Ultra-Low-Power Parametric Estimation Architectures on Microcontrollers for Decentralized, Edge-AI Continuous Screening of Obstructive Sleep Apnea

• Abiodun Okunola

  Ladoke Akintola University Technology

  Author

Obstructive Sleep Apnea (OSA) remains a pervasive yet significantly underdiagnosed sleep disorder, affecting hundreds of millions globally while relying on costly and cumbersome overnight polysomnography (PSG) as the diagnostic gold standard . This research addresses the critical gap between clinical diagnostic accuracy and the need for accessible, continuous, at-home screening by developing a novel parametric estimation architecture optimized for ultra-low-power microcontrollers. The proposed system employs autoregressive parametric modeling of respiratory effort signals, extracting key features including breath depth, frequency, and signal irregularity, which are then processed through a lightweight binarized neural network (L-BNN) classifier. Deployed on a TinyML microcontroller platform, the architecture demonstrates a classification accuracy of 89.4% in detecting OSA events, with a maximum power consumption of approximately 10 mW and memory utilization of 16.1 KB RAM and 69 KB flash . The system achieves real-time inference latency of 205 ms for ECG-derived respiratory signals and 186 ms for SpO2 data , enabling continuous overnight screening with extended battery life suitable for wearable applications. This research contributes a replicable, computationally efficient framework for edge-AI OSA screening that bridges the gap between clinical-grade accuracy and practical, decentralized healthcare delivery.

• PDF

Copyright (c) 2026 Abiodun Okunola (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.