Amplitude-Independent Machine Learning for PPG through Visibility Graphs and Transfer Learning

TL;DR

The paper introduces VGTL-net, a framework that makes PPG analysis amplitude-independent and affine-invariant by converting time-series signals into visibility graphs and treating their adjacency matrices as RGB images processed by pretrained CNN backbones. By leveraging transfer learning and minimal preprocessing, VGTL-net demonstrates strong generalization across BP waveform estimation and vascular ageing tasks, achieving state-of-the-art or competitive results while handling noisy data through robustness-enhancing augmentation. The approach provides interpretable connections between PPG structure and physiological features (breathing, HR, ageing) via the visibility graph, and shows promise as a universal, low-overhead tool for wearable-signal analysis. Overall, VGTL-net advances PPG analytics by combining graph-theoretic representations with modern computer vision, enabling efficient cross-task applicability and robustness to data quality issues.

Abstract

Photoplethysmography (PPG) refers to the measurement of variations in blood volume using light and is a feature of most wearable devices. The PPG signals provide insight into the body's circulatory system and can be employed to extract various bio-features, such as heart rate and vascular ageing. Although several algorithms have been proposed for this purpose, many exhibit limitations, including heavy reliance on human calibration, high signal quality requirements, and a lack of generalisation. In this paper, we introduce a PPG signal processing framework that integrates graph theory and computer vision algorithms, to provide an analysis framework which is amplitude-independent and invariant to affine transformations. It also requires minimal preprocessing, fuses information through RGB channels and exhibits robust generalisation across tasks and datasets. The proposed VGTL-net achieves state-of-the-art performance in the prediction of vascular ageing and demonstrates robust estimation of continuous blood pressure waveforms.

Figures (17)

• Figure 1: An example graph, with nodes which are indexed and form a circle. The lines connecting the nodes are the edges of the graph.
• Figure 2: An example of a visibility graph. (a): A PPG pulse and the red lines designate for the visibility of the signal sample 9. (b): The generated visibility graph and the red lines are the edges generated by the visibility of the signal sample 9.
• Figure 3: Pathway from a time series to an image. (top left): The original signal and the visibility between signal samples. (top right): The visibility graph generated from the input signal. (bottom right): The adjacency matrix which corresponds to the visibility graph. (bottom left): The black-and-white image generated from the adjacency matrix.
• Figure 4: (a): Adjacency matrices of the visibility graphs of the original PPG pulse. (b): The PPG pulse under moderate affine transformation. (c): The PPG pulse under heavy affine transformation. Although the PPG pulse becomes unidentifiable after affine transformations, the visibility graphs and the adjacency matrices remain the same, showing the visibility graph to be affine invariant.
• Figure 5: Example PPG waveforms and the corresponding adjacency matrices. (a): Normal breathing PPG signal and the adjacency matrix; (b): Heavy breathing PPG signal and the adjacency matrix. Red lines connect the signal sample with the highest visibility and the samples can be seen by it. The red pixels in the adjacency matrices represent edges introduced by the red lines. The baseline wandering introduced by heavy breathing gives some signal samples more visibility, thus referring to more red lines and broader wing-shape elements in the adjacency matrix.