Review and Content Analysis of Textual Expressions as a Marker for Depressive and Anxiety Disorders (DAD) Detection Using Machine Learning

doi:10.21203/rs.3.rs-3341966/v1

Download PDF

Systematic Review

Review and Content Analysis of Textual Expressions as a Marker for Depressive and Anxiety Disorders (DAD) Detection Using Machine Learning

https://doi.org/10.21203/rs.3.rs-3341966/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background

Depressive disorders (including major depressive disorder and dysthymia) and anxiety (generalized anxiety disorder or GAD) disorders are the two most prevalent mental illnesses. Early diagnosis of these afflictions can lead to cost-effective treatment with a better outcome prospectus. With the advent of digital technology and platforms, people express themselves by various means, such as social media posts, blogs, journals, instant messaging services, etc. Text remains the most common and convenient form of expression. Therefore, it can be used to predict the onset of anxiety and depression.

Methods

Scopus and Web of Science (WoS) databases were used to retrieve the relevant literature using a set of predefined search strings. Irrelevant publications were filtered using multiple criteria. The research meta data was subsequently analyzed using the Biblioshiny Tool of R. Finally, a comparative analysis of most suitable documents is presented.

Results

A total of 103 documents were used for bibliometric mapping in terms of research outcome over the past years, productivity of authors, institutions, and countries, collaborations, trend topics, keyword co-occurrence, etc. Neural networks and support vector machines are the most popular ML techniques; word embeddings are extensively used for text representations. There is a shift toward using multiple modalities. SVM, Naive Bayes, and LSTM are the most commonly used ML methods; social media is the most commonly used source of data (Twitter is the most common platform); and audio is the most commonly used modality that is combined with text for depressive and anxiety disorders (DAD) detection.

Conclusions

Asian countries are leading in the research output with China and India being the top countries in terms of the number of research publications. However, more international collaborations are needed. Limited research exists for anxiety disorders. Co-occurrence of anxiety and depressive disorders is high (33% of studies).

Depression and anxiety detection

machine learning

text analytics

natural language processing

mental disorders

Mental disorders happen to be the leading health-related burden in non-communicable diseases [1]. The aftermath of the COVID-19 pandemic has demanded the strengthening of the mental healthcare systems, as during the pandemic the occurrence of mental disorders, particularly depressive and anxiety disorders, witnessed a growth of almost 25% in just one year, as per the WHO mental health report 2022 [2]. Furthermore, WHO estimates that around 12.50% of the world’s population suffers from some sort of mental illness, with depressive and anxiety disorders being the most prevalent [2]. Like any other type of mental disorder DAD not only cause personal suffering but also lead to reduced efficiency, rising healthcare expenditures, and diminished quality of life. DAD have significant implications for individuals and societies around the world [28]. Bibliometric analysis is a part of scientometric analysis of the literature to analyze sources, impact, trends, and other aspects of the published research on a given topic, done with several tools in a variety of fields of science [3, 4, 11, 12]. It can help a wide variety of readers quickly understand the research in a more qualitative way. The modern technology landscape, including artificial intelligence, machine learning, virtual reality, wearable sensors, etc., can help in several ways, ranging from detection to management of DAD [6, 7, 8]. DAD have a correlation with language use [7] and can act as a disease marker in a variety of settings, including blog posts [8], social media posts [9], telephonic conversations [13], etc. Early detection and intervention are necessary for effectively managing DAD, and ML based text analytics plays an effective way of screening DAD [14, 15, 16]. By analyzing an extensive collection of literature sourced from Scopus and the WoS, this study aims to unravel pivotal trends, identify influential authors, pinpoint significant research themes, and shed light on the most impactful journals. Through this meticulous analysis, the paper seeks to provide a comprehensive understanding of the landscape surrounding the application of ML-based text analytics in the early detection of DAD. The research meta data is analyzed using biblioshiny to draw insights on the research output growth, most productive authors, institutions, countries & sources, highly cited documents, research collaborations, keywords analysis, and emerging research themes. Another contribution of the present work is to review the state-of-the-art research work in DAD detection from text. Transitioning from discussing the innovative advancements in ML for DAD detection through text analytics, the focus of this study is on conducting a comprehensive bibliometric study on this critical topic.

The rest of the paper is organized as follows: Section 2 describes the materials and methods. Section 3 presents the key results of bibliometric analysis and mapping. Section 4 presents a review of the relevant literature with a comparative analysis of state-of-the-art studies on DAD detection from text using ML. Section 5 presents the discussion. Section 6 contains the conclusions of the study, and at the end, references have been listed.

2.1 Research Design and Article Selection

We collected papers from two citation databases, Scopus and WoS. Scopus serves as an extensive and diverse database encompassing peer-reviewed academic content, including scientific journals, books, and conference proceedings. The WoS is another database of peer reviewed academic publications. Relevant literature was obtained when the terms in the search query appeared in the title or abstract.

Table 1

Search Strategy to obtain the articles.
Search Query	"Early diagnosis" OR "Early detection" OR "Early identification" OR "Early screening" OR "Early intervention" OR "Early-stage detection" AND "Anxiety" OR "Generalized anxiety disorder" OR "GAD" OR "Social anxiety disorder" OR "Anxiety symptom" OR "Anxiety prevalence" OR "Anxiety treatment" OR "Anxiety intervention" OR "Anxiety risk factor" OR "Anxiety assessment" OR "Anxiety management" OR "Depression" OR "Major depressive disorder" OR "MDD" OR "Clinical depression" OR "Mood disorders" OR "Depressive symptoms" OR "Treatment-resistant depression" OR "Postpartum depression" OR "antidepressants" OR "Depression prevalence" OR "Depression risk factors" OR "Depression assessment" OR "Depression interventions" OR "Depression management" OR "Depression outcomes" AND "Machine Learning" OR "ML" OR "Artificial intelligence" OR "AI" OR "Natural language processing" OR "NLP" OR "Neural networks" OR "NN" OR "Supervised learning" OR "Unsupervised learning" OR "Semi-supervised learning" OR "Transformers" OR "Text Embedding" OR "BERT" OR "Word2Vec" OR "Transfer learning" OR "Model training" OR "Model evaluation" OR "Feature selection" OR "Feature Extraction" OR "Big data analytics" OR "Pattern recognition" AND "Text"
	# Documents from Scopus: 92	# Documents from WoS: 26

Table 1 shows the search string used for looking for the relevant publications in the databases. Furthermore, only publications in English language were included for the analysis. The search was performed on May 27, 2023. A total of 103 papers published from 2009 to 2022 were obtained.

2.2 Data Extraction and Loading

The study used the data downloaded from Scopus and WoS databases in a csv format. Both files were processed to see if they contained any duplicate articles. A sum of 15 articles was found to be duplicate. These articles were eliminated, and a combined file was obtained in XLSX format. R Studio was used to access “bibliometrix” package of R. The Biblioshiny (developed by Aria and Cuccurullo [3]) was used to analyze the dataset.

2.3 Data Synthesis

The search results revealed that the earliest paper was published in 2009. Due to this, the papers included in the study are from 2009 up to 2023 for analysis. The dataset contains articles, books, conference papers, conference reviews, and early access for analysis. Table 2 shows the summary of the dataset used in this study.

Table 2

Snapshot information of the dataset used in the study.
Description	Results
OVERVIEW OF DATA
Time Duration	2009:2023
No. of Sources	67
No. of Publications	103
Publication Growth Rate (Annual) %	16.01
Publication Average Age	2.74
Avg. Citation Count	9.223
No. of References	2255
KEYWORDS STATISTICS
Keywords Plus (ID) Count	773
Author's Keywords (DE) Count	255
AUTHORS STATISTICS
Authors	345
Authors of single-authored docs	4
COLLABORATION STATISTICS
Single-author Documents	7
Avg. No. of Co-Authors per Document	3.83
International co-authorships %	13.59
PUBLICATION TYPES
No. of Articles	27
No. of Book Chapters	3
No. of Conference Papers	60
No. of Conference Reviews	4
No. of Reviews	8
Reviews; Early Access	1

3.1. Scientific Production in the Field

Figure 1 shows the yearly publication count on the topic. The annual scientific production chart shows that there has been a recent increase in interest in text analytics for early detection of anxiety and depression, as almost 41% of the research in this field has been conducted in the year 2022. Additionally, this is further backed by the fact that from 2009–2020, 45 papers have been published, and in the past 3 years (2021–2023), 58 papers have been published. A higher publication output in the year 2022 is consistent with the COVID-19 effect on mental health [2]. The field is still in the early development stage, and more research will be conducted in the future, as suggested by the average annual growth of 16.01% (Table 2).

A three-field plot has been shown in Fig. 2 that shows the relationship among cited reference (CR), authors (AU) and the author keywords (DE), wherein the cited references have been shown in the left and author keywords on the right.

The top 10 most relevant sources are presented in Fig. 3. The data that was collected in May 2023 revealed that the most productive source of information in Scopus and WoS was the Ceur workshop proceedings. The Ceur workshop proceedings are the foremost source of information. The only other notable source includes lecture notes in computer science, with both sources having 67% of the documents. The remaining 8 sources constitute a small number of documents, with each being less than 3.

Figure 4 shows that the top 4 most productive authors in the field have produced the same number of documents, which is 4, and the remaining 5 out of 6 have produced the same number of documents as well. This shows that in the field, all the top authors are almost equally productive, as they have produced a similar number of documents.

Lotka's law is derived from Zipf's law, which focuses on how often authors publish within a particular field. It states that there is an approximate inverse-square law, meaning the ratio of authors publishing a specific quantity of articles remains constant in relation to the number of authors publishing just one article [10]. Hence, as the quantity of published articles rises, authors generating the same volume of publications become less frequent. From Fig. 5, we can see that although the author productivity exactly matches Lotka’s law, it is seen to be of similar shape, implying that Lotka’s law still holds true as the graph for author productivity is symmetrical to Lotka’s law.

Figure 6 shows the corresponding authors' countries. Single country publication (SCP) refers to intra-country collaboration that is represented by authors that belong to the same country, whereas publications involving authors from various countries, known as multiple country publications (MCP), are represented by collaborative efforts between different countries [11]. From Fig. 6, India, China, and Spain are seen to have the most publications, and they are all single-country collaborations, suggesting that they have a well-established research infrastructure in this field and a network of researchers in the country. Further, international collaboration can be seen as the UK and USA have the same MCP ratio of 0.5, implying the field is not in isolation and researchers have exposure to diverse perspectives, methodologies, and ideas from researchers in other countries.

A list of top globally cited documents has been given in Fig. 7. Losada De’s paper published in 2016 in the “Lecture Notes in Computer Science” is the most cited document with 154 citations, followed by Yang M. and Cheng O.J. The top 3 globally cited documents are seen to have the most impact and influence within the field, as they consist of 39% of all citations in this field, according to the data collected from Scopus and WoS. These documents can possibly be seen as a starting point to understand key concepts, theories, or methodologies in the field and influence the research direction. Although these are the most cited documents, newer studies, emerging trends, and alternative viewpoints should also be considered to gain a complete picture of the ongoing research in the field.

Reference spectroscopy characterizes the decay in popularity of a publication in terms of the number of citations over time (Fig. 8). As a publication becomes older, it receives fewer citations. The red line refers to a deviation from the 5-year median, referring to the expected citation trajectory for a set of publications based on their publication year. A deviation from the redline indicates a difference between the actual citation count of a publication and the expected citation count based on its publication year.

The first citation was seen in 1985, then directly after that in 2000, which explains why there is a straight one going through 0 from 1986 to 1999. From 2009, the black line is seen to be above the red line, which results in overperformance as there is a higher citation count than expected based on its publication year. This suggests that the publication has received more attention than anticipated. Further, deviation from the red line can help identify influential papers as they continue to be highly cited. For example, in 2017, there was a significant deviation from the red and black lines, signifying a highly cited paper was released in 2017. Furthermore, these papers are early indicators of the growing interest in a particular area and present important research findings that resonate with the research community.

3.2 Keywords and Thematic Analysis

Analysis of authors' keywords helps researchers in understanding and identifying the most relevant and interesting topics in this field [4]. Furthermore, these words provide a means for understanding the dimensions of current research and future trends.

Figure 9 shows the most used words in text analytics for early detection of DAD, with depression and ML being the most prominent. Instead of using the frequency of their occurrence as a representation of text size, we have used the log of the frequency so that less frequent words do not become too small.

Through the log of frequency occurrences of the authors' keywords, it is evident that there is research being conducted on the intersection of natural language processing, machine learning, deep learning, social media, linguistic metadata in the field. Further, the information from which data is being collected comes from social media, which is a prominent source. These topics can help guide research efforts to address the early detection of DAD. A tree representation of the top 50 keywords has been depicted in Fig. 10, wherein depression (occurrence = 10%), machine learning (occurrence = 10%), social media (occurrence = 7%), and natural language processing (occurrence = 7%) are the top keywords.

The co-occurrence of keywords can be understood from Fig. 11, where all these top keywords have a strong association in the network (node size and edge thickness).

Figure 11, which shows the co-occurrence network analysis, could be a useful tool to understand the knowledge structure of the field. It is evidence that depression is a constant term in the field, as are almost all the prominent key terms, suggesting that there is more research involving depression. Furthermore, strong connections are seen between depression and ML, early detection, NLP, and social media, suggesting that more research is being conducted within these fields in relation to depression.

In Fig. 12, the trending topics are based on the authors' keywords in the time span 2018–2022, with the minimum word frequency of 5 and the maximum of 25. Keywords that are connected to publication content characterize topical aspects of a research field [4]. In 2018, early detection was the most used and remained in trend until 2022. From 2019 to 2022, “machine learning," “mental health," and “early risk detection” were trending, with a greater frequency of “machine learning” being used. Further in 2022, “natural language processing” seems to be the trending topic as it has a term frequency of 20, and there’s a small emphasis on “transfer learning” as it has a word frequency of 5. Through this, we can assume that more research has shifted towards natural language processing. Although “natural language processing” appeared in 2021, it appears to be the most trending topic, but terms such as “machine learning” and “social media" are still relevant.

Figure 13 provides the thematic map for early detection of DAD using text analytics. A typical thematic map has four themes with the top left being the specialized themes, the top right being the driving themes, the bottom left being the declining themes and bottom right consisting of the main themes [6]. The aim of the thematic map is to get a comprehensive view of the current condition of the field and evaluate its prospects. The information is crucial as it helps determine prospective avenues for researchers and stakeholders to explore in identifying the progression of research development. Thematic analysis groups most representative authors’ keywords into various themes which are made up of centrality and density, on x-axis and y-axis respectively.

Density is a measure of cohesiveness, while centrality is a measure of correlation among various nodes [5]. The parameters to generate the thematic map were the number of words equal to 50, the minimum cluster equal to 5, and the clustering algorithm was Louvain. Nodes are formed due to the clustering of related terms. Nodes represent individual terms or concepts, and their arrangement within the map indicates their relationships and similarities based on co-occurrence patterns in the literature.

From the thematic representation, we can make out that suicide, multimodality, and mental disorders are clusters in a node and are part of the emerging or declining themes. The motor themes in a thematic map represent the central, influential, and interconnected topics or concepts within a research domain. They help researchers identify key research areas, understand the structure of knowledge, and gain insights into the dynamics and trends shaping the field. The motor themes include CNNs and support vector machines. The field is also seen to have many basic or main themes, with there being three nodes, each significantly bigger than the other themes.

A "collaboration network" among authors refers to the network of collaborations and co-authorship relationships among researchers (Fig. 14). It helps with the identification of research communities and provides insights into the knowledge diffusion, specialization, and collaboration dynamics within a field. The collaboration network illustrates the flow of knowledge within a research community. Researchers in the network share and exchange ideas, methods, and findings through their collaborations. The network can provide insights into how knowledge spreads, disseminates, and influences the research community. In the field, researchers work in collaboration; however, they work in their own groups, with little dissemination of knowledge outside their groups. This can be seen as there are nine research groups and no communication outside any of the groups.

Table 2

Parameter values in generating the conceptual structure.
Parameter	Value
field	DE
ngrams	1
method	MCA
minDegree	2
clust	3
k.max	8
stemming	FALSE
labelsize	6
documents	5
graph	FALSE

To determine the correlation between the keywords, a factor analysis was performed. Multiple correspondence analysis (MCA) was used to create three clusters (shown in blue, green, and red colors in Fig. 15) based on the word affinity as it appears throughout the documents. Table 2 lists the values for additional parameters.

A dendrogram is a tree-like representation for showing hierarchical relationships among different data points. Figure 16 shows a dendrogram representation of different author keywords divided into three categories with a unique color-coding scheme.

The development in DAD detection from machine learning is yet to mature. However, in recent times, some progress has been made. Here, we provide a critical review of the recent work. Table 3 provides a summary of key findings of this review, where a comparative analysis among different studies has been presented in terms of disorder (depression, anxiety, or both), description of the dataset used in the study, ML technique used (for best reported performance), performance metric (accuracy, AUC score, sensitivity, specificity, recall, f-score, etc.), and remarks (any significant contribution or limitation of the study).

Although the prevalence of anxiety and mental health disorders is widespread, however, there is less focus on anxiety detection. In our comparative analysis, there are just 6% of studies solely addressing anxiety, while depressive disorders are found in 61% of the studies, and 33% have studied both.

Earlier, diagnosis was reliant on subjective measures, for example, clinical evaluations and self-reporting; however, through technological innovations, the ability to analyze enormous textual data presents new opportunities through ML-based techniques, enabling effective early detection in mental health analysis [29].

The text data is easily available from a variety of sources, such as social media, forums, and health records. ML can process and interpret text and help us with early detection of DAD. Through this, we can enhance the accuracy, efficiency, and scalability of DAD detection. The integration of ML methods in mental health research, particularly concerning DAD detection through text analytics, has ignited a new wave of exploration. Wang et al. propose the fusion of deep learning techniques involving fusion of deep learning techniques with textual data from online platforms [30], while Babu et al. (2022) provides an exhaustive review of sentiment analysis from social media data for depression detection [31]. Ortega-Mendoza et al. [32] investigate the efficacy of NLP methods in extracting mental health insights from social media text data, including analysis of personal statements. Similarly, Khan et al. [33] have used social media texts to detect aggressive and anxious behaviors on social media. DeepTweet [34] is such a system that is based on the analysis of short tweet texts for detecting mental afflictions of DAD. A review of using non-clinical data (particularly from social media posts) in mental health applications has been provided by Calvo et al. [35]. Analysis of the social media and blog text can further pave the way for providing personalized healthcare interventions, including the suggestion of therapy and self-care practices [35, 36, 37, 38].

Table 3

Comparative Analysis of Different Techniques for DAD Detection from Text using ML
Study	Year	Disorder (Depression/ Anxiety/Both)	Dataset	ML Technique	Performance / Acc.	Remarks
Shen et al. [14].	2017	Depression	They constructed two datasets on twitter: Depression and non-depression. They extracted six depression related group features.	Multimodal dictionary-based learning model	Accuracy 40–60%	The study covered online behaviors on social media and was not limited to clinical depression criteria.
Tyshchenko et al. [8].	2018	Both	Blog posts from clinical and control subjects	Support Vector Machines (SVM) and Convolutional Neural Network (CNN) classifiers	Accuracy: 78% Recall score of 0.72.	BoW, TFIDF and topic modeling techniques used for feature extraction. Claims to have one of the highest recall scores.
Al Hanai et al. [16]	2018	Depression	The data utilized consisted of 142 individuals who underwent depression screening.	LSTM-NN model	Accuracies Context modeling:0.71 Weighted modeling:1	One limitation is the sample size (just 142).
Stankevich et al. [9]	2019	Depression.	Textual messages of Russian social network VKontakte users	An implementation of logistic regression and random forest ML techniques.	66% of F1-score on the binary classification task.
Budiyanto et al. [22]	2019	Both	22,934 Facebook posts	DASS21 and naïve Bayes classification	Root cause analysis of depression to find out the influential factors for DAD	The users who are the resident of big cities, women, with a job status; are more prone to develop DAD.
Tlachac et al. [15].	2020	Depression	Response latency data in sending text messages.	ML Models built on the component of latency features	The best accuracy of 0.69, F1 score of 0.67, and AUC of 0.72	Nine features associated with response delay from meta-data of text message conversations collected from a crowd-sourced setting.
Souza et al. [17]	2020	Both	Self-diagnosed dataset extracted from Reddit	Deep learning techniques, performed a qualitative analysis using SHAP.	The model achieved a Exact Match Ratio (EMR) of 0.47, Hamming Loss (HM) of 0.29 and.	The model presented outperformed all the baseline models through Binary classifiers that make predictions about specific disorders.
Rutowski et al. [19]	2020	Both	Input from conversational speech	Deep Language Models and Transfer Learning	AUC score in the range of 0.79–0.86
Bucur et al. [20].	2020	Depression	eRisk 2018 dataset	Topic analysis and learned confidence scores	Best results for the ERDE5 metric.
Kim et al. [23]	2020	Both	Reddit posts belonging to multiple mental disorder classes, including depression and anxiety among others.	CNN	Depression accuracy: 75.13% Anxiety accuracy: 77.81%	They used a total of 6 mental afflictions and compared the performance of XGB with CNN.
Amanat et al. [12]	2022	Depression	Principal component Analysis along with text encoding.	LSTM + NN	99.0% accuracy.	No mentions of methodological flaws.
Kwon et al. [13]	2022	Both	Recordings of telephonic conversations at a call center.	Classification model using x-vectors	Achieved a sensitivity scire of 0.70 and a specificity score of 0.54 for DAD detection task.	The model’s performance is consistent with that of the questionnaire-based tests (GAD-7 and PHQ-9).
Shin et al. [18].	2022	Depression	Recordings of the Mini International Neuropsychiatric Interview	A naïve Bayes classifier model used for classification	The model achieved AUC score of 0.905, a sensitivity score of 0.699, and a specificity score of 0.964.	This study confirmed diagnosing depression using interview text.
Tejaswini et al. [21]	2022	Depression	The first dataset contains post from Reddit and the second dataset was taken from Kaggle and consists of tweets on Twitter	Hybrid Deep Learning Model (Fasttext + LSTM)	Accuracy: 87%	Fasttext embedding with a neural net classifier performs better than other combinations of word embeddings and classifiers.
Firoz et al. [24]	2023	Depression	DAIC-WOZ dataset (transcripts of patients’ interviews)	Hybrid CNN-LSTM model	Accuracy: 80%	Hyperparameter tuning to get the best model parameters. Limited dataset of native English-speaking subjects.
Alqazzaz et al. [25]	2023	Depression	Dataset contained 36587 Tweets	CNN-LSTM model	Accuracy:86.23%	Detection of depression from emotional phrases. The proposed model performs better than the traditional ML methods as well as a lexical approach.
Gupta et al. [26]	2023	Depression	SentiEmoDD dataset of Tweets containing text and emojis	SVM	Text only accuracy: 88.29% Text + Emoji Accuracy: 87.69% (using stacking of SVM, Logistic Regression and KNN)	Provides a comparative analysis of different approaches for the text-only and multimodal depression detection.
Yu et al. [27]	2023	Anxiety	Social media text retrieved from the Chinese social media platform - Sina Weibo.	XGBoostRegressor	Reliability score of 72%	Pearson correlation score between the predictions made by the model and the self-report questionnaire were moderate. The reliability of model was high (0.72).

4.1 Modalities, Data Sources, ML Techniques, and Other Aspects in DAD Detection

Based on the modality of data for DAD detection, the studies can be divided into two groups: text-only and multimodal [24, 26]. The data can be sourced from social media posts [9, 14, 21], clinical and non-clinical interviews [13, 17], patient self-descriptions [17], blogs [8, 23], and instant messaging features [15]. Twitter is the most common source of data used in the studies. When it comes to ML methods, SVM [8, 26] and naive Bayes classifier [18, 22] are the most used methods under traditional methods, while in deep learning, LSTM [12, 16, 21, 25] and CNN [8] are the popular methods, with BERT, Fasttext, and ELMO as the embeddings. Figure 18 shows a mind map around various essential components and concepts of DAD detection in text.

4.2 Multimodal Textual Data for DAD Detection Tasks

Apart from text-only data, different other modalities of data can be combined with textual data. The popular modalities that researchers have used include combinations of text + audio, text + video + audio, text + emojis, text + images, etc. [39, 40, 41]. As traditional ML methods rely on feature extraction methods for learning from these data, a variety of feature extraction [42, 43], dimensionality reduction [44], and feature fusion techniques [43, 45] are employed. Term frequency-inverse document frequency (TF-IDF), Bag of words (BoW), word embeddings (Word2Vec, GloVe, FastText, BERT, etc), n-grams and character-level features are some of the commonly used feature extraction methods for textual data. Abiodun et al. [46] present a systematic review of application of various optimization techniques in extraction and selection of textual features. The rapidly expanding landscape of ML has spurred innovative advancements in DAD detection through text analytics. Existing research highlights the capabilities of ML algorithms in detecting subtle linguistic cues and emotional patterns in textual data. Recent research emphasizes the integration of multiple modalities while delving into the potential of feature selection techniques in enhancing model performance [39]. Nonetheless, the capability of deep learning techniques (like convolutional neural networks) for end-to-end learning with and without intermediate representations (in the form of embeddings) and the ability to handle custom input and output (including multiple I/O) make them a suitable choice for training on multimodal textual data. The deep learning models are more suitable for training on multiple modalities of data and producing a multi-label output very efficiently.

This study involved a bibliometric analysis of research publications related to early detection of DAD using text analytics from 2009–2023 from two research databases, WoS and Scopus. By employing a visual and exploratory analysis on the research meta data, we mapped the bibliometric findings in terms of research output, productivity, collaborations, and trend analysis. Here, we discuss our key findings, compare them with existing literature, interpret their implications, and address the strengths and limitations of our analysis.

Although we noticed a zig-zag pattern in annual scientific production, there is a clear increase in the scientific production, indicating growing interest in the research area. The citations, however, seemed to have peaked in 2015 and 2017, implying that the work presented in these years was influential and attracted a substantial amount of recognition in that year. Although citations are noted to be a lagging indicator, citations take time to gain recognition and accumulate citations, which may explain why the peak occurred in a previous year. The Ceur workshop proceedings and Lecture Notes in Computer Science were the most productive publication sources in this field. The SRM Institute of Science and Technology and the Universidad Nacional De Educación a Distancia (UNED) are seen as being the most affiliated in the field. For individual researchers, the top 4 seem to be equally productive: Fredrich C., Koitka S., and Trotzek M. The most productive countries are China, India, and Spain, with all of them having single-country publications, which implies that there is no inter-country collaboration among the most productive countries and suggests that they have a well-established research infrastructure. Depression detection and natural language processing (NLP) were identified as the trending topics, with depression being the main topic of research. This suggests that this field needs multidisciplinary approaches to gain a better understanding of each topic. This is further highlighted through the co-occurrence word network.

The comparison to other literary works is essential to positioning our research within the existing body of knowledge on the early detection of DAD using text analytics. In this section, we analyze and discuss how our findings align with, build upon, or diverge from previous studies. We conduct a trend analysis with several seminal works in the field, including "A Test Collection for Research on Depression and Language Use" by Author Losada De, "Depression and Anxiety Detection from Blog Post Data" by Author Ye. O. Tyshchenko, and "Depression Detection from Social Media Texts" by M. Stankevich et al. These works have been widely recognized for their significant contributions to the field's understanding.

“A Test Collection for Research on Depression and Language Use” happens to be the globally most cited document meaning that it has garnered significant attention and recognition from researchers worldwide, evidenced by the large number of citations it has received. This document proposes a new novel detection method and creates a dataset using Twitter, ATL and reddit [7]. DAD detection from blog posts data which has applied support vector machines and convolutional networks for early detection of DAD. Further, they’ve gathered blog posts from clinical and control subjects. With their model, they achieved a high accuracy of 78% [8]. Depression Detection from Social Media Texts where they prepare users’ profiles dataset and propose psycholinguistic and stylistic markers of depression in text. Through their ML methods, they have achieved an accuracy score of 66% which is promising as compared to other works [9]. Subsequently, we conducted a trend analysis review on these articles, and research trends were identified. These articles employed data mainly through social media such as Reddit, blog posts, Twitter, etc. In addition, we’ve noticed that all three research papers employ approaches that are concerned with artificial intelligence for the early detection of DAD. Although the three papers use different methods, none of them appear to be highly accurate, despite being some of the highly cited papers in this field. Through this, we infer that more research requires to be conducted in this field, which could increase accuracy.

As far as access to the mental health services is concerned, there are disparities among different groups that are affected by demographic, social and financial, and other factors [47]. Digital interventions can help improve accessibility and quality of services [48, 49, 50].

ML-based text analytics for DAD detection is an evolving field with significant potential for timely interventions. This review covered research conducted over the past 14 years (2009–2023) and observed a surge in publications in 2022, largely attributed to the increased prevalence of disorders during the COVID-19 pandemic. Commonly used ML methods include SVM, Naive Bayes, and LSTM, while social media data, especially from Twitter, is the primary data source. Additionally, audio data is frequently combined with text for DAD detection. It's worth noting that while both depression and anxiety are prevalent, more research has focused on detecting depressive disorders. Future research should emphasize the detection of anxiety disorders, potentially incorporating new data sources, innovative feature extraction methods, and novel classification techniques. Ethical concerns regarding AI usage and the need for transparency are essential considerations in this field. Ensuring that classification models are more transparent and explainable will be crucial. As the field matures, interdisciplinary collaboration and methodological refinement will play pivotal roles. This study not only consolidates existing knowledge but also emphasizes the urgency of refining these technologies to enable early detection and improved management of critical mental health challenges.

Competing interests: The authors declare no competing interests.

Santomauro, D. F., Herrera, A. M. M., Shadid, J., Zheng, P., Ashbaugh, C., Pigott, D. M., ... & Ferrari, A. J. (2021). Global prevalence and burden of depressive and anxiety disorders in 204 countries and territories in 2020 due to the COVID-19 pandemic. The Lancet, 398(10312), 1700-1712.
World Health Organization. (2022). Mental Health Report 2022. Retrieved July 15, 2023, from https://www.who.int/teams/mental-health-and-substance-use/world-mental-health-report
Aria, M., & Cuccurullo, C. (2017). Bibliometrix: An R-Tool for Comprehensive Science Mapping Analysis. Journal of Informetrics, 11, 959-975.
Song, Y., et al. (2019). Exploring Two Decades of Research on Classroom Dialogue by Using Bibliometric Analysis. Computers & Education, 137, 12-31.
Esfahani, H., Tavasoli, K., & Jabbarzadeh, A. (2019). Big Data and Social Media: A Scientometrics Analysis. International Journal of Data Network Science, 3, 145-164.
Jingili, N., et al. (2023). Virtual Reality for Addressing Depression and Anxiety: A Bibliometric Analysis. International Journal of Environmental Research and Public Health, 20(9), 5621.
Losada, D. E., & Crestani, F. (2016). A Test Collection for Research on Depression and Language Use. In Lecture Notes in Computer Science, 28-39.
Tyshchenko, Y. (2018). Depression and anxiety detection from blog posts data. Nature Precis. Sci., Inst. Comput. Sci., Univ. Tartu, Tartu, Estonia.
Stankevich, M., et al. (2019). Depression Detection from Social Media Texts. In A. Elizarov, B. Novikov, & S. Stupnikov (Eds.), Data Analytics and Management in Data Intensive Domains: XXI International Conference DAMDID/RCDL, 352.
Qiu, J., et al. (2017). Author Distribution of Literature Information: Lotka’s Law, 145–183. https://doi.org/10.1007/978-981-10-4032-0_6.
Sweileh, W. M. (2020). Bibliometric Analysis of Peer-Reviewed Literature on Climate Change and Human Health with an Emphasis on Infectious Diseases. Globalization and Health, 16(1).
Amanat, A., et al. (2022). Deep Learning for Depression Detection from Textual Data. Electronics, 11(5), 676.
Kwon, N., Hossain, S., Blaylock, N., O’Connell, H., Hachen, N., & Gwin, J. Detecting Anxiety and Depression from Phone Conversations using x-vectors. In Proc. Workshop on Speech, Music and Mind (pp. 1-5).
Shen, G., et al. (2017). Depression Detection via Harvesting Social Media: A Multimodal Dictionary Learning Solution. In Proceedings of the 26th International Joint Conference on Artificial Intelligence, 3838-3844.
Tlachac, M. L., & Rundensteiner, E. A. (2020). Depression Screening from Text Message Reply Latency. In 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), 5490-5493.
Al Hanai, T., Ghassemi, M., & Glass, J. (2018). Detecting Depression with Audio/Text Sequence Modeling of Interviews. Proc. Interspeech, 2018, 1716-1720.
Souza, V., Nobre, J., & Becker, K. (2020). Characterization of Anxiety, Depression, and Their Comorbidity from Texts of Social Networks, 121-132.
Shin, D., et al. (2022). Detection of Depression and Suicide Risk Based on Text From Clinical Interviews Using Machine Learning: Possibility of a New Objective Diagnostic Marker. Frontiers in Psychiatry, 13, 801301.
Rutowski, T., et al. (2020). Depression and Anxiety Prediction Using Deep Language Models and Transfer Learning. In 7th International Conference on Behavioural and Social Computing (BESC), 1-6.
Bucur, A. M., & Dinu, L. P. (2020). Detecting early onset of depression from social media text using learned confidence scores. arXiv preprint arXiv:2011.01695.
Tejaswini, V., Babu, K. S., & Sahoo, B. (2022). Depression Detection from Social Media Text Analysis Using Natural Language Processing Techniques and Hybrid Deep Learning Model. ACM Transactions on Asian and Low-Resource Language Information Processing, Nov.
Budiyanto, S., Sihombing, H. C., & IM, F. R. (2019). Depression and anxiety detection through the Closed-Loop method using DASS-21. TELKOMNIKA (Telecommunication Computing Electronics and Control), 17(4), 2087-2097.
Kim, J., Lee, J., Park, E., & Han, J. (2020). A deep learning model for detecting mental illness from user content on social media. Scientific reports, 10(1), 11846.
Firoz, N., Beresteneva, O. G., Vladimirovich, A. S., Tahsin, M. S., & Tafannum, F. (2023, February). Automated Text-based Depression Detection using Hybrid ConvLSTM and Bi-LSTM Model. In 2023 Third International Conference on Artificial Intelligence and Smart Energy (ICAIS) (pp. 734-740). IEEE.
Alqazzaz, A., Quasim, M. T., Alshahrani, M. M., Alrashdi, I., & Khan, M. A. (2023). A Deep Learning Model to Analyse Social-Cyber Psychological Problems in Youth. Computer Systems Science and Engineering, 46(1), 551-562.
Gupta, S., Singh, A., & Ranjan, J. (2023). Multimodal, multiview and multitasking depression detection framework endorsed with auxiliary sentiment polarity and emotion detection. International Journal of System Assurance Engineering and Management, 1-16.
Yu, Y., Li, Q., & Liu, X. (2023). Automatic anxiety recognition method based on microblog text analysis. Frontiers in Public Health, 11, 1080013.
Lecrubier, Y. (2001). The burden of depression and anxiety in general medicine. Journal of clinical psychiatry, 62, 4-11.
Bzdok, D., & Meyer-Lindenberg, A. (2018). Machine learning for precision psychiatry: opportunities and challenges. Biological Psychiatry: Cognitive Neuroscience and Neuroimaging, 3(3), 223-230.
Wang, S. H., Ding, Y., Zhao, W., Huang, Y. H., Perkins, R., Zou, W., & Chen, J. J. (2016). Text mining for identifying topics in the literatures about adolescent substance use and depression. BMC public health, 16, 1-8.
Babu, N. V., & Kanaga, E. G. M. (2022). Sentiment analysis in social media data for depression detection using artificial intelligence: a review. SN Computer Science, 3, 1-20.
Ortega-Mendoza, R. M., Hernández-Farías, D. I., Montes-y-Gómez, M., & Villaseñor-Pineda, L. (2022). Revealing traces of depression through personal statements analysis in social media. Artificial Intelligence in Medicine, 123, 102202.
Khan, U., Khan, S., Rizwan, A., Atteia, G., Jamjoom, M. M., & Samee, N. A. (2022). Aggression Detection in Social Media from Textual Data Using Deep Learning Models. Applied Sciences, 12(10), 5083.
Kabir, M., Ahmed, T., Hasan, M. B., Laskar, M. T. R., Joarder, T. K., Mahmud, H., & Hasan, K. (2023). DEPTWEET: A typology for social media texts to detect depression severities. Computers in Human Behavior, 139, 107503.
Calvo, R. A., Milne, D. N., Hussain, M. S., & Christensen, H. (2017). Natural language processing in mental health applications using non-clinical texts. Natural Language Engineering, 23(5), 649-685.
Le Glaz, A., Haralambous, Y., Kim-Dufor, D. H., Lenca, P., Billot, R., Ryan, T. C., ... & Lemey, C. (2021). Machine learning and natural language processing in mental health: systematic review. Journal of Medical Internet Research, 23(5), e15708.
Yu, Y., Li, Q., & Liu, X. (2023). Automatic anxiety recognition method based on microblog text analysis. Frontiers in Public Health, 11, 1080013.
Santos, W. R. D., de Oliveira, R. L., & Paraboni, I. (2023). SetembroBR: a social media corpus for depression and anxiety disorder prediction. Language Resources and Evaluation, 1-28.
Meshram, P., & Rambola, R. K. (2023). Diagnosis of depression level using multimodal approaches using deep learning techniques with multiple selective features. Expert Systems, 40(4), e12933.
Jiang, Y., Zhang, Z., & Sun, X. (2022, August). MMDA: A Multimodal Dataset for Depression and Anxiety Detection. In International Conference on Pattern Recognition (pp. 691-702). Cham: Springer Nature Switzerland.
Solieman, H., & Pustozerov, E. A. (2021, January). The detection of depression using multimodal models based on text and voice quality features. In 2021 IEEE Conference of Russian Young Researchers in Electrical and Electronic Engineering (ElConRus) (pp. 1843-1848). IEEE.
Munoz, S., & Iglesias, C. A. (2022). A text classification approach to detect psychological stress combining a lexicon-based feature framework with distributional representations. Information Processing & Management, 59(5), 103011.
Li, W., Gao, S., Zhou, H., Huang, Z., Zhang, K., & Li, W. (2019, December). The automatic text classification method based on BERT and feature union. In 2019 IEEE 25th International Conference on Parallel and Distributed Systems (ICPADS) (pp. 774-777). IEEE.
Singh, K. N., Devi, S. D., Devi, H. M., & Mahanta, A. K. (2022). A novel approach for dimension reduction using word embedding: An enhanced text classification approach. International Journal of Information Management Data Insights, 2(1), 100061.
Zhang, J., Liu, F. A., Xu, W., & Yu, H. (2019). Feature fusion text classification model combining CNN and BiGRU with multi-attention mechanism. Future Internet, 11(11), 237.
Abiodun, E. O., Alabdulatif, A., Abiodun, O. I., Alawida, M., Alabdulatif, A., & Alkhawaldeh, R. S. (2021). A systematic review of emerging feature selection optimization methods for optimal text classification: the present state and prospective opportunities. Neural Computing and Applications, 33(22), 15091-15118.
Eaton, N. R. (2020). Measurement and mental health disparities: Psychopathology classification and identity assessment. Personality and mental health, 14(1), 76-87.
Seiferth, C., Vogel, L., Aas, B., Brandhorst, I., Carlbring, P., Conzelmann, A., ... & Löchner, J. (2023). How to e-mental health: a guideline for researchers and practitioners using digital technology in the context of mental health. Nature mental health, 1(8), 542-554.
Wu, A., Scult, M. A., Barnes, E. D., Betancourt, J. A., Falk, A., & Gunning, F. M. (2021). Smartphone apps for depression and anxiety: a systematic review and meta-analysis of techniques to increase engagement. NPJ digital medicine, 4(1), 20.
Lattie, E. G., Stiles-Shields, C., & Graham, A. K. (2022). An overview of and recommendations for more accessible digital mental health services. Nature Reviews Psychology, 1(2), 87-100.

Download PDF

Version 1

posted

You are reading this latest preprint version

Review and Content Analysis of Textual Expressions as a Marker for Depressive and Anxiety Disorders (DAD) Detection Using Machine Learning

Status:

Version 1

Abstract

Background

Methods

Results

Conclusions

Figures

1. Introduction

2. Materials and Methods

2.1 Research Design and Article Selection

2.2 Data Extraction and Loading

2.3 Data Synthesis

3. Bibliometric Analysis and Mapping

3.1. Scientific Production in the Field

4. Review of Literature on DAD Detection form Text Analysis

4.1 Modalities, Data Sources, ML Techniques, and Other Aspects in DAD Detection

4.2 Multimodal Textual Data for DAD Detection Tasks

5. Discussion

6. Conclusions

Declaration

References

Status:

Version 1