Related Experiment Video
Updated: Mar 19, 2026
06:37
Author Spotlight: Addressing Technical and Subjective Challenges in Measuring Classroom Attention
Published on: December 15, 2023
5.6K
Optimized CNN-based facial analysis for depression detection: Managing mental disorder in education
Mengqiu Wang1, Yanxia Wu1, Khosro Rezaee2
1Sichuan Vocational College of Cultural Industry, Chengdu 610000, China.
Brain Research Bulletin
|March 18, 2026
Summary
A new AI tool, Optimized CNN-Based Facial Analysis (OCFA), offers a privacy-preserving way to screen for Major Depressive Disorder (MDD) in educational settings. This lightweight, face-only system provides accurate risk estimation for early intervention.
Area of Science:
- Computer Science
- Artificial Intelligence
- Machine Learning
- Psychiatry
Background:
- Major Depressive Disorder (MDD) is often under-identified in educational settings due to resource limitations and privacy concerns.
- Studio-based programs may present unique stressors, necessitating low-burden screening tools for triage.
- Existing facial analysis methods for mental health often require multimodal data, lack robustness, and are not optimized for on-device use.
Purpose of the Study:
- To develop and evaluate OCFA (Optimized CNN-Based Facial Analysis), a lightweight, privacy-preserving, face-only machine learning pipeline for MDD risk estimation.
- To ensure the system is compute-efficient, robust, and suitable for on-device deployment in educational contexts.
- To provide a calibrated decision support tool for human-in-the-loop screening workflows.
Main Methods:
- Developed OCFA using a RobFaceNet-style backbone with Adapt-Coordinate Attention and multi-objective evolutionary model selection.
- Implemented post-hoc temperature scaling for improved probabilistic calibration.
- Aggregated frame-level facial data into session-level scores using pooling/MIL variants, ensuring subject/session separation for privacy.
- Utilized the DAIC-WOZ dataset with PHQ-8 labels and cross-domain validation on E-DAIC.
Main Results:
- OCFA achieved 82.98% accuracy, 82.61% F1, and AUROC=0.886 on the DAIC-WOZ test set.
- The system demonstrated good cross-domain generalization on E-DAIC (81.10% accuracy, 80.20% F1).
- Achieved a calibrated ECE of approximately 0.040 with low computational cost (0.065 GMac) and 3.80M parameters.
- Privacy-aligned interpretability was provided using SHAP values without revealing facial imagery.
Conclusions:
- OCFA is a calibrated, efficient, and privacy-preserving facial analysis tool suitable for on-device MDD risk estimation in educational settings.
- The system shows promise for integration into human-in-the-loop screening workflows, particularly in art and design programs.
- Further validation under real-world educational capture conditions is required before operational deployment.
