Depression Classification Based on Automatic Ontology Generation and Natural Language Processing

Depression is one of the most prevalent and debilitating mental health conditions globally, affecting individuals’ mental, physical, and social well-being. Despite its widespread impact, many individuals struggle to access or seek treatment due to various barriers, including stigma and limited resources. In response, online platforms have emerged as crucial spaces for anonymous discussion and support. A recent study published in Frontiers in Psychology explores a novel approach to identifying and understanding depression within specific communities by leveraging natural language processing (NLP) and automatic ontology generation, particularly focusing on religious individuals.

Unveiling Depression in Digital Discourse: A New Frontier in Mental Health Research

A groundbreaking study has introduced an innovative method for classifying depression by analyzing text from social media platforms, with a particular focus on religious communities. This research, published in Frontiers in Psychology, utilized advanced semantic techniques, including comparative analysis of texts through ontologies, to understand how depression manifests in the online discourse of Christians. The study also aimed to develop a natural language processing (NLP) tool capable of classifying depression to generate the necessary data for building comprehensive ontologies.

The research team meticulously analyzed text-based data from online forums and social media, specifically targeting discussions related to depression within Christian communities. By comparing the semantic structures of texts from depressed Christians, non-depressed Christians, and broader groups of depressed and non-depressed individuals, the study sought to uncover unique patterns and insights. The findings revealed significant differences in how depression is discussed and experienced across these diverse groups, highlighting the influence of religious beliefs and community norms.

Background: The Growing Role of Digital Platforms in Mental Health

Depression is recognized as a leading cause of disability worldwide, impacting millions of lives. While religious involvement has often been linked to improved mental health by providing social support and a sense of purpose, stigma within some religious communities can paradoxically exacerbate feelings of depression and deter individuals from seeking help. The anonymity offered by online platforms like Reddit has proven invaluable, allowing individuals to share their experiences without fear of judgment. This has opened new avenues for researchers to study mental health trends and identify potential indicators of depression in real-time.

The current study builds upon previous research demonstrating the efficacy of NLP techniques in analyzing social media for signs of depression. By employing ontologies—structured representations of knowledge—researchers can capture the nuanced semantic relationships between concepts related to depression, such as symptoms, emotional states, and cognitive patterns. This method offers a more profound understanding than traditional keyword analysis.

Methodology: A Multi-faceted Approach to Data Analysis

The research methodology involved several key stages:

Text Selection and Manual Classification

The study meticulously gathered text data from Reddit and public forums, focusing on users who explicitly identified as Christian and reported a depression diagnosis. A control group of general Christians without reported diagnoses was also established. The selection criteria ensured participants were at least 18 years old and had active forum accounts, while excluding individuals with other significant mental or physical disorders. This rigorous selection process yielded a corpus of 93 posts, manually classified by psychologists into two main groups: 36 posts from Christians with depression and 57 from the control group.

For broader comparisons, texts from general depressed individuals (not necessarily Christian) and general non-depressed individuals were sourced from Reddit. These large datasets, exceeding the processing capacity of the ontology generation tool, were segmented and processed in smaller batches.

NLP Training and AI Classification

To effectively analyze the large volume of text data from Christian forums, an NLP classifier was developed. This classifier was trained on a dataset of 93 posts, manually labeled by psychologists to distinguish between depressed and non-depressed individuals based on physician-reported diagnoses. The preprocessing pipeline involved removing digits and unknown characters, eliminating common English stop words, and lemmatizing words using NLTK. Text data was then vectorized using TF-IDF for machine learning algorithms.

Several baseline classifiers were evaluated, with the Linear classifier using SGD training demonstrating the highest F1 score of 85%. Further optimization using randomized search for hyperparameters resulted in an enhanced SGD classifier achieving an F1 score of 0.94 during training and 86% on the test set. This trained classifier was subsequently used to identify 205 posts indicating depression from an unlabeled set of 255 posts within Christian forums.

Automated Generation of Ontologies

The core of the study involved the automated generation of ontologies using the Text2Onto tool. This process was conducted in three steps:

1. Corpus Construction: Assembling text data for each of the four user categories.
2. Concept Extraction: Employing Text2Onto to extract relevant concepts with associated relevance scores. Given Text2Onto’s limitation of processing approximately 15,000 words at a time, larger corpora were divided and processed in parts, with results subsequently merged. Concepts with relevance scores above a dynamically determined threshold were retained.
3. Ontology Merging: For the "general" cases, a Java parser was implemented to combine concepts from sub-corpora, calculating a final relevance score as the sum of individual scores.

This process generated distinct ontologies for depressed Christians, non-depressed Christians, general depressed individuals, and general non-depressed individuals.

Construction of Ontology from Existing Literature

To provide a benchmark, a literature-based ontology was also constructed. A systematic search of scientific databases identified five relevant peer-reviewed articles published between 2015 and 2019 focusing on depression ontologies. Four of these ontologies, which addressed adolescent and general depression and aimed for diagnosis support or social media analysis, were merged using Protégé Web Ontology software. The resulting merged ontology comprised 21 classes, 43 subclasses, and 58 individuals, encompassing concepts like risk factors, symptoms, and interventions.

Results: Unpacking Semantic Similarities and Differences

The study employed a quantitative measure of ontology similarity based on the number of common terms: Similarity (Ont1, Ont2) = Nr-Com-Term / (Nr-terms-Ont1 + Nr-terms-Ont2). This concept-based similarity was chosen due to the nature of automatically generated ontologies, which excel at concept extraction but may produce less consistent relational structures.

Quantitative Similarity Analysis

Similarity was computed across ontologies, with results presented for top 800, 500, and 200 concepts, showing robust and consistent similarity measures regardless of the number of concepts considered. Key findings included:

• Highest Similarity: The greatest similarity was observed between general depressed individuals and non-depressed individuals (approximately 0.37). This suggests a shared discourse around common human experiences of sorrow and stress, even in the absence of a clinical diagnosis.
• Lower Similarity: Religious individuals exhibited lower similarity scores compared to non-religious groups. This divergence, approximately 0.18 for non-depressed individuals and 0.21 for depressed individuals, may indicate that religious beliefs and practices influence the language used to express or cope with depression.
• Intra-group Christian Similarity: The similarity between the two Christian categories (depressed and non-depressed) was approximately 0.18, suggesting distinct patterns in their online discourse.

Qualitative Analysis of Key Concepts

A qualitative examination of the most relevant terms provided deeper insights:

• General Depressed vs. Non-Depressed: Both groups shared common concepts such as "people," "game," "year," "shit," and "someone." The prominence of "game" suggests its role in contemporary life for both individuals experiencing depression and those not. The recurring use of profanity like "shit" indicates a common linguistic pattern across these groups.
• Depressed Christians vs. Literature-Based Ontology: A notable discrepancy emerged between the ontology generated from depressed Christians’ texts and the literature-based ontology. While the latter highlighted terms like "depression," "emotions," and "diagnosis," these were less prominent in the user-generated texts. Researchers hypothesized this might be because individuals actively posting online may not be as severely affected as clinical populations, or they may focus less on medical terminology.
• Depressed Christians vs. Non-Depressed Christians: Religious terms such as "God," "faith," "Christians," and "church" were prominently featured in both Christian categories, underscoring their religious identity. However, depressed Christians more frequently used terms like "depression" and "anxiety," while social relationships and temporal references remained significant for both. Profanity and gaming were less prominent among Christians compared to the general depressed group.

Psychologists reviewing the classifier’s output for depressed Christians identified symptoms like sadness, impaired decision-making, physical discomfort, and a perceived need for help. Discussions often touched upon the role of faith in recovery, the challenges of misconceptions about depression within religious communities, and the stigma associated with mental illness. For non-depressed Christians, expressions of sadness or loneliness were framed within a context of acceptance and a perceived divine presence, with a strong emphasis on personal agency in managing emotional states and an explicit rejection of suicidal ideation.

Discussion and Implications: Bridging Digital Insights and Clinical Understanding

This study represents a pioneering effort to apply automated ontology generation and NLP to understand depression within a specific religious population. The findings underscore the potential of digital discourse analysis to reveal unique patterns of expression and experience related to mental health conditions.

The observed differences between religious and non-religious groups in their discussion of depression align with existing research on religion as a coping mechanism. The findings suggest that religious beliefs can shape how individuals perceive, articulate, and manage their depressive symptoms.

Limitations and Future Directions

The researchers acknowledged several limitations. The sample size for Christian participants was relatively small, and the NLP classifier, trained on Reddit and Christian forums, may not generalize perfectly to other platforms or demographics. Future research could benefit from larger, more diverse datasets to enhance the classifier’s robustness. The study’s reliance on its own classifier for text-to-ontology analysis, while focused on concept extraction, is another area for potential refinement.

Looking ahead, the integration of such systems into counseling platforms could facilitate early depression detection and support timely intervention strategies. This research opens promising avenues for leveraging AI and digital data to improve mental healthcare accessibility and effectiveness, particularly within diverse cultural and religious contexts.

Conclusion: A Step Forward in Culturally Sensitive Mental Health Analytics

By employing sophisticated NLP and ontology generation techniques, this study has provided valuable insights into the linguistic manifestations of depression among Christians and in comparison to broader populations. The research highlights the importance of considering cultural and religious factors when analyzing mental health discourse in digital spaces. As digital health technologies continue to evolve, such data-driven approaches hold significant promise for developing more personalized and effective mental health support systems.