Melatonin Detection Model

TL;DR

Direct wearable detection of melatonin ingestion is not validated and is not possible with current consumer devices. Any melatonin-specific inference from Apple Watch, Oura, Whoop, or similar devices is proxy-only and not individually reliable. The only defensible use case is population-level signal consistency or user self-report cross-validation.

⚠️ human_signoff_required: true — All biometric inference for melatonin is proxy-only. Do not build automated detection logic that claims to identify melatonin use from wearable data alone.

Detection approach

What is being detected

Not melatonin itself (no biochemical sensor exists in consumer wearables). Instead, a melatonin detection model would infer melatonin use from downstream physiological effects observable by wearables.

Validated signals (from RCTs)

Signal	Evidence	Wearable validity
Sleep-onset latency (SOL) reduction (~7–17 min at 0.5–5 mg)	Strong (multiple RCTs/meta-analyses)	Proxy only — effect size within consumer device error margin
Elevated nocturnal HRV (lab ECG)	Moderate (lab studies)	Proxy only — consumer wearable HRV never validated for melatonin
Reduced resting heart rate	Moderate (RCTs)	Proxy only — RHR is non-specific
Sleep efficiency improvement	Moderate (some RCTs)	Proxy only — sleep efficiency is well-measured by wearables but melatonin-specific inference is not

Unvalidated signals

Signal	Problem
Sleep staging shifts (REM↑, SWS↑/↓)	Apple Watch sleep staging κ=0.20 vs PSG — essentially no agreement
HRV frequency-domain changes (LF, HF)	Consumer wearables don’t reliably measure these
Core body temperature	Consumer wearables don’t measure CBT
Melatonin concentration	No consumer wearable measures this

Inference model framework

Input signals

melatonin_detection_score = f(
  sol_delta,          # change in sleep-onset latency (user-tracked or wearable-derived)
  hrv_delta,          # change in nocturnal HRV (rMSSD)
  rhr_delta,          # change in resting heart rate
  timing_alignment,   # was melatonin taken 30min–3h before bedtime?
  light_exposure,     # was dim-light maintained after dosing?
  baseline_confound   # pre-existing SOL, HRV, RHR severity
)

Confidence boundaries

• High confidence: User self-reports melatonin use + shows SOL improvement of >15 min + took it 30min–3h before bed + maintained dim-light
• Low confidence: Any single biometric signal in isolation
• Not detectable: From wearable data alone without user disclosure

Key confounds

Confound	Direction	Impact
Sleep restriction / recovery	Can produce SOL improvements independent of melatonin	High
Alcohol use	Produces opposite HRV/SOL profile — excludes rather than confirms	High
Caffeine timing	Can delay SOL independent of melatonin	Moderate
Exercise timing	Late exercise can shift circadian phase	Moderate
Other sleep aids	Benzodiazepines, antihistamines, L-theanine, magnesium — all affect biometrics	Moderate
Device model	Different wearables have different SOL/HRV accuracy	Moderate
Circadian phase delay	User may be taking melatonin at wrong time — phase delay worsens SOL	High

Detection is not the goal

The primary Vitals use case is not detecting melatonin use (users should self-report) but validating coaching efficacy and detecting timing errors:

• A user on melatonin who shows no SOL improvement → possible timing error, insufficient dose, or wrong etiology
• A user who reports morning melatonin use → flag for timing correction (morning use can cause phase delay)
• A user showing large SOL improvement with correct timing → consistent with melatonin benefit; no automated detection required

What this model cannot do

1. Cannot detect melatonin from a blood, saliva, or breath sensor — no such consumer wearable exists
2. Cannot reliably distinguish melatonin effect from placebo at the individual level — the effect is modest and within noise
3. Cannot detect compliance — HRV/RHR changes are too non-specific
4. Cannot support product quality claims — contamination and dose variance are not visible in biometrics

Coaching integration

Scenario	Appropriate action
User self-reports melatonin use; biometrics improve	Accept self-report; log for longitudinal tracking
User denies melatonin use; biometrics suggest possible benefit	Do not infer; ask about sleep habits, timing, environment
User reports morning melatonin use	Flag: morning use risks phase delay; suggest PM timing
User shows no SOL benefit despite melatonin use	Investigate timing, dose, light environment, or wrong insomnia etiology
User on fluvoxamine (CYP1A2 inhibitor) using melatonin	Flag: severe interaction; 17-fold AUC increase; do not recommend melatonin

Product sourcing signal

Note: Melatonin product contamination (71% off-label dose; 26% contain undisclosed serotonin) is not detectable from wearable biometrics. This is a sourcing guardrail, not a detection-model signal. Wearables cannot distinguish a sub-potent product from a correctly-dosed one.

→ See Melatonin hub for product sourcing recommendations.

Related notes

• Melatonin — hub note; dosing, evidence, safety
• Melatonin Sleep Biometrics — which sleep metrics melatonin changes and how reliably
• Sleep architecture — melatonin does not meaningfully alter sleep stages at physiological doses; sleep staging via consumer wearables is unreliable
• HRV — Apple Watch Limits — Apple Watch HRV accuracy caveats relevant to any melatonin HRV inference
• Circadian Biology — SCN signaling; timing is the critical variable for melatonin efficacy
• Cannabis detection model — contrast: cannabis detection is more feasible via HRV delta + REM + motion (AUC ~0.75–0.85); melatonin detection is harder due to smaller effect sizes and absence of a distinctive biometric signature
• Alcohol detection model — contrast: alcohol has a more distinctive biometric profile (HRV suppression + elevated RHR + sleep efficiency drop)