Physics-Informed Neural Networks (PINNs) for Interpretable Parametric Estimation of SpO₂ and Respiratory Waveforms in Sleep Apnea Diagnostics

• Abiodun Okunola

  Ladoke Akintola University Technology

  Author

due to the high cost, complexity, and limited accessibility of polysomnography (PSG), the current diagnostic gold standard . While machine learning approaches using SpO₂ and respiratory signals have shown promise for automated screening, existing methods face critical limitations: they operate as black-box systems lacking clinical interpretability, fail to incorporate known physiological principles, and cannot reliably estimate the continuous parameters needed for accurate apnea-hypopnea index (AHI) calculation. This study addresses these gaps by developing a Physics-Informed Neural Network (PINN) framework that integrates physiological governing equations—specifically respiratory mechanics models and oxygen desaturation dynamics—directly into the neural network training process through physics-based loss functions. The proposed framework achieved an overall classification accuracy of 89.4% for sleep apnea detection across validation cohorts, with a sensitivity of 87.2% and specificity of 91.1%. The parametric estimation capability enabled continuous waveform reconstruction with a mean squared error of 0.043 between predicted and actual SpO₂ signals. The primary contribution is a replicable, interpretable framework that provides clinicians with traceable decision-making through physics-grounded attention mechanisms, potentially enabling cost-effective, large-scale sleep apnea screening while maintaining diagnostic rigor. Practical implications include deployment in home-based monitoring settings and integration with existing clinical workflows for preliminary risk stratification.

• PDF

Copyright (c) 2026 Abiodun Okunola (Author)

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