Predicting Mental Health
from Wearable Data.
A causal-inference study using biometric data from 17,500+ participants in the NIH All of Us dataset to identify behavioral drivers of mental health.
What can wearable data tell us about mental health, before it shows up in a survey?
[The gap: most mental health prediction relies on self-report. Wearables capture behavior continuously, but most analyses are correlational. Can we move from passive tracking to actionable early insight?]
From raw signal to causal claim, in four stages.
[The pipeline: pre-processing biometric streams, K-Means clustering on behavioral features, regression to depression outcomes, LiNGAM to test directionality.]
Sleep and steps — not heart rate — drive the signal.
[Key finding: sleep duration and daily step count emerged as the strongest behavioral predictors. LiNGAM revealed direct (not just correlational) paths to depression risk.]
Causal claims need humility, and structure.
[Reflection: where the method held up, where it didn't, and what it would take to take this from research to a real product.]