Summary: Early signs of depression can be difficult to notice, yet new research shows that artificial intelligence can detect subtle changes in facial micro-movements that correlate with depressive symptoms. In a study of Japanese undergraduates, students with subthreshold depression were perceived by peers as less expressive, friendly, and likable, even though they did not appear more nervous or artificial. AI-driven analysis identified specific eye and mouth movement patterns that matched depression scores, suggesting a non-invasive route for early screening in schools, workplaces, and digital health platforms.
Using short self-introduction videos, researchers combined subjective peer impressions with automated facial action unit analysis to reveal consistent differences in positive facial expressivity among participants with subthreshold depression (StD). These differences were subtle enough to escape untrained observers but detectable by the OpenFace 2.0 system.
Key Facts
- Peer Ratings: Students with subthreshold depression were rated as less expressive, likable, and friendly.
- AI Detection: Micro-movements in eyes and mouth—measured as specific facial action units—correlated strongly with depression scores.
- Early Screening Potential: The method provides a non-invasive, scalable approach that could aid early detection of depression before clinical symptoms appear.
Source: Waseda University
Background
Depression is among the most common mental health conditions and is often associated with reduced facial expressivity and altered social communication. Subthreshold depression (StD) refers to a set of depressive symptoms that fall short of a formal diagnosis but increase risk for developing clinical depression. Whether StD also affects facial expressions and first impressions has been unclear.
Associate Professor Eriko Sugimori and doctoral student Mayu Yamaguchi of Waseda University’s Faculty of Human Sciences explored this question by combining brief video recordings with artificial intelligence-based facial analysis and peer evaluations.
Their study, published in Scientific Reports on 21 August 2025, recorded 10-second self-introduction videos from 64 Japanese undergraduates (ratees). A separate group of 63 students (raters) provided impression ratings for expressivity, friendliness, naturalness, and likability while both groups completed the Beck Depression Inventory–II (BDI-II).
Findings
Peer ratings showed that ratees with StD (BDI-II scores 11–20) received lower scores on positive items—expressive, natural, friendly, and likable—than healthy ratees (BDI-II scores 1–10). Notably, raters did not judge those with StD as more stiff, fake, or nervous, indicating that StD is associated with a muted positive expressivity rather than overtly negative or awkward behavior.
Automated analysis using OpenFace 2.0 identified a profile of facial action units (AUs) that were more frequent or pronounced in participants with StD. These included AU01 (Inner Brow Raiser), AU05 (Upper Lid Raiser), AU20 (Lip Stretcher), and mouth-opening actions (AU25/26/28). Five of these AUs remained significantly associated with BDI-II scores after correcting for multiple comparisons, showing that micro-movement signatures corresponded with depressive symptom severity.
Implications
The study suggests that short, naturalistic video recordings combined with automated facial-action analysis can reveal subtle emotional expression differences linked to early depressive symptoms. Because the changes are subtle and not readily noticed by casual observers, automated tools can add sensitivity to screening efforts.
Potential applications include incorporation into mental health technology, digital health platforms, university wellness programs, and workplace screening initiatives where non-invasive, scalable methods are needed to monitor psychological well-being and flag individuals who may benefit from support or further assessment.
Limitations and considerations
The authors note that the study was conducted with Japanese students, and cultural norms influence emotion expression and impression formation. As a result, findings may not generalize directly across different cultural or age groups. Furthermore, while AI tools can detect correlates of StD, they do not replace clinical evaluation and should be used as part of a broader assessment strategy.
“Our approach of short self-introduction videos combined with automated facial-expression analysis offers an accessible, non-invasive way to detect potential early signs of depression,” says Sugimori. “It may enable earlier intervention and support before clinical symptoms emerge.”
About this AI and depression research news
Author: Armand Aponte
Source: Waseda University
Contact: Armand Aponte – Waseda University
Image: The image is credited to Neuroscience News
Original Research: Open access.
“Subthreshold depression is associated with altered facial expression and impression formation via subjective ratings and action unit analysis” by Eriko Sugimori et al., Scientific Reports. DOI and original article details are available from the publishing journal.
Abstract
Subthreshold depression is associated with altered facial expression and impression formation via subjective ratings and action unit analysis
Depression is often linked to reduced facial expressivity and to biases in recognizing others’ emotions. Whether subthreshold depression (StD)—a potential early stage—shows similar alterations remained uncertain.
The study recorded 10‑second self‑introduction videos from Japanese undergraduates (ratees; n = 64) and collected subjective impression ratings from a separate group of raters (n = 63). Both groups completed the Beck Depression Inventory‑II (BDI‑II).
Raters’ own depressive tendencies did not predict their impression ratings after correction. Ratees classified with StD (BDI‑II = 11–20) received significantly lower scores on positive items—expressive, natural, friendly, likeable—compared with healthy ratees (BDI‑II = 1–10; partial η² = 0.18–0.70).
Automated OpenFace 2.0 analysis revealed higher presence or intensity of AU01 (Inner Brow Raiser), AU05 (Upper Lid Raiser), AU20 (Lip Stretcher), and mouth‑opening AUs (AU25/26/28) in StD faces; several of these AUs correlated with BDI‑II after false‑discovery‑rate correction (q < 0.05).
Overall, subthreshold depression was associated with muted positive expressivity and distinct eye‑ and mouth‑movement patterns but did not markedly alter first‑impression judgments by observers. These observable AU signatures may help identify individuals at risk for clinical depression and support earlier intervention efforts.