by the collection of more data. The latter option requires substantial resources.
Convolutional neural networks are the most common types of neural networks for classifying the presence or absence of a depressive disorder. They rely on a two-dimensional or one-dimensional convolution operation, or hybrid models of convolutional and recurrent neural networks
[83][19].
The input of the neural network is either the raw EEG signal (after pre-processing in the form of filtering and denoising) or its converted version. Given the existence of well-known models of neural networks used for image classification, the transformation often consists of the formation of an “image” from the EEG signal, which is then fed to the input of a two-dimensional convolutional network. To form an “image”, the power values of brain rhythms can be used, which are spatially projected onto a plane, following the location of the electrodes
[84][20].
Since deep learning algorithms are demanding towards an amount of training data, and the existing samples of training data are often of limited size, the augmentation method is applied. Augmentation is a method of artificially increasing the amount of data used for training
[85][21]. The use of data augmentation in training deep neural networks can reduce the effect of overfitting and improve accuracy and stability. EEG data augmentation uses sliding window data sampling, data from generative models, noise addition (generally Gaussian noise is used), sampling, segment recombination, and Fourier transform
[86][22]. The architectures of neural networks and the achieved accuracy of depression detection according to EEG data are presented in
Table 3.
Table 3.
EEG classification accuracy for depressive disorder.
It should be noted that the claimed high accuracy of depression classification, exceeding 90%, may be associated with testing models on a limited set of initial data and possible incorrect partitioning of training data by patients, which might have led to the implicit leakage of training data into the test subset. Leakage of this kind leads to overestimated accuracy and reduces the generalizability of the model due to the effect of overfitting
[99][35].
The application of methods for automated assessment of depressive disorder severity involves the selection of relevant features of the received signal and the development of complex metrics of the depressive state based on them. Isolation of non-linear spectral characteristics of EEG signals in combination together with the support vector machine methods demonstrated high classification accuracy in several studies
[100][36]. The combination of linear discriminant analysis and genetic algorithms also demonstrated high discrimination performance when classifying patients with the depressive disorder
[101][37]. This line of research can be developed further by incorporating new diagnostic categories as comparison groups and the development of classifiers with higher ecological validity.
The development and validation of composite diagnostic indices, as well as ensembles of algorithms for solving specific diagnostic problems, are most often carried out on small samples in the absence of independent external validation. This creates a high risk of overfitting the algorithms and reduces the generalizability of the results. To increase the validity in the development of EEG-based diagnostic indices, attention should be paid to the composition of a training sample, its size, and its quality. The use of the combination of neurophysiological data labelled according to a single protocol from different clinical sites could be an optimal solution.
2. Treatment of Depressive Disorders
The main methods of treatment of depressive disorders include pharmacological therapy and psychosocial interventions (psychoeducation and group and individual psychotherapy)
[102][38]. Biofeedback and TMS are less commonly used. In individuals with depression, relevant therapeutic targets include maladaptive thoughts and beliefs, lowered mood, quality of life, and others. Reducing the risk of suicide in severe depression is also an important therapeutic goal.
2.1. Pharmacological Therapy
The most common medication used for treating depression is serotonin reuptake inhibitors (e.g., sertraline and paroxetine), tricyclic antidepressants (citalopram and fluoxetine), and monoamine oxidase inhibitors (moclobemide and pirlindole). The mechanism of action of antidepressants is based on a change in the concentration of neurotransmitters available for binding in the brain, which leads to long-term potentiation or depression of synaptic connections
[103][39]. The choice of medication is made based on their safety and the observed therapeutic effect in a particular patient. Doses are adjusted by the medical professional during treatment. Antipsychotics can also be used in depressive states that have arisen as part of psychotic disorders (e.g., bipolar disorder)
[104][40]. Pharmacological therapy is usually only required for moderate to severe depressive disorders. The combination of pharmacological therapy and psychotherapy is optimal for the treatment of such forms of depression
[105][41].
2.2. Psychosocial Interventions
The main psychosocial intervention for working with patients with mild depression is psychoeducation, providing information about the symptoms and the possible progression of the disorder, recommendations for self-help, and lifestyle changes (if necessary). Other suitable psychosocial interventions for patients with depressive disorders are group and individual psychotherapy. The most studied approaches for the treatment of depression are cognitive behavioral therapy and its variations, psychodynamic therapy, schema therapy, and decision-oriented therapy
[106,107][42][43]. In general, psychotherapeutic approaches do not differ significantly in effectiveness, so the choice of a particular approach in each case depends on the availability of a specialist and the individual preferences of the patient. Besides psychotherapy, regular physical exercise could also be considered a part of a psychosocial intervention. Regular exercise has been shown to be effective in reducing depressive symptoms
[108][44]. Positive changes may result from the metabolic and hormonal changes that accompany regular exercise.
2.3. Biofeedback
Biofeedback methods have also been shown to be effective in reducing depressive symptoms. A meta-analysis of 14 randomized controlled trials in 794 subjects showed significant improvements in symptoms in patients with depression
[109][45]. Neurofeedback using EEG and fMRI demonstrated promising results in reducing depressive symptoms, but the number of published studies is still limited
[110,111][46][47].
2.4. Brain Stimulation
Studies demonstrated that transcranial magnetic stimulation had a potentiating effect on antidepressant intake and was associated with improved clinical outcomes
[112][48]. TMS is recommended by several national medical agencies for use in the treatment of depression as a procedure with potentially positive results and no side effects
[113][49]. However, there is currently insufficient data to draw definitive conclusions about the effectiveness of TMS in reducing depressive symptoms.
2.5. Application of Machine Learning
Machine learning is also used to assess the treatment effectiveness in depressive disorders
[114][50]. The effectiveness of depression treatment using antidepressants or TMS can be assessed by using and applying classical machine learning methods (support vector machine
[115][51] and random forest
[116][52]) or deep neural network
[97][33] algorithms to EEG data
[117][53].
The machine learning methods used and the accuracy achieved in predicting depression treatment based on EEG data are presented in Table 4.
Table 4.
EEG predictive accuracy of treatment for depressive disorder.
Studies on the predictive treatment of depression, as well as studies on the diagnosis of depression, are characterized by a limited number of participants, which does not allow us to assert that the methods used have sufficient generalizability. At the moment, the most promising areas for the development of applied methods of machine learning for the therapy of depression are dynamic monitoring of the symptoms and treatment of pharmacoresistant depression.