A causal discovery framework for digital phenotyping.

Area of Science:

• Computational psychiatry
• Digital health
• Behavioral science

Background:

• Digital phenotyping quantifies behavior for mental health prediction.
• Current models face a 'predictive plateau' with limited mechanistic insight.
• Opaque 'black box' models hinder understanding of well-being dynamics.

Purpose of the Study:

• Propose a paradigm shift from predictive classification to structural causal modeling.
• Develop a framework for inferring interpretable causal graphs of behavior and psychological states.
• Identify candidate causal biomarkers for mental health outcomes.

Main Methods:

• A two-stage computational framework using CNN-based encoders for behavioral embeddings.
• Neuro-symbolic causal discovery to infer directed graphs of behavioral-psychological dynamics.
• Principal Component Analysis (PCA) for dimensionality reduction and interpretation.

Main Results:

• Deep embedding models showed limited predictive accuracy for stress (AUC = 0.532), confirming a predictive plateau.
• Causal approach identified time-lagged associations, e.g., reduced sleep and mobility preceding stress.
• PCA revealed five components explaining 85% of variance, including 'Stationary Social Engagement.'

Conclusions:

• Structural causal modeling offers a more interpretable alternative to predictive models in digital phenotyping.
• Candidate causal biomarkers, like specific behavioral components, can generate hypotheses about behavior-stress relationships.
• This approach moves beyond prediction to understanding underlying mechanisms in mental health.