Cannabis detection model

Core idea

Cannabis is one of the best wearable detection candidates after alcohol. Tolerance is the main enemy of accuracy.

Primary feature stack

  1. Delta resting HR from personal baseline — strongest anchor signal
  2. Minimum HR / plateau pattern during intoxication window (~10–30 min post-inhale)
  3. REM suppression same night — strongest overnight corroborator
  4. Sleep efficiency drop that night
  5. Motion context — separates from exercise (cannabis HR event is plateau-like, not spike-decay)
  6. Tolerance-adjusted priors — frequency-adjusted priors for heavy users

Published evidence

  • MobiFit (Fitbit + smartphone fused): 99% AUC, 0.85 F1 for moderate-to-intense intoxication (naturalistic)
  • Apple Watch-specific model: does not exist yet
  • Expected watch-only performance: AUC ~0.75–0.85 in occasional users

Harder problems

  • Heavy daily users: acute HR spike muted by tolerance → false negatives
  • Distinguishing from alcohol: cannabis has sharper immediate HR event + REM suppression; alcohol has stronger next-day HRV suppression + elevated RHR
  • Distinguishing from exercise: cannabis plateau HR vs exercise spike-decay curve + different motion/thermal context

Confounders

Stress/anxiety, sleep deprivation, stimulants — all can elevate HR similarly.

Cannabis, HRV signatures, REM suppression, Sleep architecture, Cannabis withdrawal