Daily Language as an Objective Indicator of Depressive Affective States: A Two-Week Ecological Momentary Assessment...

Workers performing emotional labor are at increased risk for depression, yet conventional self-report assessments often lack objectivity and temporal sensitivity. This study aimed to examine whether longitudinal analysis of daily natural language collected through ecological momentary assessment (EMA) can serve as an indicator of depressive affective states and to compare its temporal sensitivity with traditional self-report measures.

A total of 400 call center employees completed three voice-recorded free-text entries per day for two weeks using an EMA application. Transcriptions were analyzed using a lexicon-based sentiment approach (Linguistic Inquiry and Word Count [LIWC]) and three large language models (LLMs; GPT-4o-mini, Qwen, and Mistral) under zero-shot prompting. Depressive symptoms were assessed using the Patient Health Questionnaire-9, and neuroticism was measured using both self-reported questionnaires and language-inferred scores.

LLM-derived sentiment scores significantly differentiated groups across levels of depressive symptom burden and consistently outperformed LIWC. Longitudinal analyses demonstrated clear group-level separation, particularly in morning entries, whereas self-reported mood ratings failed to distinguish groups and showed lower adherence over the two-week period. Language-inferred neuroticism exhibited stronger associations with depressive symptoms than self-reported neuroticism. A cumulative model based on morning sentiment scores showed progressively improved discrimination over time, reaching its highest performance on Day 10 (area under the curve=0.75).

Daily natural language captures meaningful longitudinal affective dynamics associated with depressive symptoms and may complement conventional self-report assessments as a low-burden and scalable indicator of depressive affective states in real-world occupational settings.