Depression is a common mental disorder primarily characterized by a depressed mood and anhedonia, or loss of interest or pleasure. Other secondary symptoms such as appetite or weight changes, sleep difficulties (insomnia or hypersomnia), psychomotor agitation or retardation, fatigue or loss of energy, diminished ability to think or concentrate, feelings of worthlessness or excessive guilt and suicidality have also been reported [1]. Intensive research is being carried out on its possible causes (genetic, neurological, stress-related, inflammatory, immunological, microbiome-related, etc.), using a variety of methodological and disciplinary approaches (electroencephalography or EEG, neuroimaging, clinical medicine, psychiatry, biology, animal models, drugs, etc.). Yet, no conclusive unifying criteria or standardized biomarkers have been found [2].
It is estimated that 3.8% of the global population is affected by depression, including 5% of adults and 5.7% of adults over 60. Globally, 280 million people suffer from depression, which has been included as one of the critical/imperative conditions covered by WHO’s Mental Health Gap Action Programme (mhGAP) [3]. The Global Burden of Diseases, Injuries and Risk Factors Study showed that depression caused 34.1 million of the total years lived with disability (YLD), ranking as the fifth largest cause of YLD overall [4]. Furthermore, depression is a concomitant factor in different mental health conditions, like Adjustment Disorder [5], Alzheimer’s Disease [6], and others. With depression increasingly being a public health issue, its early detection is an important factor in reducing overloaded primary care systems and improving quality of life [7].
Depression is mainly diagnosed through questionnaires and clinical interviews based on the expertise of professionals with the guidance of WHO’s International Classification of Diseases (ICD 11) or the DSM-V in the USA. However, such an approach does not provide an entirely reliable diagnosis, since patients’ answers about their well-being or expectations of improving their mood are quite subjective, and also because emotional links between patients and doctors could be involved. All that can lead to a biased interpretation of questionnaires, making it necessary to advance in early depression detection through biomarkers that could help both in the diagnosis and in the evaluation of treatment [8, 9].
The quest for objective diagnosis has led to the development of EEG-based analysis protocols like Event Related Potentials, Band Power, Signal Features, Functional Connectivity, and Alpha Asymmetry, with average classificatory accuracies of 90% [10]. At the same time, the large amount of data provided by EEG calls for the use of fast, automatic machine learning algorithms.
The computer implementation of Machine Learning (ML) algorithms to automatically classify large amounts of data is derived from early proposals such as the brain learning mechanisms hypothesized by Hebb in 1949 [11], or the development of the perceptron by Rosemblatt in 1957 [12]. Since then, the number of algorithms and neural architectures has grown and their use has spread into almost every domain of daily life. Deep Learning (DL), as a new area of ML, combines multiple layers of the same or different basic architectural modules that have achieved goals such as speech recognition, translation or even the creation of online chatbots that help people to improve their mental health [13, 14]. The success of DL in these fields has encouraged many researchers to implement these methods for EEG data classification in the field of mental health [15].
A recent review on the use of resting-state EEG data classification for depression [16] pointed out the principal pitfalls and successes of research studies in this domain: 1) All used ML or statistical techniques for classifying, 2) Non-linear features worked best, 3) Electrode selection was based on theoretical criteria of the researchers and ranged from 1 to 30 electrodes, 4) Classification accuracy ranged from 80 to 99.5%, 5) None reported capacity of generalization of their classifiers to data from other sources, 6) The lack of public EEG databases with resting-state data on depressive individuals is an obstacle.
In our study, the ML and DL classifiers were initially applied to the resting-state EEG data obtained from a sample of university students with high and low scores in a depression questionnaire (BDI). We used a data-driven approach to electrode selection, to avoid a priori decisions about whether few or many electrodes are most appropriate. We selected only non-linear features, since this approach seems the most efficient strategy in depression research. Finally, we searched publicly available databases with EEG resting state to test whether the classifier trained with our participants’ data could be generalized to data from different populations.