Abstract
Early detection of depression is a critical public health challenge, as many individuals remain undiagnosed due to social stigma or limited healthcare access. This paper proposes a novel screening framework that treats a user’s social media history as a "bag" of multimodal instances through Multiple Instance Learning (MIL). By leveraging a Transformer architecture, the model captures temporal dependencies between text and images, providing both improved classification accuracy and clinical explainability.
Introduction
Depression affects over 300 million people globally, yet a significant portion of the population does not receive timely diagnosis. Social media platforms offer a unique window into daily behaviors and psychological states, providing a potential avenue for early screening. However, existing machine learning (ML) approaches often fail to align with clinical diagnostic standards, such as the DSM-5, which require observing symptoms over specific timeframes. This research addresses these gaps by introducing a multimodal MIL-Transformer framework designed to model the evolution of user behavior over time.
Challenges in Automated Depression Screening
Current models for depression detection face three primary obstacles:
- Temporal Context Neglect: Clinical diagnosis relies on the persistence of symptoms over time (e.g., a two-week period). Most ML models treat posts as isolated events or aggregate them into a static representation, losing the temporal dynamics essential for accurate assessment.
- Weak Supervision: While researchers may have labels for a user (depressed vs. non-depressed), they rarely have labels for individual posts. Treating every post from a depressed user as a positive instance introduces significant noise, as not every interaction reflects a depressive state.
- The "Black Box" Problem: For a screening tool to be clinically viable, it must be interpretable. Clinicians need to understand which specific behaviors or time periods led to a model's classification.
Methodology: The MIL-Transformer Framework
To overcome these hurdles, the authors propose a framework based on Multiple Instance Learning (MIL) and Transformer architectures.
Multiple Instance Learning (MIL)
In this framework, a user is represented as a "bag" ($P^i$) containing multiple "instances" (individual posts $p_k$). Instead of forcing a label onto every post, the model learns to predict the label for the entire bag. This approach acknowledges the reality of social media behavior: the "signal" of depression is often buried within a "noise" of everyday, non-symptomatic posts.
Multimodal Transformer Architecture
The core of the model is a Transformer designed to capture dependencies between posts across time.
- Feature Extraction: The model utilizes ELMo for generating text embeddings and ResNet34 for extracting image features.
- Cross-Modality Attention: By feeding one modality into the encoder and another into the decoder, the model uses an attention mechanism to weigh specific posts or combinations of text and images. This attention mechanism serves a dual purpose: it improves predictive performance and provides "restricted explanation" by highlighting the most indicative posts for clinicians to review.
Evaluation and the Depression CheckList
The authors utilize an Instagram Student Dataset, which includes both image data and text captions. Recognizing that standard NLP metrics (like accuracy or F1-score) may not capture the psycholinguistic nuances of mental health, the authors propose the DEPRESSION CHECKLIST. Inspired by the "CheckList" methodology in NLP, this suite of behavioral tests evaluates the model’s ability to recognize specific markers associated with depression, ensuring the model is learning clinically relevant patterns rather than spurious correlations.
Preliminary Results and Future Work
In preliminary experiments using standard neural network classifiers on multimodal data (IMAGE+TEXT), the authors achieved an F-measure of 0.79. The proposed MIL-Transformer framework serves as a roadmap to exceed this baseline by better exploiting the hierarchical and temporal nature of social media data.
While promising, the research faces challenges regarding data scarcity, as high-quality "gold standard" annotated datasets in the mental health domain are rare. Future work will focus on validating the approach on external multimodal benchmarks and exploring the generalizability of the MIL-Transformer framework to other mental health conditions where temporal sequences and weak supervision are prevalent.
Summary updated