Mental Health User Classification from Chronological Reddit Posts Using Transformer–Recurrent Hybrid Models

Abstract

Social media platforms contain valuable longitudinal information about users' linguistic patterns and behavioural changes, enabling automated mental health analysis. However, most existing approaches classify individual posts independently and fail to capture the chronological dependencies present in user-generated content. This study proposes a user-level mental health classification framework, MentalBERT–Temporal-BiLSTM, that fuses domain-specific contextual post embeddings with normalized inter-post time intervals and processes the resulting chronological representation with a bidirectional Long Short-Term Memory (BiLSTM) network. The framework is evaluated on 16,336 Reddit users across six classes—ADHD, Asperger's syndrome, depression, OCD, PTSD, and normal—achieving 90.64% accuracy and a 90.10% macro F1-score, and outperforming non-temporal and unidirectional temporal recurrent baselines. The results indicate that jointly modelling semantic content, posting-interval behaviour, and bidirectional sequential dependencies improves user-level classification over post-level and unidirectional approaches.