Falls are a leading cause of injury in older adults, and the time spent on the floor after a fall matters. Wearable detectors
promise automatic alerts — but published algorithms are usually tested on simulated falls by young volunteers. When they
were evaluated on real falls of older adults, sensitivity dropped sharply (≈ 57 % on average) and false alarms ranged from a few to
dozens per day. FallGuard lets you see why.
1 · Signals
A waist-worn tri-axial accelerometer at 50 Hz measures specific force: ≈ 1 g upwards when still. A fall has phases:
- Descent — the magnitude |a| drops (partial free-fall) while the body rotates.
- Impact — a short spike of several g.
- Aftermath — the body lies roughly horizontal and still.
The lab profile produces fast descents and hard impacts (3–6.5 g). The real-world profile reflects older adults:
slower, often without free-fall, softer impacts (1.5–3.8 g), and more slow collapses. Daily activities — sitting down heavily,
flopping onto a bed, stumbling, stairs — produce confusing spikes and postures.
2 · Algorithms
- Impact: alarm when |a| exceeds a threshold.
- + Posture: after an impact, wait 2.5 s and require a lying posture (tilt of the gravity vector) and stillness.
- State machine: free-fall → impact within 1 s → posture → stillness, the classic low-power design for microcontrollers.
- Learned: logistic regression on six window features (peak, pre-impact minimum, time below 0.8 g, orientation change,
post-impact motion, final uprightness), trained on either lab or real-world data.
3 · Evaluation
sensitivity = detected falls / falls
false alarms per day = Σ_activities P(alarm | activity) × occurrences per day
Occurrences per day approximate an older adult's routine (80 walking bouts, 25 sit-downs, 20 bends, 3 stumbles…). The benchmark sweeps
each algorithm's threshold to draw its trade-off curve.
4 · Alerting
A detected fall starts a countdown on the device (“Are you OK?”). Cancelling avoids a false call; otherwise the caregiver is notified.
This human-in-the-loop step is how real products keep false alarms tolerable.
Things to try
- Switch the benchmark between lab and real-world falls: every algorithm loses sensitivity.
- Use the state machine on the older adult and trigger a slow collapse — no free-fall, no alarm.
- Lower the impact threshold until sensitivity recovers and watch the false alarms per day climb.
- Train the learned model on real-world falls instead of lab falls.