AI Reads Between the Lines: How Daily Video Diaries Could Replace Depression Questionnaires

Mental health monitoring may be on the verge of a significant transformation. Researchers have demonstrated that artificial intelligence can accurately assess depression severity by analyzing transcripts of brief daily video diaries—and remarkably, the AI detects clinically relevant information that patients don't report on standard questionnaires.

The validation study, published in npj Digital Medicine, tracked 108 participants over two to three weeks as they completed clinical interviews, daily depression surveys, and video diary entries. Nearly half of the participants were currently receiving mental health treatment, providing a clinically diverse sample for testing the AI approach.

Strong Correlation with Traditional Measures

The large language model (LLM) analysis of diary transcripts showed substantial agreement with established depression assessment methods. When comparing day-to-day fluctuations, LLM-derived depression scores correlated at r = .45 with participants' self-reported depression levels. When averaged across the study period, this correlation strengthened to r = .61 with overall self-reported scores and r = .53 with depression severity determined through clinical interviews.

Perhaps most striking, the AI ratings achieved profile correlations of r = .98 within individuals and r = .94 between individuals when compared against self-reports of daily functioning variables—suggesting the technology captures nuanced patterns in how depression manifests across different life domains.

Capturing What Surveys Miss

The study's most clinically significant finding emerged from multivariable regression analysis: LLM ratings predicted both daily functioning and interview-rated depression severity even after accounting for traditional self-report measures. The AI contributed unique predictive information with partial effect sizes ranging from |.12| to |.28| for daily functioning variables and .33 for interview-rated depression.

This suggests that analyzing natural speech patterns reveals aspects of depression that structured questionnaires fail to capture—potentially including subtle changes in language use, emotional expression, or cognitive patterns that patients themselves may not consciously recognize or report.

Implications for Mental Health Care

For healthcare providers managing depression treatment, this technology offers a potential solution to a persistent challenge: tracking symptom changes without overwhelming patients with frequent surveys. The approach could enable real-time monitoring of treatment response and earlier detection of symptom deterioration, potentially allowing for more timely intervention adjustments.

Digital mental health platforms may find particular value in integrating natural language processing capabilities that extract clinical insights from patient-generated content—whether video diaries, audio journals, or text entries—without requiring structured questionnaires. This could substantially reduce patient burden while improving the granularity and accuracy of longitudinal monitoring.

The research, supported by multiple grants from the US National Institutes of Health, establishes a foundation for AI-powered depression assessment that balances clinical validity with practical feasibility. As mental health care increasingly incorporates digital tools, the ability to derive meaningful clinical information from brief, unstructured patient narratives may represent a significant advance in making intensive monitoring both effective and sustainable.

Sources

1. Large language models for depression assessment from brief daily diaries — npj Digital Medicine