Yusuke Fukazawa

Abstract In this paper, we propose a hybrid approach that combines Small Language Model (SLM)-based interpretation with machine learning (ML)-based prediction to analyze stress levels and related factors from step-count data. While several datasets exist for predicting mental health conditions from sensor data, most do not explicitly address the underlying factors associated with stress. To explore this issue, we collect step-count data from 30 nurses, together with stress assessments (QIDS: Quick Inventory of Depressive Symptomatology) and stress factor ratings based on six questionnaire items measured on a 4-point Likert scale, collected over 8 days within 4 weeks. We evaluate the proposed approach through two tasks. The first task examines how intermediate textual interpretations relate to stress presence estimation. Under our baseline experimental settings, BERT (Bidirectional Encoder Representations from Transformers) with intermediate stress interpretations achieved the highest accuracy (0.74), compared with BERT using raw step-count representations (step count: 0.63, distance: 0.59) and a prompt-based approach. The second task evaluates the association between intermediate interpretations and stress factor ranking. In this setting, BERT with intermediate stress interpretations achieved a ranking accuracy of 0.60, compared to 0.56 when using step-count sequences without interpretation. Higher correlations were observed for work-related stress factors such as “workplace relationships,” “busy work,” “heavy work responsibilities,” and “lack of time off.” Overall, these results suggest that intermediate textual representations derived from step-count data can be useful for stress analysis under baseline conditions, while avoiding causal claims about stress determinants.

Detecting mental illness from short social media posts is challenging because these texts are often brief, fragmented, and lack explicit descriptions of the user’s mental state. Prior studies using encoder-based models such as BERT show promise but struggle when key contextual information is missing. To address this, we propose a method that augments posts with interpretive sentences generated by MentaLLaMA-chat, a generative model specialized in mental health, and fine-tunes BERT on the augmented dataset. We curated 1,525 Japanese posts containing the word “mental” (in katakana) from X (formerly Twitter) and manually annotated them according to Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition criteria, labeling 557 posts as positive and 968 as negative. Our method improved recall by 2.4 percentage points compared to models trained on the original posts alone, while maintaining comparable accuracy and precision. Shapley Additive Explanations analysis revealed that tokens introduced by the interpretive sentences—including both negative and positive expressions—enhanced the model’s ability to identify mental-distress posts. These results demonstrate that generative-model-based text augmentation effectively provides additional context, enabling more accurate detection of mental illness indicators in short, ambiguous social media posts.