Abstract
This paper introduces VGTL-net, a novel framework designed to overcome the inherent limitations of photoplethysmography (PPG) signal processing in wearable health technology. By transforming 1D PPG signals into 2D Visibility Graphs and subsequently into adjacency matrix images, the proposed method achieves complete amplitude independence. This approach allows for the application of advanced transfer learning from computer vision models, resulting in state-of-the-art performance in predicting vascular age and blood pressure without the need for extensive manual feature engineering.
Introduction
Photoplethysmography (PPG) is the foundational technology for heart rate and blood oxygen monitoring in modern wearables. Despite its ubiquity, PPG-based machine learning models face significant hurdles. Primarily, signal amplitude is highly sensitive to external factors such as skin tone, age, and sensor contact pressure, leading to poor generalization across diverse populations. Furthermore, traditional methods often require heavy preprocessing and manual feature extraction, which can be computationally expensive and prone to error.
The authors address these challenges by shifting the focus from numerical signal values to the underlying structural geometry of the pulse. By utilizing Visibility Graphs, the model becomes robust against amplitude variations, focusing instead on the topological features of the cardiac cycle.
Methodology
The VGTL-net Pipeline
The core of the VGTL-net framework is a multi-step process that converts temporal data into a visual format suitable for deep learning architectures.
- Visibility Graphs (VG): The 1D PPG signal is mapped onto a graph where each data point serves as a node. An edge is established between two nodes if they possess a "line of sight"—that is, if a straight line connecting them does not intersect the signal curve. This transformation ensures affine invariance; stretching or shifting the signal vertically does not alter the connectivity of the graph.
- Adjacency Matrices as Images: The resulting graphs are represented as 2D adjacency matrices. This step effectively recontextualizes a signal processing task as a computer vision problem, allowing the model to interpret physiological patterns as visual textures.
- RGB Fusion: To maximize information density, the researchers developed a three-channel "color" image representation. The Red channel contains the adjacency matrix of the original PPG, the Green channel represents the amplitude-inverted signal, and the Blue channel incorporates a slope-weighted version of the inverted signal. This fusion captures complex geometric relationships that a single channel might miss.
- Transfer Learning: Rather than training a new architecture from scratch, the framework utilizes pre-trained models such as VGG19, ConvNeXt v2, and Coatnet. These models, originally trained on the ImageNet dataset, are adept at identifying the subtle "textures" and edges within the adjacency matrices that correspond to different cardiovascular states.
Experimental Setup and Datasets
The efficacy of VGTL-net was evaluated using two distinct public datasets:
- Blood Pressure Estimation: The UCI/MIMIC II dataset provided a combination of PPG, ECG, and invasive arterial blood pressure (ABP) signals for rigorous testing.
- Vascular Aging: The Mendeley Real-World PPG Dataset was used to classify subjects into four age groups (ranging from 10 to 75 years) based on signal morphology.
Key Results and Insights
The experimental results demonstrate the superiority of the graph-based approach over traditional 1D models:
- Superior Accuracy: For vascular aging, VGTL-net achieved a 97.2% accuracy rate, a significant leap from previous benchmarks of approximately 87%.
- Clinical Precision: In blood pressure estimation, the model recorded a Mean Absolute Error (MAE) of 3.11 mmHg for Systolic and 1.94 mmHg for Diastolic pressure. These results meet the British Hypertension Society's "Grade A" standard, indicating high potential for clinical utility.
- Physiological Visualization: The adjacency matrices provide a unique way to visualize arterial health. For instance, younger subjects typically exhibit "larger wings" in their matrices, reflecting the sharper dicrotic notches and peaks associated with more elastic arteries.
Conclusion and Future Work
VGTL-net represents a paradigm shift in PPG analysis by prioritizing structural topology over raw amplitude. While the results are promising, the authors acknowledge that future work must address the impact of motion artifacts in real-world settings and the computational demands of running large computer vision models on low-power wearable devices. Nevertheless, this "signal-to-graph-to-image" framework offers a robust, universal architecture that could potentially be extended to other biological signals like EEG or EMG.
Summary updated