JMIR Publications

ABSTRACT

Background:

Depression has become a major global public health challenge, and early intervention is critical for improving patient outcomes. Current depression detection techniques based on social media data (Traditional Risk Detection, TRD) rely heavily on users’ complete historical information, which cannot meet the timeliness requirements of early intervention. This underscores the need for Early Risk Detection (ERD) methods emphasizing early-stage and real-time warning. However, existing ERD studies have notable limitations: (1) they overlook sleep–wake rhythms hidden in posting timestamps, missing vital warning signals; and (2) they depend on static templates or resource-intensive sequence models, resulting in limited interpretability and inefficient use of early data, ultimately constraining their clinical applicability for early intervention.

Objective:

To address these issues, this study aims to develop an efficient, reliable, and interpretable ERD model. The core objectives are: to extract sleep–wake rhythm features from posting timestamps, thereby to enrich the feature dimensions for risk warning; to leverage large language models (LLM) for improved text filtering precision and depression-related factor analysis; and ultimately to achieve accurate early detection of depression, supporting early clinical intervention.

Methods:

We propose the Monitoring of Individual Nighttime Dynamics and LLM Analysis (MIND) model, which integrates two key innovations: (1)Sleep Dynamics: posting timestamps are transformed into sleep–wake rhythms, analyzing fluctuations in posting frequency and time to derive sleep-related features, thereby compensating for the limitations of text-only approaches. (2)LLM Depression Profiler: LLM is used for dynamic text filtering, automatically removing irrelevant noise and focusing on potential depression-related cues. Based on LLM semantic understanding, latent depression risk factors are identified, enhancing interpretability for clinical treatment and robustness to noise.

Results:

Experiments on the eRisk2017 benchmark dataset demonstrated that MIND significantly outperformed existing baseline models in early detection sensitivity, specificity, and accuracy. By combining sleep features with text analysis, the model achieved interpretable, traceable predictions that can support clinical treatment. Relevant experimental code is publicly available

Conclusions:

The MIND model innovatively combines sleep–wake rhythm features with LLM-based text analysis, addressing the challenges of poor interpretability and inefficient use of early-stage data in existing ERD methods. It significantly enhances early detection performance, offering a new paradigm for applying social media data in ERD task, thereby enabling earlier intervention and reducing the public health burden of depression.