2025 · Preprint · Machine learning
Scalable Early Detection of Mental Health Signals in Social Media Text
This manuscript has not been peer reviewed.
A study of BERT-based encoder adaptation for identifying mental health signals in social media text. The work compares frozen encoders, full fine-tuning, and parameter-efficient approaches, including LoRA, across seven categories.
Read abstract
Early detection of mental health signals in social media text can enable timely interventions and reduce the risk of crises. This paper presents a systematic study of encoder adaptation strategies for automated screening of mental-health indicators in social media posts, with a primary focus on BERT-based encoders. Using a curated dataset labeled across seven clinically informed categories, we compare frozen-encoder classifiers, full encoder fine-tuning, and parameter-efficient approaches including low-rank adaptation (LoRA). Through extensive hyperparameter sweeps and controlled ablations we evaluate layer-wise unfreezing schedules, encoder/classifier head learning-rate splits, and regularization schemes. We show that full encoder fine-tuning achieves the highest classification performance with a peak F1-score of 0.83, substantially outperforming frozen and low-rank alternatives while maintaining strong generalization and balanced class sensitivity. Partial fine-tuning and LoRA offer notable compute and memory savings but require careful parameterization to avoid instability at large LoRA ranks. Overall, the findings demonstrate that a fine-tuned BERT encoder can serve as a reliable foundation for early diagnostic support, provided its deployment remains grounded in ethical design and clinical responsibility.