Detecting Bipolar Disorder from Misdiagnosed Major Depressive Disorder with Mood-Aware Multi-Task Learning

Daeun Lee, Hyolim Jeon, Sejung Son, Chaewon Park, Ji Hyun An, Seungbae Kim, Jinyoung HanNAACL 2024 · DOI

TL;DR

This paper introduces a novel multi-task learning approach to detect the risk of Bipolar Disorder (BD) in individuals initially misdiagnosed with Major Depressive Disorder (MDD). By leveraging social media data and a new expert-validated dataset (BD-Risk), the model accurately estimates mood levels and variations, offering crucial insights for early intervention and suicide prevention.

What Problem Does This Paper Address?

Bipolar Disorder (BD) is often misdiagnosed as Major Depressive Disorder (MDD) due to overlapping symptoms, leading to delayed treatment and increased suicide risk. While early BD detection is vital, a significant challenge lies in identifying individuals who transition from MDD to BD. This transition is often missed because BD patients may not perceive manic episodes as abnormal, and they frequently seek help during depressive phases. Existing research often focuses on detecting MDD or BD in isolation, neglecting the critical aspect of diagnosis shifts. Furthermore, understanding the nuances of mood fluctuations in real-world scenarios, beyond limited clinical interactions, is crucial. This paper aims to bridge this gap by developing a model that can detect BD risk in individuals with a prior MDD diagnosis, using their social media activity to capture dynamic mood patterns.

Methods

The core of this work is a multi-task learning (MTL) model designed to simultaneously predict BD risk, estimate BD mood level, and assess mood variation. The model comprises three key components:

  1. Temporal Convolution Layer: This layer uses a pre-trained Sentence-BERT (SBERT) to generate semantic representations of individual posts. A 1D convolution then captures dynamic mood patterns by analyzing neighboring elements in a post sequence.
  2. Dynamic Mood-Aware Attention Layer: This layer employs a Transformer encoder with positional encoding to capture historical context. Crucially, it incorporates a "time-aware parameter" that fuses self-attention with a mechanism to emphasize critical mood states influencing BD risk, considering the time elapsed between posts. This allows the model to better understand the temporal dynamics of mood.
  3. Prediction Layer: This layer decodes the learned representations to output predictions for BD risk, BD mood level, and mood variation.

The MTL approach is further enhanced by uncertainty weighting, which assigns adaptive weights to each task's loss function based on its inherent uncertainty. This allows the model to learn effectively across tasks with different scales and granularities.

Datasets

The paper introduces a novel dataset called BD-Risk. This dataset is unique in that it:

  • Contains information on specific mental disorders (MDD and BD, including subtypes).
  • Includes expert-validated BD mood levels on a scale of -3 to 3.
  • Is derived from social media posts (Reddit).
  • Specifically captures users who transitioned from an MDD diagnosis to a BD diagnosis.

The BD-Risk dataset consists of 7,346 posts from 1,025 users, with labels validated by clinical experts (a psychiatrist and a clinical psychologist). This dataset is made publicly available to benefit the research community.

Key Insights and Ideas

  • Social Media as a Window into Diagnosis Shifts: The study demonstrates that social media data, particularly longitudinal analysis of posts, can effectively capture the subtle behavioral shifts indicative of a transition from MDD to BD, which might be missed in clinical settings.
  • Mood Variation is Key: Beyond just the absolute mood level, the variation in mood is a critical indicator for BD risk. The model's ability to capture this dynamic aspect significantly improves prediction accuracy.
  • Multi-Task Learning Synergy: Jointly learning BD risk, mood level, and mood variation enhances performance across all tasks. This suggests that these tasks are not independent and benefit from shared representations.
  • Interpretability for Clinical Insight: The dynamic mood-aware attention mechanism provides interpretable insights into which posts and mood states most strongly influence the prediction of BD risk, potentially aiding clinicians in understanding patient trajectories.

Key Results

  • The proposed multi-task learning model significantly outperforms state-of-the-art baselines in predicting BD risk, achieving an F1 score of 57.8%.
  • For BD mood level prediction, the model achieved a Mean Absolute Error (MAE) of 0.701, demonstrating high accuracy.
  • The ablation study showed that the Dynamic Mood-Aware Attention layer was crucial for improving BD risk prediction performance, highlighting the importance of temporal context.
  • Jointly learning mood level and variation tasks (MTL) yielded superior performance compared to single-task learning (STL).
  • The average diagnosis shift duration from MDD to BD was found to be approximately 563.17 days, indicating a substantial period where early intervention might be possible.

Limitations and Future Work

  • Subjectivity of Social Media Data: Evaluating moods on social media is inherently subjective, and there's a possibility of including unreliable data if users misunderstand their diagnoses.
  • Generalizability: While the model shows promise, its effectiveness in diverse clinical settings and with different demographics needs further investigation.
  • Future Work: The authors suggest incorporating additional features like intensity variations or sentiment analysis to enrich the model's understanding of mood swing patterns. They also plan to improve the sequence model to better capture temporal dynamics and their impact on predictions. Further research into the trade-off between predicting mood level and variation is also suggested.

Summary updated

BibTeX
@inproceedings{lee2024detecting,
  author = {Daeun Lee and Hyolim Jeon and Sejung Son and Chaewon Park and Ji Hyun An and Seungbae Kim and Jinyoung Han},
  title = {{Detecting Bipolar Disorder from Misdiagnosed Major Depressive Disorder with Mood-Aware Multi-Task Learning}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2024},
  doi = {10.18653/v1/2024.naacl-long.278},
}
Figure