Multimodal Foundation Models for Early Detection of Depression and Anxiety

Abstract

Depression and anxiety are major mental health conditions
globally with substantial social, academic and economic costs.
Typical assessment tools such as self-reported symptom and
clinical observation are subjective, time consuming and can
be difficult to scale, but early identification is critical for timely
intervention and improved treatment outcomes. The latest
developments in AI include multimodal AI foundation models
which learn rich representations from various data types,
providing fresh opportunities for automated and accurate mental
health screening.Recent advancements in artificial intelligence
have given rise to the concept of multimodal AI foundation
models, which can learn rich representations from a variety
of data types, presenting new opportunities for automated
and accurate mental health screening. This study investigates
the use of multimodal foundation models for early detection of
depression and anxiety by combining textual, speech, facial,
physiological and neuroimaging modalities. The paper presents
an overview of the present multimodal learning frameworks,
foundation model architectures, and fusion strategies used
in mental health analytics. Additionally, it reviews how well
multimodal representations are able to capture complex
behavioral, emotional and cognitive indicators of depression
and anxiety disorders. Comparative evaluation shows that
multimodal foundation models continually outperform traditional
unimodal and conventional multimodal models through the
use of cross-modal interactions and large-scale pretraining.
Heterogeneity, privacy, interpretability, and ethical deployment
of data in clinical settings are also discussed. The results
suggest that multimodal foundation models hold significant
promise for advancing scalable, reliable, and clinically relevant
mental health screening systems, leading to earlier diagnosis
and better care decisions in psychiatric treatment.