TinyML SpO₂ and HRV On-Device Screening
The challenge
Ambulatory devices must balance clinical fidelity for arrhythmia/HRV metrics with long battery life and reliable artifact rejection. Clinicians require compact, annotated event summaries rather than continuous raw traces. This demanded a high-efficiency approach to product engineering to create a scalable solution for remote health monitoring.
What we built
EurthTech delivered a production-ready ambulatory SpO2 + HRV device, built on AI-powered embedded systems. The solution runs robust PPG preprocessing and a TinyML classifier locally (apnea/AF screening), combining local short-window FFT/HRV analysis. By uploading only event-summaries, the system minimizes data, preserves patient privacy, and enables fast clinician triage.
How it works
The custom hardware uses synchronous LED timing and high-SNR AFE. Our embedded systems development team deployed firmware that performs matched filtering, beat detection, and robust artifact masking. HRV features are computed on-device and compressed. Event detection uses a small quantised CNN and rule-based fusion. Results & Impact: This engineering validated clinical claims via bench testing against pulse-ox and ECG reference standards, providing flagged summaries for efficient clinician use. The on-device intelligence drastically reduces measurement drift and enables per-unit calibration.
Validation and security
Bench validation against clinical reference is required for claims; for wellness SKUs the device provides flagged summaries. Security & Update Path: We use secure OTA with signed images and a small recovery bootloader. Looking ahead, planning for NIST PQC moves ensures crypto-agility.