AI-Driven Fall Detection Wearable IoT Device Development

A primary difficulty in accurately detecting falls is the problem of "False Positive", differentiating a severe fall (a fall that poses a risk to the elderly) from the elderly simply sitting down quickly or the user dropping the device. Our solution employs Edge-based Machine Learning technology to provide fall detection that is both cost-effective and real-time.

The device has been designed as either a small pendant or wristband that contains advanced motion sensors (accelerometers, gyroscope, magnetometer).

Sensing System

The Bosch BMI270 device (6 degree of freedom Inertial Measurement Unit) was used for its very low power needs and its ability to recognize gestures.

Processor

The Nordic Semiconductor nRF52840 was selected for this application as it can support local inference models while also providing battery run time of ~10+ days.

Connectivity

This application uses two types of cellular connections, LTE-M (Cat-M1) and NB-IoT, to enable greater penetration into buildings (important for falls that occur in a bathroom), and at a lower power requirement than traditional 4G networks.

Instead of relying on basic threshold triggers that can generate false alarm, we used a much more involved multi-stage detection system consisting of:

Stage 1

Impact Detection- The accelerometer monitors for a high-G event

Stage 2

Orientation analysis- The gyroscope checks to see if the device's position has gone from vertical to horizontal and has remained in the horizontal position for 10 seconds or longer.

Stage 3

Edge AI inferences- A TensorFlow Lite Micro model looks at 3 seconds of movement data prior to and after the impact and searches for specific movement patterns (called "fall signatures", which are different based on whether a slip or trip occurred).

The unit is a stand-alone "Safety Hub," meaning no need for your cell phone to be paired to it.

Two-Way Voice

The unit features a high gain microphone and speaker which facilitates immediate voice communication. After a fall is detected, the device will automatically open a "Live-Voice" session with either a 24/7 monitoring center or a designated family member.

Precision Location

The device uses GPS/GNSS to determine outdoor location and employs Wi-Fi sniffing to determine indoor location of the user within a building (e.g., which floor, which room, etc).

Haptic Cancellation

The unit vibrates very strongly for 15 seconds after detecting a fall in order to minimize the chances of generating a false alarm. If the user is safe, they simply need to tap the unit to cancel any alarm before the local emergency services are dispatched.

The backend will be hosted using AWS IoT Core, which provides users (family and clinical staff) with a centrally managed platform.

Health Telemetry

The App will track the Activities of Daily Living (ADLs) for the individual (e.g., steps taken, sleep quality, gait stability trends). An acute decrease in gait speed can be an indicator of an impending fall.

Battery Alerts

Caregivers will receive automated SMS/Push notifications when the device's battery drops to 15%.

Device Health

The device will allow for real-time monitoring of both the signal strength and sensor calibration status.

98% Accuracy

"High-Impact" activities such as clapping or sitting were accurately filtered.

Fast Response Time

The average time from when a resident has an incident to when a caregiver receives notification has reduced to less than 30 seconds after the incident.

Scalable Architecture

The system provides "White Label" integration for home care agencies and insurers.