[Paper Review] Detection of the Prodromal Phase of Bipolar Disorder from Psychological and Phonological Aspects in Social Media

Yen-Hao Huang, Lin‐Hung Wei+1 more|arXiv (Cornell University)|Dec 26, 2017

Mental Health via Writing3 references18 citations

TL;DR

This study proposes a novel method to detect the prodromal phase of bipolar disorder using psychological and phonological features from social media, leveraging a time-specific subconscious crowdsourcing approach to collect diagnosis-timed user data. The model achieves over 91% precision in classifying BD onset using only text-based features, including phonological energy patterns, demonstrating strong potential for early, low-cost mental health screening in primary care settings.

ABSTRACT

Seven out of ten people with bipolar disorder are initially misdiagnosed and thirty percent of individuals with bipolar disorder will commit suicide. Identifying the early phases of the disorder is one of the key components for reducing the full development of the disorder. In this study, we aim at leveraging the data from social media to design predictive models, which utilize the psychological and phonological features, to determine the onset period of bipolar disorder and provide insights on its prodrome. This study makes these discoveries possible by employing a novel data collection process, coined as Time-specific Subconscious Crowdsourcing, which helps collect a reliable dataset that supplements diagnosis information from people suffering from bipolar disorder. Our experimental results demonstrate that the proposed models could greatly contribute to the regular assessments of people with bipolar disorder, which is important in the primary care setting.

Motivation & Objective

• To identify early, pre-manic symptoms of bipolar disorder (BD) before full onset, addressing high misdiagnosis and suicide rates.
• To develop a scalable, low-cost method for regular mental health assessment using publicly shared social media content.
• To introduce a new data collection method—time-specific subconscious crowdsourcing—that captures users' self-reported diagnosis timing.
• To integrate phonological features (e.g., word energy) into text-based models for improved BD onset prediction.
• To differentiate prodromal from acute symptoms using linguistic and behavioral patterns in social media.

Proposed method

• Time-specific subconscious crowdsourcing: identifies BD users on Twitter who explicitly state their diagnosis and date, enabling time-anchored data collection.
• Constructs a classifier using psychological features (e.g., sentiment, emotion, word choice) and novel phonological features based on word-level energy from speech prosody.
• Applies a sliding window approach over time to segment user data and predict BD onset probability per time frame.
• Employs a Conditional Random Field (CRF)-based language model (CLM) to analyze linguistic style changes over time, identifying prodromal patterns.
• Uses a prodrome filter to detect initial prodromal periods by analyzing shifts in linguistic and phonological features.
• Validates results via a blind evaluation with 8 psychologists, using a relaxed consensus criterion to account for diagnostic ambiguity.

Experimental results

Research questions

• RQ1 Can psychological and phonological features from social media texts reliably predict the prodromal phase of bipolar disorder?
• RQ2 How does the temporal evolution of linguistic features reflect the transition into a BD episode?
• RQ3 Can phonological features—derived from speech prosody—improve text-only BD detection models?
• RQ4 To what extent can time-specific subconscious crowdsourcing yield reliable, diagnosis-timed data for mental health research?
• RQ5 How does the model perform in distinguishing prodromal from acute or non-BD states in real-world social media data?

Key findings

• The model achieved an average precision of 0.934 when predicting BD onset over a 2-month time frame, improving from 0.809 at 12 months.
• Phonological features alone, when combined with a text-only ensemble model, enabled over 91% precision in detecting BD onset.
• The CLM analysis revealed that word expression patterns shift meaningfully before BD onset, with increased use of negative and overwhelmed-related language.
• The prodrome filter successfully identified initial prodromal periods in user timelines, as visualized in Figure 6.
• Psychologist evaluations showed 75% agreement rate for classifying regular users as not having BD symptoms, and 45% agreement for BD onset cases, indicating diagnostic ambiguity but strong model performance under relaxed criteria.
• The study confirms that linguistic and phonological features can reflect mental state changes, enabling regular, non-invasive monitoring of at-risk individuals.