Personal tracking tool, not a medical device

Our Research

How Moodery detects episodes — the methodology, signals, strategies, and clinical research behind every confidence score.

Multi-Strategy Ensemble Detection

The Problem

Single detection algorithms are fragile. They over-weight certain signals, miss atypical presentations like irritable hypomania or high-functioning depression, and give binary yes/no answers without transparency into what triggered the detection.

Our Approach

Four independent detection strategies analyze the same daily data, each emphasizing different biomarkers. An episode is confirmed when multiple strategies agree — like getting a second opinion built into the algorithm.

Daily Data

(Energy, Drive, Sleep, HRV, Activity, ...)

Balanced

Physiological

Self-Report

Sleep-Circadian

Consensus Engine

2 of 4 detect hypomania → Confirmed

DSM-5 Aligned

Episode duration thresholds, both-pole requirements, and pattern classification follow diagnostic criteria.

Research-Backed

Signal importance drawn from peer-reviewed studies on digital biomarkers for mood episodes.

Transparent

Every threshold and weight is visible and adjustable in-app. No black boxes.

11 Biomarkers, Two Sources

Moodery combines what you report with what your body says. When signals are missing, the algorithm normalizes available weights — it adapts rather than breaks.

User Input

EnergyUser Input

Core 1–5 scale. Sustained low energy is the most self-reportable depression marker; sustained highs correlate with hypomania onset.

DriveUser Input

Motivation independent of energy. The mismatch between energy and drive reveals transitional states invisible to single-scale trackers.

Self-ReportUser Input

Composite of mood valence, social battery, arousal, and emotionality. Captures subjective experience that physiology can miss.

Catches high-functioning depression where activity and heart rate appear normal.

IrritabilityUser Input

1–5 scale tracking irritation intensity. Key for irritable hypomania — negative mood with elevated energy — which gets misclassified without this signal.

Risky BehaviorUser Input

Optional flag for impulsive actions (spending, substance use, sexual impulsivity). Hypomania-only signal — contributes zero to depressive scoring.

Apple Watch

Sleep DurationApple Watch

Deviation from your personal baseline. Asymmetric thresholds: 15% decrease for hypomania (even mild undersleep matters), 25% increase for depression (20% oversleep is normal variation).

Sleep variability predicts hypomania onset with 0.87 accuracy.

HRVApple Watch

Heart rate variability (SDNN) measures autonomic nervous system balance. Drops indicate dysregulation. Differentiated using resting HR: elevated RHR + low HRV suggests sympathetic activation (hypomania); normal RHR + low HRV suggests parasympathetic withdrawal (depression).

Resting Heart RateApple Watch

Deviation from personal baseline. Elevated RHR reflects sympathetic nervous system activation, common in both mania and anxiety.

Top single predictor of hypomania with AUROC = 0.852.

ActivityApple Watch

Step count deviation from baseline. Both extremes matter: hyperactivity in hypomania, withdrawal in depression.

Sleep StagesApple Watch

REM-to-deep sleep ratio compared to baseline. Depression increases REM percentage and reduces deep sleep — an established finding in sleep medicine.

Altered sleep architecture is a DSM-5 specifier for mood episodes.

Circadian RegularityApple Watch

7-day standard deviation of sleep onset times. Circadian disruption is a prodromal marker — it often fires before episodes fully manifest.

Identified as an early warning signal for mood episodes in recent research.

Four Detection Strategies

Each strategy runs independently on the same daily data but weighs signals differently. This isn't redundancy — each strategy is designed to catch what others might miss.

Balanced

All signals contribute moderately — good all-rounder

Best at catching: General-purpose detection when all data sources are available

Depressive Weights

Energy
0.20
Drive
0.15
Sleep
0.15
Self-Report
0.13
Activity
0.10
HRV
0.09
Circadian
0.07
Resting HR
0.06
Sleep Stages
0.05

Hypomanic Weights

Energy
0.18
Sleep
0.15
Self-Report
0.10
Activity
0.10
Drive
0.09
HRV
0.08
Resting HR
0.08
Risky Behavior
0.07
Irritability
0.06
Circadian
0.05
Sleep Stages
0.04

Physiological

Emphasizes passive markers — best with consistent Apple Watch data

Best at catching: Catching episodes the user doesn’t self-report — denial, alexithymia, or anosognosia

Depressive Weights

Sleep
0.22
HRV
0.16
Activity
0.16
Resting HR
0.12
Sleep Stages
0.10
Circadian
0.10
Energy
0.08
Drive
0.04
Self-Report
0.02

Hypomanic Weights

Sleep
0.20
Resting HR
0.16
HRV
0.14
Activity
0.14
Circadian
0.08
Sleep Stages
0.08
Energy
0.08
Drive
0.04
Risky Behavior
0.04
Self-Report
0.02
Irritability
0.02

Self-Report

Emphasizes your subjective experience — catches high-functioning depression

Best at catching: Detecting episodes with normal physiology but altered inner experience

Depressive Weights

Energy
0.25
Self-Report
0.25
Drive
0.22
Sleep
0.08
Activity
0.05
HRV
0.05
Circadian
0.04
Sleep Stages
0.03
Resting HR
0.03

Hypomanic Weights

Energy
0.22
Self-Report
0.15
Irritability
0.15
Risky Behavior
0.12
Drive
0.12
Sleep
0.08
Activity
0.06
HRV
0.04
Resting HR
0.03
Sleep Stages
0.02
Circadian
0.01

Sleep-Circadian

Based on research showing sleep patterns are the strongest episode predictors

Best at catching: Early warning and prodromal detection — circadian disruption often precedes episodes by days

Depressive Weights

Sleep
0.25
Circadian
0.18
Sleep Stages
0.15
Energy
0.10
Resting HR
0.08
HRV
0.08
Activity
0.06
Drive
0.05
Self-Report
0.05

Hypomanic Weights

Sleep
0.22
Circadian
0.18
Sleep Stages
0.12
Resting HR
0.10
Activity
0.08
HRV
0.08
Energy
0.08
Risky Behavior
0.05
Drive
0.04
Irritability
0.03
Self-Report
0.02

All weight tables are visible and adjustable in Settings → Algorithm. Every table sums to exactly 1.00.

The Consensus Engine

A single algorithm can be wrong. Multiple independent algorithms agreeing is much harder to fool — the same principle behind ensemble methods in machine learning and second opinions in medicine.

01

Run

All 4 strategies analyze your daily scores independently, each producing its own episode candidates with confidence scores.

02

Compare

The engine finds overlapping time windows where multiple strategies detected the same episode type.

03

Confirm

Episodes where 2+ strategies agree become confirmed. Single-strategy detections remain visible but marked as lower confidence.

Example: 10-Day Detection Window

Balanced
Physiological
Self-Report
Sleep-Circadian
Consensus
12345678910
3–4 of 4 agree

Reduces False Positives

A sick day might fool the Physiological strategy — elevated resting heart rate, poor sleep. But Self-Report and Balanced won't fire if your energy and mood are explained by illness, not depression.

Reduces False Negatives

High-functioning depression might not show in physiology — normal steps, normal heart rate. But the Self-Report strategy catches the low drive and negative mood. If Sleep-Circadian also fires from disrupted sleep onset, that's 2 of 4 — confirmed.

How It Works In Practice

Three real-world scenarios showing how different strategies respond to the same data — and why a multi-strategy approach catches what single algorithms miss.

Scenario

The Sick Week

Four days of flu. Low energy, poor sleep, elevated resting heart rate — a naive single-algorithm detector would flag this as a depressive episode.

Energy: 2Sleep: −20%RHR: +12%HRV: DownMood: NeutralDayFact: Sick
Balanced
Not detected
Physiological
Weak signal (0.35)
Self-Report
Not detected
Sleep-Circadian
Not detected

Consensus: 0 of 4 — No episode. The illness confound tag prevented misclassification.

Without confound awareness, this would be a false positive depressive episode.

Scenario

Irritable Hypomania

Energy through the roof, sleeping 4 hours and feeling fine, but mood is irritable and dark — not euphoric. Traditional trackers see negative mood and call it depression.

Energy: 5Drive: 4Sleep: −40%Mood: −2Irritability: 4Activity: +60%RHR: +8%
Balanced
Hypomania (0.78)
Physiological
Hypomania (0.82)
Self-Report
Hypomania (0.71)
Sleep-Circadian
Hypomania (0.75)

Consensus: 4 of 4 — Hypomania confirmed (0.77 confidence)

The irritability signal prevents misclassification. Without it, the negative mood would pull toward a depressive reading.

Scenario

High-Functioning Depression

Going through the motions — hitting the gym, showing up to work — but internally empty. Sleep duration is normal but quality is poor. Heart rate and steps look fine.

Energy: 3Drive: 2Sleep: NormalREM/Deep: +30%Mood: −2Social: 1Activity: NormalRHR: Normal
Balanced
Weak signal (0.48)
Physiological
Not detected
Self-Report
Depressive (0.68)
Sleep-Circadian
Depressive (0.58)

Consensus: 2 of 4 — Depressive episode confirmed (0.63 confidence)

Self-Report catches what physiology misses. The sleep stage signal adds the second vote despite normal sleep duration.

What We Filter Out

Certain conditions mimic mood episodes. Moodery applies confound modifiers to prevent misclassification.

Illness

Composite scores reduced by 50%. Illness confound flag added to episode factors.

Travel

Sleep and activity signal weights reduced by 50% for that day. Energy and self-report remain full.

Medication Changes

Confidence reduced by 30%. Medication adjustment flag added to factors.

Menstrual Cycle

Depressive confidence reduced by 20%. Hormonal cycle confound flag added.

Research & Sources

Sleep Variability as a Predictor of Hypomania Onset

medRxiv, 2025

Sleep variability predicts hypomanic episode onset with 0.87 accuracy in a longitudinal study of bipolar spectrum participants.

SleepCircadian
View source

Resting Heart Rate as a Digital Biomarker for Hypomania

JMIR Medical Informatics, 2025

Resting heart rate emerged as the top single predictor of hypomania with AUROC = 0.852, outperforming activity and sleep metrics alone.

Resting HRHRV
View source

Circadian Disruption as a Prodromal Marker for Mood Episodes

PMC, 2025

Circadian rhythm disruption precedes mood episode onset by 3–7 days, suggesting utility as an early warning signal in digital monitoring.

CircadianSleep
View source

Diagnostic and Statistical Manual of Mental Disorders, 5th Edition (DSM-5)

American Psychiatric Association

Defines diagnostic criteria for cyclothymia (2+ years of sub-threshold episodes without 2-month symptom-free periods), bipolar II (4+ day hypomania, 2+ week depression), and rapid cycling (4+ episodes per year).

Pattern Classification
View source

Heart Rate Variability and Autonomic Dysregulation in Bipolar Disorder

Psychophysiology literature review

Reduced HRV correlates with mood episode severity across both poles. Combined with resting heart rate, HRV direction helps differentiate depressive from hypomanic autonomic profiles.

HRVResting HR

Moodery is a personal tracking tool, not a diagnostic instrument. Detected patterns are statistical estimates based on your data and should not replace professional clinical assessment. Always consult a qualified healthcare professional for diagnostic and treatment decisions.