Abstract
Detecting depression on social media is often hindered by the overwhelming amount of irrelevant daily content. We propose E2-LPS, an end-to-end framework that utilizes clinical psychiatric scales to screen for risky posts. Unlike traditional two-stage methods where screening is static, E2-LPS employs a straight-through estimator to make the screening process learnable and differentiable. This allows the final detector to provide feedback to the screener, ensuring the selection of the most clinically relevant content. Our experiments on three benchmark datasets demonstrate that E2-LPS significantly outperforms state-of-the-art baselines and large language models.
Introduction
Depression detection from social media history is a "needle in a haystack" problem. Users post frequently about mundane topics, creating a high noise-to-signal ratio for mental health monitoring. Current research typically addresses this with a two-step pipeline: a "static" screener filters posts based on keywords or frozen models, and a detector classifies the filtered subset. However, this decoupled approach prevents the screener from learning which specific nuances the detector finds most useful. We argue that screening and detection should be optimized jointly to capture subtle psychological cues.
Methods: The E2-LPS Framework
The proposed E2-LPS (End-to-End Learnable Psychiatric Scale Guided Risky Post Screening) architecture bridges the gap between screening and detection.
Psychiatric Scale-Guided Screening
To ensure clinical relevance, we derive 24 symptom templates from the Beck Depression Inventory-II (BDI-II). User posts are encoded via Sentence-BERT, and their cosine similarity to these templates determines a "risk score" for every post.
Differentiable Selection
A major challenge in end-to-end training is that "Top-K" selection is a discrete, non-differentiable operation. We overcome this using a Straight-Through Estimator (STE), which allows gradients to flow back from the detection error to the screening parameters during training.
Detection and Attention
The screened "risky" posts are processed by a Transformer-based encoder. We utilize a modified self-attention mechanism to focus interactions solely among the selected posts, refining the signal before final classification determines the likelihood of depression.
Experimental Results
We evaluated E2-LPS on three Reddit-based benchmarks: eRisk2017, eRisk2018, and SMHD. To ensure the model learns symptoms rather than explicit disclosures, we removed "anchor points" (e.g., "I was diagnosed with depression").
As shown in the results table, E2-LPS consistently achieves the highest F1-scores. On eRisk2017, it reached 0.714, surpassing the previous state-of-the-art DeCapsNet (0.699) and significantly outperforming zero-shot GPT-4.1. Furthermore, the model exhibits strong cross-dataset generalization, indicating it captures universal depressive signals rather than community-specific jargon.
Discussion and Conclusion
E2-LPS provides both high performance and interpretability. By explicitly selecting a subset of posts, the model offers a built-in explanation for its predictions, which can be reviewed by clinicians. Expert validation by board-certified psychiatrists confirmed that the posts flagged by E2-LPS possess higher clinical validity than those from automated baselines.
While promising, we emphasize that E2-LPS is a screening tool intended to assist, not replace, human experts. Ethical considerations regarding privacy and potential stigmatization remain paramount in the deployment of automated mental health monitoring systems.