ACuRE: Accurate Continuity-Regularized SpO2 Estimation Using Liquid Time-Constant Networks

Shahzad Ahmad, Divya Mishra, Sania Bano, Sukalpa Chanda, Yogesh Singh Rawat

TL;DR

This paper introduces ACuRE, a novel framework for non-contact SpO2 estimation from facial videos. By combining advanced AC/DC signal separation, continuous-time modeling with Liquid Time-Constant (LTC) networks, and a physics-informed loss based on mass conservation, ACuRE significantly improves SpO2 accuracy and robustness, especially under challenging motion and illumination conditions.

What Problem Does This Paper Address?

Monitoring blood oxygen saturation (SpO2) is crucial for assessing respiratory and circulatory health, particularly for conditions like COPD and heart failure. Traditional pulse oximeters are invasive and can be uncomfortable, making non-contact methods using RGB cameras a promising alternative, especially for telemedicine and low-resource settings. However, existing non-contact SpO2 estimation methods, primarily based on remote photoplethysmography (rPPG), struggle with motion artifacts, varying illumination, and limited ability to model the complex, continuous-time dynamics of physiological signals. This leads to inaccurate and unstable SpO2 readings in real-world scenarios.

Methods

ACuRE tackles these challenges with a multi-pronged approach:

  • AC/DC Signal Separation: The framework employs a two-branch 3D-ResNet-18 architecture to robustly separate the pulsatile (AC) component, which reflects blood volume changes, from the non-pulsatile (DC) component, representing static tissue absorption. This separation is crucial for isolating the signal of interest and mitigating noise.
  • Liquid Time-Constant (LTC) Networks: To capture the continuous and nonlinear temporal dynamics of physiological signals, ACuRE integrates LTC layers. These networks, based on neuron-specific Ordinary Differential Equations (ODEs) with adaptive time constants, are known for their stability and ability to model fine-grained temporal patterns, outperforming traditional recurrent neural networks.
  • Physics-Informed Loss (PDE Loss): A novel physics-informed loss function, derived from the principle of mass conservation in fluid dynamics, is introduced. This loss regularizes the AC-component features, ensuring their temporal continuity and physical consistency, thereby suppressing flicker and abrupt changes caused by artifacts.
  • Synergistic Design: The core innovation lies in the integration of these components. The AC/DC separation provides clean signals, the 3D-ResNet-18 extracts spatio-temporal features, the LTC layer models continuous dynamics, and the PDE loss enforces physical consistency. This combination aims to create a robust and accurate SpO2 estimation pipeline.

Datasets

The proposed ACuRE framework was evaluated on three publicly available datasets:

  • PURE: Contains 60 face videos from 10 subjects with varied head motions.
  • BH-rPPG: Includes 105 videos from 35 subjects under low, medium, and high lighting conditions.
  • VIPL-HR: Features over 3,000 visible and near-infrared videos from 107 subjects, captured with multiple devices and under various conditions.

Key Insights and Ideas

  • The Power of Continuous-Time Modeling: The paper strongly advocates for continuous-time models like LTC networks for physiological signal estimation. They demonstrate that these models are inherently more stable and expressive for capturing the subtle, dynamic changes in blood volume that drive SpO2 signals, especially when compared to discrete RNNs.
  • Physics as a Regularizer: Incorporating a physics-informed loss (mass conservation PDE) is a key insight. It acts as a powerful regularizer, forcing the learned features to adhere to physical principles, which significantly improves robustness against noise and artifacts that would otherwise disrupt signal continuity.
  • AC/DC Separation is Foundational: The paper reinforces the importance of separating AC and DC components. This initial step is critical for isolating the pulsatile signal that carries SpO2 information and for reducing interference from static or slowly varying illumination changes.
  • Component Synergy is Key: The results clearly show that no single component is a silver bullet. The significant performance gains are achieved through the synergistic combination of AC/DC separation, advanced feature extraction (3D-ResNet-18), continuous-time modeling (LTC), and physics-based regularization (PDE loss).

Key Results

ACuRE demonstrates state-of-the-art performance across multiple datasets:

  • PURE Dataset: Achieved a Mean Absolute Error (MAE) of 0.28 ± 0.17, representing a 44% improvement over baselines.
  • BH-rPPG Dataset: Achieved an MAE of 0.39 ± 0.18, showing robustness under varying lighting conditions.
  • VIPL-HR Dataset: Achieved an MAE of 0.72 ± 0.20, demonstrating generalization across devices and motion.
  • Robustness: The framework shows strong performance even under challenging conditions like dark frames, motion, and illumination variations, outperforming a 3D-CNN+LSTM baseline.
  • Efficiency: While the 3D-ResNet-18 backbone dominates computational cost, the LTC layer adds minimal parameters, and the PDE loss incurs only a small training overhead, making the inference efficient.

Limitations and Future Work

  • Limitations: The paper acknowledges that while ACuRE significantly improves robustness, it still faces challenges with extreme motion or illumination shifts. The performance might also be influenced by individual skin pigmentation and physiological variations not fully captured.
  • Future Work: The authors suggest exploring real-time deployment of ACuRE and conducting extensive clinical validation to further establish its efficacy in practical healthcare settings. They also hint at investigating more advanced architectures and potentially incorporating multi-modal sensor fusion for even greater accuracy.

Summary updated

BibTeX
@inproceedings{ahmad2026acure,
  author = {Shahzad Ahmad and Divya Mishra and Sania Bano and Sukalpa Chanda and Yogesh Singh Rawat},
  title = {{ACuRE: Accurate Continuity-Regularized SpO2 Estimation Using Liquid Time-Constant Networks}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2026},
}
Figure