Alcohol detection model

Detection rationale

Alcohol produces a characteristic multi-signal biometric signature: HRV suppression, elevated resting heart rate, and disrupted sleep architecture. The combination is more discriminative than any single signal.

Primary signals

Signal	Direction	Magnitude	Duration
RMSSD	↓	30–50% below baseline	24–48 h
Resting HR	↑	+10–25 bpm above baseline	12–24 h
REM sleep	↓	20–50% reduction	1–2 nights
Sleep efficiency	↓	Measurable drop	Same night
WASO	↑	Increased arousals	Same night

Feature stack

Optimal multi-signal model combines:

1. HRV delta from personal baseline (most important)
2. RHR delta from personal baseline (corroborating)
3. Sleep efficiency drop (context)
4. Motion absence (rules out exercise-induced HRV suppression)
5. REM suppression (strongest sleep-stage signal)

Wearable accuracy

• Oura Gen 4: CCC=0.99, MAPE=5.96% for HRV — most accurate consumer wearable
• Oura Gen 3: MAPE=7.15%
• WHOOP: MAPE=8.17%
• No wearable can specifically attribute HRV suppression to alcohol without contextual features

Differentiation from other stressors

Stressor	HRV	RHR	REM	Motion
Alcohol	↓↓	↑	↓↓	↓
Vigorous exercise	↓	↑	↑ (recovery REM)	↑↑
Illness/infection	↓	↑	Variable	↓
Poor sleep (behavioral)	↓	Variable	↓ REM	Normal

Expected AUC

• Occasional users: AUC ~0.75–0.85
• Heavy users: AUC worse unless tolerance-adjusted
• Acute signals (same night): stronger than next-day signals

Limitations

• Cannot specifically identify alcohol — only detects “physiological stress pattern consistent with alcohol”
• Heavy drinkers with tolerance: acute HRV spike muted; must weight sleep signals more
• Combining with self-report or contextual data (evening motion = low, no gym) dramatically improves specificity

Related

Alcohol, HRV signatures (alcohol), Sleep architecture (alcohol)