Depression assessment stops at one summary: EDA pins each behavioral cue to its source clip, lifting 7B F1 by 7.61 and 9.88 points
Synopsis
Video-based depression assessment previously condensed long recordings into a single behavioral description, leaving it hard to check which footage an observation came from or how it shaped the judgment. EDA localizes five acoustic and four visual cues from prior literature into evidence records with time intervals and measurements, has frozen multimodal agents check them against the original media before admission, and compares the result with PHQ-9-guided text analysis. On D-Vlog and LMVD, EDA Support with a frozen 7B backbone raises F1 over direct prediction by 7.61 and 9.88 percentage points.
Interpretation
The output of depression assessment shifts from a narrative about a whole recording to individual behavioral records that trace back to specific time intervals, each carrying the audio or video it came from. Earlier multimodal methods summarized long recordings into global behavioral descriptions; a phrase such as frequent pauses names a behavior but does not identify where or when it occurs, nor how it affects the final judgment. On D-Vlog and LMVD, AI cue annotations agree with human consensus labels on shared source segments at 91.02% accuracy, a figure pooled across both datasets that measures cue agreement rather than classification.
The system proceeds in three stages: observe the behavior, verify it against the original media with frozen models, and only then use admitted records for comparison. Verification splits into Support, Counter and Evidence roles: Support judges whether the media supports the phenomenon, Counter uses an expanded interval to test whether the pattern persists and departs from the person's baseline, and Evidence adjudicates admission or exclusion per cue while keeping the grounds. The final Support policy keeps positively verified records directly and retains 1.26 more percentage points of candidate intervals than extended Evidence review under D-Vlog Leave-Query-Out and 7.40 more on LMVD.
Behavioral evidence and PHQ-9-guided text analysis each produce a judgment, and only their disagreement triggers a review of the original transcript. Text agents retrieve transcript passages under the nine PHQ-9 symptom domains and distinguish current personal experience from history, denial, third-party statements and general discussion, without assigning questionnaire scores. With a frozen 7B backbone, Support improves F1 over direct prediction by 7.61 percentage points under D-Vlog Leave-Query-Out and 9.88 on LMVD, and exceeds 7B DepressionAgent by 6.14 and 5.28 points.
Perspective
The demonstrated scope is two video cohorts, D-Vlog and LMVD, evaluated with frozen Qwen2.5-Omni-7B and 30B-A3B backbones, and cue agreement is measured only on 3-second source segments shared by the two datasets. These cohorts were used while developing cues and decision rules, so training references alone do not amount to an untouched prospective evaluation.
Behavioral cues are nonspecific and annotations can contain errors; source links allow inspection but do not establish clinical validity. Recording-level folds do not guarantee speaker separation, and single-run scores do not measure inference variability or isolate the effect of model size. Among changed predictions under conflict review, 87.50% are beneficial under D-Vlog Leave-Query-Out and 58.97% on LMVD, rates conditional on outputs that changed.
