LLMs Track Depression in Daily Video Diaries With Clinical Fidelity
Depression monitoring has long depended on self-report questionnaires that were built for convenience, not for the messy cadence of real life. Brief daily diaries create a different problem: they capture change, but they also demand repeated, low-burden measurement that patients can realistically sustain. In this study, large language models were tested as a way to score depression from daily video diary transcripts, and the result was not a novelty exercise. It was a direct comparison with self-report and interview-based severity in a cohort of 108 participants, 45% of whom were already in treatment.
The design matters because intensive longitudinal assessment is only useful if it can preserve signal across days. Participants completed a clinical interview and then 14 to 21 days of depression surveys alongside daily video diaries. The model-generated ratings were then compared with self-report depression at the within-person level, where day-to-day variation is clinically meaningful, and at the between-person level, where overall burden and illness severity are more easily recognized. The two methods converged moderately within individuals, with _r_ = .45 and 95% CrI [.40, .49], and more strongly when daily scores were averaged into a single estimate of overall depression, with _r_ = .61 and 95% CrI [.43, .75].
Speech, Symptoms, and the Clinical Interview Signal
A diary transcript is not just a text artifact. It contains pauses, hedges, self-evaluations, emotional intensity, and the kind of subtle language shifts that often accompany anhedonia, guilt, hopelessness, and psychomotor slowing. LLMs are well suited to this setting because they can read semantic nuance at scale rather than forcing every symptom into a fixed checkbox. In that sense, the model is not merely counting words; it is estimating clinical state from narrative structure and content.
The link to clinician-rated severity reinforced that possibility. Averaged LLM ratings correlated with depression symptom severity established through clinical interview at _r_ = .53, with 95% CrI [.35, .68]. For a digital measure to reach that level against an interview standard is important, because interviews remain the closest thing psychiatry has to a reference measure in routine practice. It also suggests that the LLM output is not duplicating self-report in a superficial way. It is capturing a related, but not identical, construct.
One finding stood out immediately.
LLM ratings and self-report produced nearly identical association profiles with daily functioning.
Within-person profile correlation reached _r_ = .98, and the between-person profile correlation reached _r_ = .94. Those values indicate that when depression fluctuated across days, both methods tracked the same practical impairments in a strikingly similar pattern. Functioning is where depressive symptoms become visible at the bedside: sleep disruption, reduced activity, social withdrawal, work difficulty, and a narrowed behavioral repertoire. A measure that mirrors those shifts closely has obvious value for research, but also for longitudinal care where clinicians need more than a monthly questionnaire snapshot.
Added Information Beyond Self-Report
The more consequential result is that the LLM did not simply echo what participants already said on surveys. Multivariable regression showed that LLM ratings predicted daily functioning variables with partial β values between |.12| and |.28|, and predicted interview-rated depression with partial β = .33 and 95% CrI [.07, .58], above and beyond self-reports. That pattern implies incremental validity. In practical terms, the model extracted clinically relevant information that the survey response set failed to capture fully.
Why might that happen? Self-report scales are constrained by the items they ask and the response options they permit. A diary transcript is open-ended, which allows a participant to describe fatigue, emotional blunting, interpersonal strain, and moment-to-moment mood in a way that standardized scales may compress. Depression is not always expressed as a clean endorsement of sadness. It often appears as flattened language, reduced specificity, and a global sense that ordinary tasks require effort. An LLM can detect those features if the transcript preserves enough context.
The study therefore points toward a useful division of labor between measurement modes. Surveys still offer efficiency and comparability, while diary-based language analysis may recover clinically meaningful variance that structured items miss. That is especially relevant in intensive longitudinal designs, where the goal is not just to classify illness but to follow its movement day by day. The broader promise is a measurement system that is less brittle than single-time-point scales and less burdensome than repeated clinician interviews.
What This Approach Can and Cannot Claim Yet
The sample was modest, and the work focused on participants selected for mental health treatment status rather than a broader community population. That setting strengthens clinical relevance, but it also leaves open how well the approach performs in milder depression, in people with different communication styles, or in settings where diary adherence is lower. The measurements were based on brief daily video diaries, so the model’s performance may depend on the richness of spoken self-expression and the quality of transcription.
The study also stops short of showing that LLM scoring can replace established clinical assessment. It does not establish treatment response thresholds, diagnostic cutoffs, or safety performance in high-risk scenarios such as suicidality monitoring. What it does show is more specific and more immediate: language-based modeling can match self-report closely, align with interview-rated severity, and carry additional prognostic information about daily functioning. That combination is enough to justify further validation in larger and more diverse cohorts, especially if the goal is real-time depression tracking with lower patient burden.
For clinicians and researchers, the most useful implication is straightforward. Depression assessment does not have to rely entirely on fixed questionnaires when patients are already speaking about their symptoms every day. In the right workflow, brief diary language may become a measurable clinical signal rather than an unstructured side note. The remaining work is to define where that signal is strongest, where it fails, and how it behaves when the clinical stakes are higher than a study diary.