Home Product Hardware Facilities Science Research About Log in Get Started

Evidence-based

The science
behind Guardian.

Every metric Guardian monitors is grounded in peer-reviewed research. We chose what to track based on what the evidence says actually predicts health events in elderly people.

The core principle

Personal baseline.
Not population average.

Most clinical thresholds are population-derived. "Normal" resting HR is 60–100bpm. But for a person whose personal baseline is 58bpm, a reading of 72bpm is a significant elevation — clinically "normal" but personally concerning.

Guardian uses exponentially weighted moving averages (EWMA) to build a personal baseline for each metric, with recent data weighted more heavily than older data. Anomalies are detected using z-scores against that personal baseline.

When two or more metrics trend in concerning directions simultaneously, Guardian issues a compound warning — because co-occurring anomalies are far more predictive than any single metric alone.

Anomaly detection method
Step 1 — EWMA Baseline
Build personal mean and standard deviation using 30-day half-life exponential weighting
Step 2 — Z-Score Detection
Flag readings that deviate from personal baseline by more than 1.5–2.5 standard deviations
Step 3 — Compound Warning
Issue elevated warning when 2+ metrics trending badly simultaneously over 72-hour window
Step 4 — Tier Routing
Severity score determines which of the 5 alert tiers fires — proportionate, not alarming

Evidence by metric

What the research shows.

Gait VariabilityHighest Evidence
The single most validated fall predictor in peer-reviewed literature. Inconsistency in step timing — even at normal walking speed — means the brain and motor system are struggling to coordinate. This manifests measurably in accelerometer data 24–72 hours before a fall event.
Signal: Step timing variance increasing >15% from personal baseline over 3+ consecutive days
Heart Rate Variability (HRV)Highest Evidence
HRV is one of the most sensitive early indicators of physiological stress. It declines measurably before infection, cardiac events, and significant illness — often 24–36 hours before symptoms appear. In elderly people, a personal HRV decline of more than 20% sustained over 3 days is a strong early warning signal.
Signal: 7-day average HRV dropping >20% below personal 30-day baseline
Nocturnal MovementHigh Evidence
Increased restlessness during sleep — measured by accelerometer — predicts infection and discomfort 12–24 hours before clinical presentation. In elderly people, UTIs frequently present as nocturnal agitation before any urinary symptoms appear.
Signal: Nocturnal movement events increasing >40% above personal baseline for 2+ nights
Resting Heart Rate TrendStrong Evidence
A sustained upward trend in resting HR — rather than any single elevated reading — is a reliable predictor of developing infection, dehydration, or cardiac stress. The trend matters more than the absolute value.
Signal: 3-day average resting HR trending upward >8% above personal baseline
Sleep ArchitectureStrong Evidence
Disruption to deep sleep (slow-wave) is associated with cognitive decline, immune suppression, and fall risk. The link between sleep fragmentation and next-day physical performance in elderly people is well-established. Guardian tracks sleep efficiency and fragmentation against personal norms.
Signal: Sleep efficiency dropping >15% below personal average for 4+ consecutive nights
Social Circadian RhythmStrong Evidence
Disruption to established daily movement patterns — getting up later, moving less, skipping the morning walk — is a leading indicator of depression onset and early cognitive episodes.
Signal: Daily routine regularity score dropping >25% for 4+ consecutive days
Evidence strength
Gait Variability
Highest
HRV
Highest
Nocturnal Movement
High
Resting HR Trend
Strong
Sleep Architecture
Strong
Circadian Rhythm
Strong
Guardian's advantage
Compound detection — when 2+ metrics co-occur, predictive accuracy increases significantly beyond any single metric alone.

Our approach

How Guardian
learns your person.

📡
Continuous passive collection
Data is collected passively from Apple Watch or Withings via Apple Health. No interaction required. No buttons to press. No logs to fill in.
🧮
EWMA baseline modelling
Exponentially weighted moving averages build a living model of each person's health. 30-day half-life means recent data matters more — naturally adapting to seasonal change and ageing.
🔍
Z-score anomaly detection
Readings are compared to the personal baseline using z-scores. Statistically robust, avoids false positives, and scales across every metric simultaneously.
📈
72-hour trend analysis
Guardian looks for trends over 72 hours — not just today's readings. A single bad night means nothing. Three bad nights in a row is a signal worth acting on.
🤖
AI interpretation
Raw anomaly data is interpreted by AI into plain English summaries. No jargon. Just: "Mum's sleep has been disrupted this week — might be worth a call."
🎯
Proportionate alerting
The 5-tier system ensures every alert is proportionate. Minor deviations get gentle nudges. Serious compound anomalies escalate to family, GP, or emergency services.

See it in action.

Early access is free. Connect an Apple Watch and Guardian starts learning within 14 days.

Get started free