AI-Based Non-Verbal Behavioral Biomarkers for Differentiating Anxiety and Depression in Telemedicine

Abstract

1. Introduction

Telemedicine has broadened access to psychiatric care, yet remote assessments still struggle to capture non-verbal cues that reveal affective and psychomotor states. Given that anxiety disorders and major depressive disorder(MDD) share neurochemical mechanisms in serotonergic and noradrenergic pathways, quantitative behavioral biomarkers may enable objective differentiation of overlapping symptom profiles.

2. Objective

This study aimed to implement an AI-based video analysis pipeline capable of automatically extracting and quantifying non-verbal behavioral features from patient videos to distinguish anxiety from depression, and to explore its potential applicability in medical populations under psychological stress.

3. Method

Using a custom Python-based system integrating Mediapipe, OpenCV, and NumPy, we directly implemented a video analysis pipeline that extracted non-verbal metrics from Symptommedia clinical videos(10 anxiety and 10 MDD cases).

(1) Facial Expression Mean — frame-to-frame change in the mouth-to-eye ratio, indicating facial expressivity;

(2) Voice Percentage — proportion of speech frames determined by a 30-ms energy-based voice activity detector using a 75th-percentile threshold;

(3) Movement Mean — robustly normalized head movement derived from nose-tip displacement and interocular tilt variation.

Each metric was interpolated across sampled frames, and one-way ANOVA was performed to evaluate group differences.

4. Results

Movement significantly differed between the anxiety and MDD groups(p = 0.021, η² = 0.29), indicating greater psychomotor activation in anxiety and retardation in depression. In contrast, Facial Expression(p = 0.549) and Voice Percentage(p = 0.120) were non-significant, suggesting generalized affective blunting, likely influenced by similar pharmacologic modulation(e.g., SSRIs or SNRIs).

5. Conclusion

The successful implementation of this AI-based behavioral analysis pipeline demonstrates the feasibility of quantifying psychomotor and affective features from clinical videos. The system may extend to identifying and monitoring stress-related psychiatric symptoms in medical students and residents, who often experience subclinical anxiety or depression while facing barriers to mental health care. Integrating such analytics into telemedicine could enable early detection and personalized support for well-being in high-stress healthcare environments.

6. Keywords:

telemedicine, non-verbal behavior, psychomotor activity, depression, anxiety