PaperlayerView exact revision on arXiv

Source-linked AI summary

Semi-Supervised Approach to Monitoring Clinical Depressive Symptoms in Social Media

Amir Hossein Yazdavar, Hussein S. Al-Olimat, Monireh Ebrahimi, Goonmeet Bajaj, Tanvi Banerjee, Krishnaprasad Thirunarayan, Jyotishman Pathak, Amit Sheth

arXiv:1710.05429v1cs.CL

TL;DR

The paper addresses whether clinical depressive symptoms can be detected unobtrusively from Twitter rather than relying only on questionnaires. It develops a semi-supervised, symptom-guided topic model for longitudinal tweet analysis and reports 68% accuracy and 72% precision for capturing symptoms per user.

  • Problem

    Questionnaire-based depression studies can suffer from sampling and reporting limitations, motivating unobtrusive detection of clinical symptoms in social-media content.

  • Method

    The study combines a depression-symptom lexicon with semi-supervised topic modeling to extract and monitor symptom-related topics and words over time.

  • Results

    68% accuracy and 72% precision were reported for capturing depression symptoms per user over a time interval.

  • Takeaways & Limitations

    The approach complements questionnaire-driven diagnostic tools by monitoring depression symptoms continuously and unobtrusively from tweets.

Abstract

from arXiv · show

With the rise of social media, millions of people are routinely expressing their moods, feelings, and daily struggles with mental health issues on social media platforms like Twitter. Unlike traditional observational cohort studies conducted through questionnaires and self-reported surveys, we explore the reliable detection of clinical depression from tweets obtained unobtrusively. Based on the analysis of tweets crawled from users with self-reported depressive symptoms in their Twitter profiles, we demonstrate the potential for detecting clinical depression symptoms which emulate the PHQ-9 questionnaire clinicians use today. Our study uses a semi-supervised statistical model to evaluate how the duration of these symptoms and their expression on Twitter (in terms of word usage patterns and topical preferences) align with the medical findings reported via the PHQ-9. Our proactive and automatic screening tool is able to identify clinical depressive symptoms with an accuracy of 68% and precision of 72%.

I. INTRODUCTION

The study targets limitations of questionnaire-based and existing social-media approaches by modeling depressive symptoms over time from tweets. It proposes a semi-supervised model guided by clinical symptoms and evaluates whether social-media language can support automatic depression screening.

  • Motivation: Questionnaire-based depression detection can suffer from underrepresentation, sampling bias, incomplete information, reporting bias, and delays between data collection and findings.Twitter may provide real-time expressions of users’ feelings, behaviors, and mental-health experiences.
  • Limitations of prior approaches: Existing lexicon-based methods have low recall and depend heavily on lexicon quality, while supervised methods require labor-intensive annotation of large datasets.The paper also distinguishes clinical depression from short-term feelings of being down because diagnosis involves symptoms lasting for a fixed period.
  • Proposed contribution: The study develops a statistical model that extracts, categorizes, and unobtrusively monitors depressive symptoms by modeling social-media content as topics evolving over time.The dataset contains 23 million tweets from more than 45,000 Twitter users who self-reported depressive symptoms in their profiles.
  • Research questions: The research asks whether social-media text can reliably capture clinical-depression symptoms over time and whether depressed users share underlying themes.These questions frame the goal of a proactive, automatic screening tool and the search for common thematic patterns.
  • Proposed contribution: The approach combines a PHQ-9-based depression lexicon with data-driven discovery of latent topics from users’ tweets.Informative lexicon terms are ranked for each user and used as seeds for discovering symptom-related topics.
  • Contributions: The study contributes a depression-symptom lexicon and a semi-supervised model for continuous temporal analysis of individual tweets.Empirical evaluations report superior learned-topic quality compared with five baselines.

II. RELATED WORK

Prior work detects depression in social media using linguistic, emotional, topical, network, engagement, and lexicon-based signals. These approaches provide useful indicators but can overlook symptom categories or face ambiguity in depression-related language.

  • Machine-learning approaches: Earlier studies characterize depression using language, emotion, style, ego-network, and user-engagement features to predict depression in posts or individuals.These methods include retrospective analyses of tweets from depressed and non-depressed users.
  • Lexicon-based approaches: Dictionary-based methods may assign an overall depression score by counting matched indicators without separating distinct symptom categories.Other work examined tweets containing the keyword “depression” and reported evidence of mental-health disclosures and treatment updates.
  • Illustrative examples: Table I presents sample tweets together with their associated symptoms.The table connects social-media text examples to symptom categories used in the study.
  • Lexicon limitations: Depression-related words can be ambiguous because terms such as “depression” may refer to economic, historical, meteorological, or clinical concepts.This ambiguity complicates automatic screening based only on surface word matches.

III. PROPOSED APPROACH

The proposed approach emulates the PHQ-9 by detecting clinical-depression symptoms in Twitter data through complementary bottom-up and hybrid top-down processing. It uses related word clusters alone or guides extraction with a depression lexicon.

  • Clinical framing: Clinical depression is operationalized as a set of symptoms that persist over a fixed period, following the DSM framing.The PHQ-9 provides the questionnaire framework that the approach emulates.
  • Bottom-up processing: The bottom-up approach uses distributional semantics to uncover depressive symptoms through related word clusters in users’ tweets.It seeks symptom signals without initially imposing a predefined symptom lexicon.
  • Hybrid processing: The hybrid approach combines bottom-up discovery with top-down processing that uses lexicon terms to guide symptom extraction from tweets.This design links clinically articulated symptoms with patterns discovered from social-media language.

A. Bottom-up processing:LDA

The bottom-up component applies LDA to user tweets to discover latent topics representing depressive symptoms. Because ordinary LDA can produce broad or poorly interpretable topics, the proposed model supervises topic formation with symptom-related seed terms.

  • LDA foundation: LDA represents each document as a mixture of latent topics, with each topic defined as a distribution of co-occurring words.The study applies LDA to extract topics discussed by users.
  • Limitations of standard LDA: Traditional LDA may produce topics that are insufficiently granular for depressive symptoms and may correlate poorly with human judgments.This motivates incorporating domain knowledge into topic discovery.
  • Seeded topic modeling: The seeding model supervises topics at the token level by restricting relevant tokens to specified topics rather than measuring distributions over a predefined term list.This token-level constraint distinguishes the model from the cited alternative.

B. Hybrid processing: Proposed ssToT Model

The ssToT model combines a clinically grounded depression lexicon with supervised topic-seeding strategies to identify symptom categories in social-media language. It addresses varied phrasing, ambiguous words, and the need to monitor symptoms over time.

  • Hybrid processing: The model uses PHQ-9 symptom categories and a clinician-refined lexicon as predefined seed clusters for topic discovery.The lexicon incorporates synonyms and colloquial expressions to represent nine clinical symptom categories.
  • Challenges: Social-media language creates challenges because users express the same symptom with diverse metaphors and use words whose meanings depend on context.Examples include multiple expressions of low energy and different senses of “cut.”
  • Hybrid processing: The algorithm generates personalized seed terms for each user from the lexicon and retains ambiguous terms only when their enclosing sentence has negative sentiment.This approach selects user-specific expressions while using sentiment polarity for word-sense disambiguation.
  • Temporal monitoring: Clinical depression requires symptoms to persist for weeks, making temporal monitoring necessary rather than relying on transient expressions of feeling down.The framework therefore analyzes social-media content across time intervals.

IV. ALGORITHM

The ssToT algorithm divides each user’s timestamped tweets into time buckets and models depression symptoms as seeded latent topics. Gibbs sampling estimates symptom and word distributions while constraining predefined seeds to their designated topics.

  • IV. ALGORITHM: The framework continuously monitors each user’s tweets by dividing them into buckets defined by a time interval of d days.The generative process operates over each user’s collection of timestamped tweet buckets.
  • IV. ALGORITHM: ssToT generates nine symptom topics, draws topic distributions for buckets, and assigns each word to a symptom topic and its word distribution.Algorithm 1 represents symptoms as topics and words as emissions from those topics.
  • IV. ALGORITHM: Gibbs sampling estimates topic-word and bucket-topic distributions while incorporating prior knowledge through predefined seed assignments.The model’s distinctive design assigns each seed term to only one symptom topic during inference.

V. EXPERIMENTAL RESULTS

The study collected and anonymized Twitter data from users who self-declared depression and from randomly sampled undeclared users. After filtering and preprocessing, the authors formed a 4,000-user evaluation subset and normalized lexicon phrases for topic modeling.

  • Dataset: The dataset began with 45,000 self-declared depressed users and 2,000 randomly collected undeclared users.Users with fewer than 100 tweets were removed before the main tweet collection was summarized.
  • Dataset: The evaluation subset contained 2,000 self-reported users and 2,000 randomly selected users without depression terms in their profiles.The authors denote this 4,000-user subset by U.
  • Preprocessing: Preprocessing converts lexicon phrases into underscore-separated tokens and applies platform filtering, normalization, stopword removal, lemmatization, spelling correction, and repetition handling.Phrase replacement is necessary because topic modeling operates at the word level while many depression seeds are multiword expressions.

A. Qualitative Results

Qualitative results indicate that ssToT produces more interpretable symptom topics than LDA and captures social-media-specific expressions. The analysis also reveals temporal symptom trends and community-level themes involving relationships and academic stress.

  • A. Qualitative Results: ssToT constrains seed terms to symptom categories, producing more relevant and interpretable topics than LDA’s mixed symptom clusters.LDA topics combine terms from sleep, eating, and suicidal-thought categories, whereas ssToT keeps these categories more distinct.
  • A. Qualitative Results: ssToT captures acronyms and expressive interjections used in social media within their related symptom categories.Examples include “ugw,” “mfp,” “idec,” “argh,” and “aw.”
  • A. Qualitative Results: Table II compares sample symptom topics learned by ssToT and LDA using associated coherency scores.
  • A. Qualitative Results: Community-level themes include family and friend problems, need for support, and school or academic stress, extending beyond the PHQ-9 symptom categories.Illustrative topics contain terms related to relationships, school, exams, classmates, and assignments.
  • A. Qualitative Results: The model visualizes when and for how long specific depressive symptoms occur by tracking symptom topics across time buckets.Human annotators manually assessed buckets to check the validity of topical trends.

VI. QUANTITATIVE RESULTS

The study evaluates ssToT against topic-modeling and clustering baselines using coherence measures designed to assess semantic interpretability. Across UMass, UCI, and NPMI, ssToT learns more coherent topics associated with depressive symptoms.

  • Coherence Measures: Topic coherence measures assess semantic similarity among high-scoring topic words to distinguish interpretable topics from statistical artifacts.The study considers intrinsic and extrinsic evaluation measures.
  • Coherence Measures: UMass measures word co-occurrence within the training corpus, whereas UCI uses Pointwise Mutual Information from an external dataset such as Wikipedia.These metrics represent intrinsic and extrinsic coherence evaluation, respectively.
  • Quantitative comparison: ssToT topics are more coherent than LDA topics across UMass, UCI, and NPMI measures.The evaluation uses tweets for UMass and Wikipedia as the external corpus for UCI and NPMI.
  • Baselines: The study compares ssToT with existing unsupervised and semi-supervised approaches, including k-means, LSA, LDA, and BTM.The baselines represent clustering, distributional-semantic, Bayesian topic-modeling, and related approaches.
  • Implementation: Fifteen topics are selected because this setting has the lowest perplexity while providing suitable granularity for the task.The models use 80% of the data for training and 20% for testing in perplexity selection.

VII. SYMPTOM PREDICTION (MULTI-LABEL CLASSIFICATION)

The ssToT model is evaluated as a multi-label classifier that assigns depressive-symptom labels to 14-day tweet buckets. It achieves moderate overall performance, competes with supervised baselines, and avoids labeled data during training, while symptom ambiguity limits some categories.

  • Evaluation dataset: The evaluation uses 10,400 tweets grouped into 192 buckets, with each bucket containing a user’s tweets from a 14-day span.The ground-truth data includes randomly sampled self-reported depressed and random users.
  • Model evaluation: 0.68 accuracy across nine depressive symptoms yields 0.72 average precision against human judgments.The model assigns symptom sets to tweet buckets without using labeled data during training.
  • Per-symptom results: 0.90 F-Measure is achieved for “Lack of Interest,” while “Concentration Problems” produces the weakest result.Descriptive or metaphorical tweets and ambiguous symptom wording make some categories harder to classify.
  • Baseline comparison: ssToT improves F-score over supervised approaches for five of nine symptoms and provides better averaged accuracy.The supervised comparisons use binary relevance and classifier-chain methods with Multinomial Naive Bayes and SVM models.
  • Limitations: The study cannot assess users who post little content or avoid publicly revealing depressive symptoms, and it measures presence, duration, and frequency rather than severity.The sample may also be biased toward more severely depressed individuals who publicly express depression.

VIII. CONCLUSION AND FUTURE WORK

The study combines lexicon-based and semi-supervised topic modeling to extract and monitor depression symptoms from social-media text. It reports 68% accuracy and 72% precision for capturing symptoms per user over time.

  • The model combines lexicon-based analysis with semi-supervised topic modeling to extract per-user depression-related topic distributions.It analyzes tweets across different time windows and derives symptom-indicating word distributions.
  • The approach supports continuous and unobtrusive monitoring as a complement to questionnaire-driven diagnostic tools.
  • 68% accuracy and 72% precision were achieved for capturing depression symptoms per user over a time interval.
Loading 1710.05429v1…