Decoding Blood Volume Loss: Can Non-Invasive Signals Match the Gold Standard?

The potential of advanced waveform analysis of physiological signals to diagnose varying levels of hypovolemia has not been fully explored. This study extends previous work by comparing the discriminative ability of invasive and non-invasive signals to classify levels of ongoing hypovolemia using a deep learning (DL) framework. Hypovolemia was simulated via a dynamic lower body negative pressure (LBNP) model among healthy volunteers, allowing for real-life-like fluctuations in intravascular blood volume caused by bleeding and resuscitation. The analysis incorporated photoplethysmography (PPG) as a non-invasive signal and invasive arterial blood pressure (ABP) waveforms, processed using the same DL-based framework. A supervised DL model was developed to perform ternary classification, segmenting the signals and labeling them with corresponding LBNP target levels: Mild (Class 1), Moderate (Class 2), and Severe (Class 3) blood volume loss. The model utilized time-frequency representations of the waveform segments and late fusion of latent space features to enhance classification performance. Results revealed that while both PPG and ABP demonstrated the ability to detect hypovolemia levels, the PPG signal showed comparable or superior performance in classification accuracy with AUROC 0.8861 and F₁ 0.7216, offering a non-invasive alternative with high diagnostic potential.