Depression, or depressive disorder, is a common disease. According to the World Health Organization (WHO), the number of people with depression was estimated at more than 300 million affected worldwide [1]. Depression may severely impact well-being and functioning at work, school, and family, and can even lead to self-harm. Adolescent depression is associated with mood disorders and severe mental illness in adult life [2, 3]. Nearly 0.8 million people die from suicide each year and suicide is the fourth leading cause of death in 15–19-year-olds, according to WHO [1]. Amongst the top major diseases causing disability or incapability, five are mental illnesses—depression being the most prominent of these [4]. Hence, the disease burden due to depression is vast. The prevalence of depression in the adult population is approximately 5% across cultures, and 20% in its milder forms (i.e., partial symptoms, mild depression, and probable depression) [5]. Among adults, those most at risk are within the middle-aged population. Also, the world-wide occurrence of depression is increasing, with a rise of 18% between 2005 and 2015. However, early professional intervention can improve mental symptoms (e.g., absence of self-confidence and rumination) and resolve somatic problems (e.g., gastrointestinal problems and sleeping disorders) in most of the cases [6, 7].
Early detection of depressive symptoms followed by assessment and treatment can considerably improve chances for curbing symptoms and the underlying disease; mitigate negative implications for well-being and health as well as personal, economic, and social life [7,8,9,10]. However, detection of depressive symptoms is challenging and resource demanding. Current approaches are mainly based on clinical interviews and questionnaire surveys by hospitals or agencies [11], where psychological evaluation tables are utilized to make predictions on mental disorder. This approach is mostly based on one-to-one questionnaires and can roughly diagnose the psychological disorder for depression.
An alternative approach to interview or questionnaire-based predictions of depression is the analysis of informal texts provided by users. Previous studies in clinical psychology have shown that the relationship between the user of a language (e.g., speaker or writer) and their text is meaningful and has potential for the future [12]. A recent study by Havigerová et al. indicate a potential for text-based detection of persons at risk for depression, using a sample of informal text written about a holiday [12]. Hence, online records and data are increasingly seen as a valuable data source in supporting health care with decision support. The approach to identify depression symptoms from informal texts is promising, as it allows for benefitting from recent advances in natural language processing and Artificial Intelligence (AI). AI applied for natural language processing employs linguistics and computing techniques to help machines to understand underlying phenomena such as sentiments or emotions from texts. In that case, the core intent is to analyse opinions, ideas, and thoughts via the assignment of polarities either negative or positive.
Previous work has found that automatic analysis of depression symptoms from texts can be applied in, for example, sentiment retrieval from suicide notes and detecting insulting or depressive words or sentences in conversations or blog posts [13,14,15,16,7,18]. However, there is still substantial untapped potential in research on extracting depressive symptoms from texts. Key challenges include portraying significant cues of depression from texts. Also, there is a substantial hurdle in detecting depression symptoms from short texts.
To contribute towards solving these challenges, we aim to develop an automatic algorithm for detecting depression symptoms in texts, using a text-based sample of young people seeking advice about self-perceived depressive symptoms. We believe our automatic detection approach, describing the problems of the users in natural language, can be a substantial contribution to this research field. Hence, the current study focuses on how symptoms of depression are manifested through text in natural language using AI.
To visualize sample data of different groups in different applications, Linear Discriminant Analysis (LDA) is a good tool for data visualization based on discriminations [19,20,21,22]. It works on grouping of samples of similar classes. It tries to find the directions where the classes are best separated by considering minimizing the within-class scatter while maximizing the between-class scatter. LDA has already been used in various practical applications such as facial emotion recognition and human activity recognition. LDA projects the sample data of different classes onto a lower-dimensional vector space. Thus, the ratios of the between-class scatter and the within-class scatter is maximized to achieve highest discrimination.
Deep neural network has been contributing a lot recently in enormous fields of research, especially in pattern recognition and AI [23,24,25,26,27,28,29,30,31,32,33,34]. Though it is more robust than typical neural networks, it however consists of two major disadvantages. The first disadvantage is overfitting problem most of the time. The last one is taking much time for modelling the underlying data. The first successful deep learning algorithm was deep belief network that consisted of Restricted Boltzmann Machines (RBMs) that made the training quite faster than other previous learning approaches. Later, convolutional neural networks (CNN) was proposed and got popular especially in image processing fields. It showed better discriminative power compared to other approaches. CNN also extracts features alongside training the data. It has some convolutional stacks to generate a progressive hierarchy of abstract features via convolution, pooling, tangent squashing, rectifier, and normalization [24]. CNN is mostly applied for image and video pattern analysis rather than temporal information decoding. Hence, it has not been adopted for time-sequential data analysis. Recurrent Neural Networks (RNNs) is however a better choice than CNN since it consists of better discriminative power over others in case of sequential data and pattern analysis [30]. Since the basic RNNs usually consist of vanishing gradient problem due to long-term dependencies when it handles high-dimensional and time-sequential data, Long Short-Term Memory (LSTM) was introduced in RNN to overcome it. Hence, this work utilizes the advantage of LSTM-based RNN to model different emotional states in text data.
Among different approaches to analyse physical and mental states of human being from different data sources, machine learning has been very widely used [35,36,37,38,39,40,41]. Since machine learning models are progressively being employed to do significant predictions in crucial contexts day by day, the demand of transparency rises in such contexts from the various stakeholders in AI industry [42]. The high risk in this regard is making and applying the AI decisions that are unjustifiable and lacks explanations of the models' behaviour. Hence, explanations of the output of a model are vital. For example, specialists in precision medicine fields need further information from the machine learning models than simple prediction for supporting their diagnosis. Such necessities may also arise in other fields as well, such as medical emergencies. Hence, focusing merely on the performances of the AI models, gradually makes the systems towards unacceptance in some cases. Therefore, current research has highlighted the importance of explainable Artificial Intelligence (XAI) for establishing trust in machine learning-based decisions through the explanations of the black-box models. Popular state-of-the-art explanation algorithms include Local Interpretable Model-Agnostic Explanations (LIME), SHapley Additive exPlanations (SHAP), and layer-wise relevance propagation (LRP). From which, LIME is very light-weight and yet tries to generate quick and satisfactory post-hoc explanations. Hence, this work adopts LIME to see the explanations (i.e., importance of the features) once the decision is provided by the model.
Contribution
This work focuses on processing text data, features, and depression symptoms text recognition with the target of chatbot as a smart application. Figure 1 shows a schematic setup of a text-based depression symptoms text detection system in a smart application where a user provides a query in text format and a server processes the text to apply feature extraction and deep learning. Based on the results, the server can suggest further advices to the user. Figure 2 shows the basic architecture of the proposed system consisting of training and testing procedure for the classification of texts describing symptoms of depression. In the training part, text data from all the users is obtained and then the features are trained using RNN. In the testing part, features from a sample test are applied to the trained model to take the decision whether the user describes depression or not. LDA is applied to show the robustness of the proposed features compared to other traditional ones. Finally, we apply one of the most popular algorithms (i.e., LIME) for post-hoc, local, and meaningful explanations of the machine learning decision regarding the existence of a potential depression or not, in the text. The contribution of the paper can be summarized as bellow:
-
A large dataset of text is obtained from a public Norwegian online information channel: ung.no.
-
Novel features are extracted representing the possible symptoms of depression defined by the experts from medical and psychology domains.
-
RNN is applied based on LSTM, attention, and dense layers for modelling the emotional states.
-
The machine learning decisions are explained using a state-of-the-art XAI approach, LIME to see the importance of the features.