On-Device Early Fall-Detection Wearable
US-based aging-in-place / digital-health wearable startupA wearable continuously analyzes motion data locally to identify movement patterns associated with potential falls, enabling rapid detection without continuously transmitting raw sensor data to the cloud.
Where AI accelerated vs. where our engineers led.
Every project follows the same Rivix Split — AI drafts the mechanical work, a named engineer reviews and owns every decision that touches safety, cost, or a regulatory limit.
Where AI accelerated
- IMU driver implementation & firmware modules
- Signal-processing & feature-extraction code
- Model integration & inference pipeline
- Mobile implementation
- Automated test generation
Where our engineers led
- Fall-detection methodology & dataset strategy
- Model selection, accuracy/recall tradeoffs
- False-positive management
- Real-time & power constraints
- Validation
Technical deep-dive
Additional detail on architecture, stack and implementation specifics for this engagement.
Architecture & stack
The wearable pairs an IMU (accelerometer/gyroscope) with a Cortex-M-class MCU capable of running a lightweight inference model directly on-device, avoiding a dependency on continuous cloud connectivity for the time-critical detection path. A BLE link and companion mobile app handle configuration, event upload and caregiver alerting.
- IMU (accelerometer + gyroscope) feeding a Cortex-M-class MCU with on-device inference
- Lightweight classifier running fully on-chip — no cloud dependency on the detection path
- BLE link to a companion app for configuration, alerts and event history
Connectivity & data pipeline
Raw IMU samples are processed locally through a feature-extraction and classification pipeline running entirely on-device; only detected events (not raw streams) are pushed over BLE to the phone and onward to the cloud, keeping the safety-critical detection loop independent of network conditions.
- Feature extraction (windowed acceleration/orientation) computed entirely on-device
- Only classified events, not raw IMU streams, cross the BLE link
- Event payloads timestamped locally to preserve ordering before upload
Key engineering challenges
The central challenge was tuning the detection model for recall (catching real falls) while keeping false positives low enough that the device stays trustworthy for daily wear — a tradeoff validated against a labeled motion dataset covering both falls and fall-like everyday movements (sitting quickly, dropping the device, etc). Doing this inference within the power and compute budget of a small wearable, while keeping detection latency sub-second, shaped the model and firmware architecture.
- Tuning recall vs. false-positive rate against a labeled fall / near-fall motion dataset
- Fitting classifier inference inside a sub-second, low-power compute budget
- Distinguishing genuine falls from fall-like everyday motion (sitting fast, device drops)
Why this approach
On-device inference was chosen deliberately over a cloud-inference model so that detection latency and reliability don't depend on network availability — a hard requirement for a safety use case where the device may be relied upon in areas with poor connectivity.
- On-device inference chosen specifically to decouple detection latency from network availability
- Event-only uploads (not raw streaming) kept both bandwidth and battery draw minimal
- Classifier kept intentionally lightweight to fit real-time constraints on constrained hardware