A raised heart rate is not proof of danger. A missed response is not proof of safety. We are exploring how motion, context and human judgment could work together—with their limits made visible.
Idea stageNo field validation yetOpen research agenda
THE AMBITION
Research target: 99% fewer false alarms. A target to test.
The original concept proposes a 99% reduction in false alarms. That is an aspirational research target, not a measured result, accuracy rating, or assurance that an escalated alert represents a real threat.
Three unanswered prompts do not make an emergency 99% certain. Repeating the same evidence does not create three independent confirmations.
Define the baseline before testing. Measure both systems on the same held-out population, exposure hours and event definitions. A zero baseline makes this percentage undefined.
Track user prompts, contact escalations and emergency-service escalations separately. Lowering one is not evidence that every stage improved.
A BALANCED SCORECARD
Fewer interruptions. Without hiding missed emergencies.
False alerts per hour
Count incorrectly triggered events per worn user-hour, with confidence intervals and a clear definition of each alert stage.
Emergency recall
Measure the share of labeled target events detected. Report missed events alongside false alerts; a silent system can appear precise while failing its purpose.
Time to assistance request
Measure detection, user check-in and contact-delivery latency separately. Report typical and slow cases, including degraded connectivity.
Positive predictive value
Report how many escalations correspond to target events. Evaluate realistic low event prevalence, not only balanced laboratory datasets.
Human burden
Measure dismissals, abandoned setup, alert fatigue, retention and contact availability. Usability is part of a safety evaluation.
Performance across people
Break results down by hardware, wear position, activity, accessibility needs and relevant participant characteristics, with uncertainty.
SCENARIO LIBRARY
Design for the inconvenient cases.
An evolving engineering threat model, from a crowded commute to a broken connection. These are proposed safeguards and test cases—not validated product capabilities or an exhaustive account of every emergency.
Loading the scenario library…
ADAPTATION, WITH GUARDRAILS
Learn routines. Keep uncertainty.
A proposed personalization period could establish normal motion and sensor patterns. One month of wear alone does not establish reliable emergency detection.
Exercise, dancing and martial arts may resemble concerning movements. Activity context should inform a decision, never blanket-disable protection. Explicit SOS should remain available while personalization is incomplete.
Bounded learning
Use confidence and signal-quality checks before interpreting a pattern.
Separate routine recognition from emergency escalation decisions.
Do not automatically label every dismissal as a harmless event.
Review model changes, keep version history and enable rollback.
Treat unusual, missing or conflicting data as uncertainty—not an “all safe” result.
Test silent danger and inability to respond, as well as obvious movement.
VALIDATION ROADMAP
Earn each next step.
01 / DEFINE
Scope and hazard analysis
Specify intended users, target events, excluded uses, hardware constraints and severity-dependent escalation. Predefine the baseline, outcome labels and acceptance criteria.
02 / BENCH
Signals and failure injection
Test recordings, sensor placement, clock drift, dead batteries, duplicate messages and network interruption. Use consented ordinary activities and safe simulations; no dangerous reenactments.
03 / SHADOW
Observe before alerting
With consent and appropriate ethics review, assess a frozen candidate model without dispatch. Split data by participant and time to reduce leakage. Independent adjudication should examine missed events too.
04 / LIMITED PILOT
Supervised, opt-in evaluation
Only after predefined gates are met: a bounded pilot with trained contacts, explicit limitations and stop criteria. Any authority connection requires a separate approved operational pathway.
Current position: idea stage. This is a proposed research roadmap, not work reported as completed. No hardware prototype, trained classifier, collected study dataset, enrolled pilot or emergency-response integration is presented.
THE RESPONSE MATTERS TOO
An alert is not a completed rescue.
The proposed system must distinguish an alert created, delivered, acknowledged and acted on. An unanswered guardian notification should not be mistaken for a confirmed emergency—or successful assistance.
Sharing location with police or SHE Teams is a future integration question. No partnership, dispatch access or guaranteed response exists here. A cancellation should update recipients already notified; it cannot undo an emergency call or guarantee that responders stop.
Location can be stale, inaccurate or unavailable. Future interfaces should show its age and uncertainty rather than presenting a dot as certainty.
External references inform the research agenda. They do not validate this concept or imply a relationship with Pulse IQ Technologies.
PRODUCT DOCUMENTATION
Apple Fall Detection
Apple documents user confirmation and escalation behavior, and explicitly notes that not every fall is detected and high-impact activity may trigger an alert.
A distinct smartwatch study illustrates why real-world false-call burden and sensitivity must be evaluated together. Its condition, hardware and results do not transfer to threat detection.
Consult the official Emergency Response Support System for current public emergency information. This website is not an emergency service or an ERSS-connected application.
We welcome discussion with researchers in wearable sensing, human factors, privacy and safety engineering. Collaboration would begin with scope, evidence and responsibilities.