CTMC-LSTM: A Markov-Based Hybrid Model for Depression Severity Modelling with an Expert-Annotated Longitudinal Dataset

ACL ARR 2025 May Submission5241 Authors

Abstract:

Depression severity is a clinically important indicator for assessing mental health status and guiding treatment, yet it remains challenging to infer reliably from user-generated text. Existing NLP research has largely focused on binary depression detection or isolated symptom classification, with limited attention to modelling severity progression over time.
We present the DepSy Severity dataset, a novel resource for modelling depression severity from social media, fully annotated by psychologists with weekly severity scores for users who self-report a depression diagnosis. To address this task, we propose CTMC-LSTM, a hybrid model that combines LSTM-based predictions with temporal dynamics modelled through a Continuous-Time Markov Chain (CTMC), capturing user-level severity progression over time. We frame severity estimation as both regression and multi-class classification, evaluating a range of architectures and feature combinations. Our experiments show that models incorporating structured features outperform text-only baselines, and the CTMC-LSTM model yields the highest performance in severity classification, particularly for underrepresented classes such as moderate and severe depression. These results highlight the importance of integrating temporal context for robust mental health modelling.

Paper Type: Long

Research Area: Human-Centered NLP

Research Area Keywords: Depression Severity, Markov chains, Temporal Modelling, longitudinal data.

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Publicly available software and/or pre-trained models, Data resources, Data analysis

Languages Studied: English

Submission Number: 5241