L'IA peut-elle détecter la dépression à partir de changements subtils dans les micro-expressions faciales lors d'appels vidéo ?

La reconnaissance des émotions à partir de vidéos a progressé rapidement grâce aux modèles d'apprentissage profond. Ces systèmes analysent des micromouvements faciaux que les humains oublient souvent de remarquer. Ils sont corrélés avec des échelles de dépression clinique et le suivi durable de l'humeur. Cette technologie soulève des questions éthiques concernant le consentement et la surveillance dans les interactions numériques.

Background

Emotional recognition from video has advanced rapidly due to deep learning models, which analyze minute facial movements often missed by humans. These systems correlate with clinical depression scales and sustained mood tracking, though they raise ethical questions about consent and surveillance in digital interactions. Current systems can reliably recognize basic facial action units and coarse emotions, but detecting depression from subtle, real-time micro-expressions in ordinary video calls remains unreliable in clinical settings. Research prototypes using 3D facial meshes, frame-level attention, and multimodal signals (voice, typing cadence) show modest correlations with PHQ-9 scores in controlled studies, yet generalization to diverse lighting, angles, and backgrounds is poor. Privacy, consent, and algorithmic fairness concerns further limit large-scale deployment, and no certified device is approved for diagnosis via video alone. (Enriched May 12, 2026; Source: National Academies of Sciences, Engineering, and Medicine)