Context Aware Health

On-device health monitoring for Wear OS — fall detection and heart rate anomaly alerts running entirely on a Fossil Gen 6 smartwatch, powered by Edge Impulse machine learning.

This project is a full wearable health stack: collect labeled sensor data from the wrist, train ML models in Edge Impulse, deploy them as native C++ libraries on the watch, and send caregiver alerts over Wi‑Fi when something looks wrong. No cloud inference required at runtime.

Hardware tested: Fossil Gen 6 · Wear OS · ARM (arm64-v8a / armeabi-v7a)

---

Why "Context Aware"?

The watch doesn't just read raw numbers — it understands context:

Signal	What the watch knows	What triggers an alert
Motion context	Gravity vector at 50 Hz over 1.5 seconds	Fall-like impact + orientation change + ML confidence
Physiological context	BPM pattern over 30 seconds	Deviation from learned "normal" baseline (K-means anomaly)
Connectivity context	Wi‑Fi available or not	Alert only sent when network is reachable

Two models run in parallel inside a foreground monitoring service. Each uses post-ML gating (impact detection, consecutive confirmation, cooldowns) so normal daily activity doesn't spam false alarms.

---

Features

• Fall detection — Gravity sensor → spectral neural network → email alert
• Heart rate anomaly detection — BPM time series → K-means anomaly score → email alert
• Live monitoring UI — On-watch Fall % / Normal % scores while monitoring
• Data collection mode — Record labeled sessions and upload directly to Edge Impulse for retraining
• Email alerts — Wi‑Fi email to caregiver (no phone companion required)
• On-device inference — Edge Impulse C++ MCU libraries via Android NDK (INT8 quantized)

---

Architecture

┌──────────────────────────────────────────────────────────────┐
│                     Fossil Gen 6 Watch                        │
├──────────────────────────────────────────────────────────────┤
│  Gravity (50 Hz) ──► libei_fall.so ──► Fall decision logic   │
│  Heart Rate (1 Hz) ──► libei_hr.so  ──► HR anomaly logic     │
│                              │                                │
│                    MonitoringService (foreground)             │
│                              │                                │
│                    PhoneAlertSender (Wi‑Fi email)             │
└──────────────────────────────────────────────────────────────┘
         ▲                                    │
         │ training data                      ▼
   Edge Impulse Cloud                   Caregiver Gmail
   (ingestion API)                     notification

Modules

Module	Purpose
:app	Watch app — data collector + health monitor + native ML
:phone	Optional Android phone companion (Wearable Data Layer listener)

---

ML Models

1. Fall Detection (fossil-fall-detection)

Type	Supervised binary classifier
Labels	Fall · Non-Fall
Input	75 samples × gravX/Y/Z @ 50 Hz (1.5 seconds)
Pipeline	Spectral Analysis (FFT) → 39 features → INT8 neural network
Output	P(Fall), P(Non-Fall)
Accuracy	~90% on Edge Impulse test set

Post-ML gating (reduces false alarms):

• Impact / orientation change must be present
• Inference every 0.5 s (not every sample)
• 2 consecutive high-confidence windows required
• Fall score must beat Non-Fall by 12%+
• Threshold: 72% normally, 55% on strong impact
• 2-minute cooldown between alerts

2. Heart Rate Anomaly (fossil-hr-activity)

Type	Unsupervised K-means anomaly detection
Input	30 BPM readings @ 1 Hz (30 seconds)
Pipeline	Spectral Analysis → 13 features → 4-cluster K-means
Output	Anomaly score (≥ 0.5 triggers alert)
Training	64 Normal samples + synthetic anomaly data

The watch firmware derives BPM from the optical PPG sensor on the back of the watch. The app reads Sensor.TYPE_HEART_RATE — it does not process raw PPG waveforms on Fossil Gen 6.

---

Sensors

Sensor	Used for	Rate	Notes
Gravity	Fall model	50 Hz	Primary input — isolates tilt/orientation
Accelerometer	Fallback	50 Hz	Only if gravity sensor unavailable
Heart rate	HR model	~1 Hz	OS-processed BPM from PPG hardware
Gyroscope	Data collection	50 Hz	Available, not in production models

IMU explained: The accelerometer measures force (including gravity). The gravity sensor strips out motion so the fall model sees orientation changes cleanly. A real fall shows a sudden impact spike plus a large shift in the gravity vector direction.

BPM explained: Green LED on the watch back shines into the wrist → photodiode measures reflected light → blood volume changes with each heartbeat → watch firmware converts the pulsating signal to beats-per-minute → app reads one BPM value per second.

---

Setup

Prerequisites

• Android Studio (Ladybug+) with Wear OS support
• Android SDK 35, NDK, CMake 3.22+
• Edge Impulse account
• Fossil Gen 6 (or compatible Wear OS watch) on Wi‑Fi for alerts

Configure API key

cp gradle.properties.example gradle.properties

Edit gradle.properties:

EI_API_KEY=ei_your_api_key_here

Get your key from Edge Impulse Studio → Dashboard → Keys.
Never commit gradle.properties — it is gitignored.

Configure alert email

Edit app/src/main/java/com/example/wizdatacollector/presentation/PhoneAlertSender.kt:

private const val ALERT_EMAIL = "your-email@gmail.com"

First run: FormSubmit sends a one-time activation link to that address — click it once.

---

Usage

Health monitoring (launcher app)

1. Connect watch to Wi‑Fi
2. Open the monitor app
3. Tap Test phone alert — confirm email arrives
4. Tap Start monitoring
5. Watch live Fall X% / Normal Y% on screen
6. Fall or abnormal HR → email alert to caregiver

Data collection (for retraining)

1. Select Gravity sensor (matches fall model)
2. Enter label: Fall, Non-Fall, Normal, etc.
3. Record 2–60 seconds
4. Uploads automatically to Edge Impulse

---

Edge Impulse Integration

Export models to watch

1. Train in Edge Impulse Studio
2. Deployment → C++ library (MCU)
3. Extract into:
   • Fall → app/src/main/cpp/ei_fall/
   • HR → app/src/main/cpp/ei_hr/
4. Rebuild

---

Tech Stack

Kotlin · Jetpack Compose · Wear OS · Android NDK · C++ · Edge Impulse · TensorFlow Lite INT8 · FormSubmit

---

Author

Vikramadhitya — 2026

Built with Edge Impulse on a Fossil Gen 6 smartwatch.