It started as a sleep-aware remote. Now it's an AI health tracker.
PulseApnea started as a sleep-aware wristband that turned your devices off once you fell asleep. After months of research and development it is now an AI-powered sleep apnea wristband, aimed at a condition that affects an estimated 1 billion people worldwide.
From first idea to where it stands today
-
Phase 01May 2026Concept
The sleep-aware remote
Another IoT sleep remote, built for convenience: a wrist device that notices when you have fallen asleep and switches off nearby electronics to save electricity.
Concept: watch to Home Assistant to IR blaster to TV -
Phase 02June 2026Prototype
A physical watch that talks to the room
Serious prototyping began. The watch became a physical device and the infrared remote integration was completed, so it could turn off the devices around the user.
First prototype model: enclosure, board and battery -
Phase 03August 2026Pivot
From convenience to health
While building sleep tracking we found that the methods that reliably tell whether you are asleep also produce detailed physiological data. We wanted that data to deliver more value, and began linking it to health conditions. Obstructive sleep apnea (OSA) affects an estimated 1 billion people worldwide, so we pivoted to researching a device that could help.
-
Phase 04September 2026PulseApnea
First polished prototype
The first polished PulseApnea prototype was built. We upgraded the sensors and microcontroller for the new purpose, contacted local hospitals for training data, and began developing and training our own machine learning models to push detection accuracy further.
The new hardware centres on a Seeed XIAO nRF52840 connected over I²C to a pulse sensor. We drew the schematic and routed a custom two-layer board in KiCad.
Schematic: microcontroller and sensor wiring 
Routed board layout, front and back copper 
KiCad PCB editor -
Phase 05 · NowOctober 2026Current build
Companion app and ongoing model training
We have started building our own app to pair with the device, while continuing to improve and train the machine learning model in parallel.
-
NextDate to be setRoadmap
Validate on more patients
The model now reaches 80% accuracy on the four-class severity task, but on a small labelled set. Next: confirm it holds with participant-grouped cross-validation and additional patient data, and keep improving the CNN.
What the device measures, and what it does with it
Data flow
Core functions
- Pairs with
- PulseApnea wristband
- Status
- In development
- Devices
- Fans, TVs, speakers
- Codes stored
- Most common fan, TV and speaker codes
- Inputs
- SpO₂, heart rate, motion, microphone
- Output
- Night-level severity: normal, mild, moderate or severe
- Sensor
- MAX30102 (PPG)
- Error vs ref.
- ± 5 bpm vs Apple Watch
- Tracks
- Wrist movement and activity level
- Inputs
- Accelerometer
- Capacity
- 200 mAh
- Charge time
- 1 h
Training data
Model performance
| Metric | Value | What it means |
|---|---|---|
| Accuracy | 80% | Nights placed in the correct severity class |
| Sensitivity (severe) | 75% | Severe nights correctly flagged |
| Specificity (severe) | 68.8% | Non-severe nights correctly cleared |
| F1 score | 0.718 | Macro and weighted F1 are both 0.718 |
What sleep apnea looks like in real overnight data
Interactive charts built from open, clinically scored polysomnography recordings on PhysioNet. They illustrate the signals our model works with.
Severity of the 25 patients
Patients grouped by apnea-hypopnea index (AHI), the number of breathing events per hour of sleep.
Who has apnea: age vs BMI
Each dot is one patient, colored by severity. Hover for their AHI.