Fewer Alarms.
More Prevention.
Better Outcomes.
Four features, built on 2,000+ analysed falls, that adapt to each resident every day. 19 individually calibrated alarm rules. A 21-feature prediction model. Walking aid monitoring for the highest-risk population. Autopilot to keep it all current.
From alarm fatigue to alarm intelligence
Traditional fall alarms treat every movement the same. We analysed over 2,000 confirmed falls across the US, Denmark, the UK, and Switzerland. For each fall, we identified the exact transition in the 5-minute window before it happened — what the resident was doing, what time of day, how long they had been inactive. Then we built rules targeting those transitions, measured against real outcomes.
What Optimised Alarms does
Four features that work together — presets define the rules, prediction updates the risk, walking aid alarms fill the gap, autopilot keeps everything current.
Match alarms to risk
Each risk level — Low, Medium, High — maps to a distinct set of ML-calibrated transition rules. The High profile alone contains 19 individually tuned rules, each specifying which movement to alert on, at what time of day, after how long inactive, and at what alert level. Every rule was tested against 2,000+ confirmed falls with measured catch-rates and false-positive rates.
Predict risk daily
A logistic regression model trained on real patient outcomes analyses 21 features per resident every day — fall history, safe-to-ground events, walking aid usage, sleep regularity changes, respiration shifts, and mobility patterns. It produces a probability score mapped to risk thresholds calibrated for 53% recall at the High level and 80% recall at Medium.
Monitor walking aids
The sensor detects when a walking aid or wheelchair has moved more than one metre from the bed. Walking aid users represent 32% of the population but account for 52% of all falls — and 67% of those falls happen without the aid. This single alert targets the largest preventable fall category.
Keep settings current
Autopilot runs at 3:50 AM every day. It takes each resident’s ML-predicted risk level and applies it to their alarm preset automatically. If sleep has fragmented, respiration has shifted, or a fall happened recently, sensitivity increases overnight. Staff manual decisions are never overridden.
Product walkthrough video
90–120s · Team-led
Staff trust the alarm
Every alarm rule is measured against real outcomes. The highest-priority night rule — bed edge movement after 1+ hour inactive — catches 35 falls per rule with just 0.97 false positives per bed per day. When the alarm goes off, it means something.
Catch risk shifts overnight
The prediction model detects changes in sleep regularity, respiration rate, and mobility the moment they shift. A 30-day deterioration in walking speed or a recent safe-to-ground event both feed directly into tomorrow’s risk score — not a quarterly care review.
Every resident covered, every day
Autopilot closes the loop between prediction and action. Risk levels update at 3:50 AM every morning, without staff needing to check a single setting. Department-level and patient-level toggles give managers full control. Complete audit trail of every change.
The data behind it
Every alarm rule in Optimised Alarms is calibrated against real data from over 2,000 confirmed falls across care settings in four countries.
| Finding | Data |
|---|---|
| Getting out of bed | 2.5x more likely to result in a fall |
| Night-time transitions (19:00–07:00) | 63% more likely to result in a fall |
| After 1+ hour inactive in bed | 2.4x more likely to result in a fall |
| Walking aid users | 3.2x risk — 67% fell without their aid |
| Previous fall in last 3 months | 5x more likely to fall again within 30 days |
| Wheelchair users getting out of bed | 5.3x fall risk |
Under the hood
How alarm presets are built
Each alarm preset is a set of transition rules. A rule defines a specific state change — what the resident was doing, what they moved to, how long they were inactive, and what time of day. Every rule was measured against real fall data.
Here are five rules from the High-risk profile, showing how the system distinguishes between transitions that warrant an immediate alarm and those that warrant a warning.
Lying in bed → Sitting on bed edge
After 1+ hour inactive · Night (19:00–07:00)
The single most precise rule — catches the most dangerous night-time transition with less than 1 false alert per bed per day.
35
falls caught
0.97
FP / bed / day
Lying in bed → Sitting on bed edge
Under 1 hour inactive · Night (19:00–07:00)
Catches the most falls of any single rule. Warning-level because shorter inactivity means higher false-positive rate.
42
falls caught
3.23
FP / bed / day
Sitting on bed edge → Standing on floor
1–5 min on edge · Night (19:00–07:00)
The bed-exit moment. Staff have a short intervention window — the resident is upright but not yet walking.
31
falls caught
1.28
FP / bed / day
Sitting in chair → Standing on floor
10–60 min seated · Night (19:00–07:00)
Night chair transfers after extended sitting are rare but dangerous. Alarm-level with very low false-positive rate.
11
falls caught
0.44
FP / bed / day
Wheelchair → Standing on floor
Any duration · Any time
Wheelchair-to-standing carries 5.3x fall risk. Always-on monitoring regardless of time of day.
14
falls caught
1.86
FP / bed / day
The High-risk profile contains 19 rules like these. The Medium profile uses a focused subset. Every rule was tested against 2,000+ real falls.
The prediction engine
21 features, updated daily
The fall risk prediction model analyses 21 data points per resident every day. It groups them into five categories, each contributing to a single probability score.
Fall history
Fall in last 3 months, safe-to-ground events
Strongest predictor — recent fall increases risk by +1.25 log-odds
Mobility profile
Walking aid user, wheelchair user, lift user, bedridden status
Walking aid users: +0.62. Lift users: −1.10 (consistently supervised)
Sleep patterns
Sleep duration, sleep regularity index, 30-day changes
Deteriorating sleep regularity is an early indicator of increased risk
Vital signs
Respiration rate, respiration 30-day trend
Rising respiration rate correlates with physiological instability
Activity metrics
Walking speed, stationarity score, 30-day trends
Declining walking speed flags mobility deterioration before a fall happens
How thresholds work
High threshold
Top 20% of risk scores. Catches 53% of future fallers. These residents get the full 19-rule alarm profile — the most sensitive monitoring available.
Medium threshold
Top 50% of risk scores. Catches 80% of future fallers. These residents get a focused rule set covering the highest-precision night-time transitions.
How teams are using Optimised Alarms
The features that help care teams act on the right signals, every shift.
New resident setup
Staff select a risk level and walking aid status during admission — two clicks. The system activates the corresponding rule set immediately. No manual threshold tuning needed.
Overnight monitoring
Night-specific rules activate at 19:00. A resident moving to the bed edge after lying still for an hour triggers an immediate alarm. The same movement at 2 PM triggers a warning or nothing at all, depending on risk level.
Post-fall response
After a fall, the ML model recalculates risk using the strongest coefficient in the model (+1.25 log-odds for recent fall). Autopilot increases alarm sensitivity that night. Protection increases within hours, not days.
Care reviews and family updates
Managers review predicted risk levels, the factors driving them (with coefficient weights), alarm history, and trend data. Every prediction comes with the specific features that influenced it.
“[Staff quote placeholder — CS to collect]”
Night nurse or nurse assistant · [Community name]
Early results
We rolled out Optimised Alarms at [TBD] communities. Here’s what the data shows.
Fewer notifications
per resident per day
More residents covered
with active notifications
Fewer falls
per 1,000 patient-days
Faster response
average staff response time
Fewer alarms. More residents covered. Fewer falls. Faster responses. Not a trade-off — all four improved at once.
Feature details
Optimised Alarm Presets
19 individually calibrated rules per high-risk resident.
Each alarm preset maps to a specific set of transition rules derived from analysing 2,000+ confirmed falls. A rule specifies: the start state (e.g., lying in bed), the end state (e.g., sitting on bed edge), the inactivity duration before the transition, the time of day (day: 07:00–19:00 / night: 19:00–07:00), and the alert level (warning or alarm). The High-risk profile contains 19 active rules. The Medium profile uses a subset focused on the highest-precision night-time rules. Every rule carries measured performance: how many real falls it would have caught, and the expected false-positive rate per bed per day. Smart 5-minute deduplication prevents alert floods from the same transition.
Fall Risk Prediction
21 features. Trained on real outcomes. Updated daily.
A logistic regression model (v1.1, AUC-ROC 0.74) analyses 21 features per resident daily. The top coefficients: recent fall (+1.25), recent safe-to-ground event (+0.74), walking aid user (+0.62), wheelchair user (+0.29). Protective factors reduce risk: lift users (−1.10, consistently supervised transfers) and bedridden residents (−0.18, rarely attempt unsupervised mobility). The model outputs a probability score. Two calibrated thresholds convert it to a risk level: High (top 20% of scores, 53% recall — catches every other future faller) and Medium (top 50%, 80% recall). Staff see the predicted level, the driving factors, and a one-click Apply button.
Walking Aid & Wheelchair Alarms
Targeting the 52% of falls that come from 32% of residents.
Walking aid users are 3.2x more likely to fall — and 67% of those falls happen when the aid is out of reach. The sensor detects walking aid displacement greater than 1 metre from the bed and alerts staff before the next transition. For wheelchair users, the system monitors all wheelchair-to-standing transfers (5.3x fall risk) around the clock. These two populations represent the single largest category of preventable falls in our dataset.
Autopilot
Risk levels recalculated and applied at 3:50 AM every morning.
Autopilot runs as a scheduled job (daily, 3:50 AM Copenhagen time). It takes the ML-predicted fall risk level and applies it to each resident’s alarm preset. If yesterday’s data shows fragmented sleep, a respiration shift, or a recent fall, alarm sensitivity increases before the next shift starts. When conditions improve, it scales back down. Manual staff decisions are never overridden — the system only modifies the fall risk field. Department-level and patient-level toggles. Full audit trail.
“[Staff quote placeholder — CS to collect]”
Manager or general manager · [Community name]
Häufig gestellte Fragenzu Optimised Alarms & Teton
Wie unterscheidet sich Optimised Alarms von Standardalarmeinstellungen?
Wie wurden die Alarmregeln erstellt?
Was sagt das ML-Modell tatsächlich voraus?
Trifft das System Entscheidungen ohne Personaleingriff?
Funktioniert das mit meiner bestehenden Teton-Installation?
Ist dies ein Medizinprodukt?
Wie genau ist das Sturzvorhersagemodell?
Können wir Optimised Alarms ohne Autopilot nutzen?
See it in your community