Attractor Image-Based Deep Learning of Arterial Pulse Waves for Age Classification

This study investigates whether arterial pulse waveform morphology, which evolves with age and reflects structural and functional changes in the cardiovascular system, can be used to classify chronological age in healthy adults using a deep learning approach. The authors transformed pulse wave time-series data from photoplethysmography (PPG) and arterial tonometry into images using the Symmetric Projection Attractor Reconstruction (SPAR) method, which condenses time-series data into a single density plot image. These SPAR images were used to train a convolutional neural network (CNN) based on a simplified TinyVGG architecture to classify healthy subjects into two closely spaced age groups: 35–40 years and 50–55 years.

Two datasets were used: the Asklepios Study, comprising 2,524 individuals (30–59 years, 52% female) with arterial tonometry waveforms (20 s, 200 Hz), and the Vortal dataset, containing finger PPG recordings from 56 subjects (40 labelled as 'Young' aged 18–35 years and 16 as 'Elderly' aged 70+ years). From the Asklepios population, only participants aged 30-40 and 50-59 years were included, excluding obese subjects (BMI ≥ 30 kg/m2) and those with high blood pressure (systolic ≥ 140 mmHg or diastolic ≥ 90 mmHg). For the Asklepios dataset, one 20-second segment per subject was used, generating one attractor image per participant; for the Vortal dataset, multiple non-overlapping 20-second segments were extracted from the 10-minute recordings, generating around 30 images per subject.

The model was trained on 80% of the Asklepios data from the 35-40 and 50-55 age groups, with the remaining 20% split equally into validation and test sets. Ten-fold cross-validation was performed to improve robustness. The model was then tested on different test sets: the Asklepios 35-40 and 50-55 group, the broader Asklepios 30-40 and 50-59 group, and the Vortal dataset. Performance was evaluated using sensitivity, specificity, and F1 score, with the positive class corresponding to the 50-55 years group.

The model achieved average F1 scores of at least 70%, with sensitivity >67% and specificity >79% across all test sets. Specifically, on the Asklepios 35-40 and 50-55 test set, the F1 score was 70.9 ± 8.6%, sensitivity 67.0 ± 12.3%, and specificity 85.0 ± 6.3%. On the broader Asklepios 30-40 and 50-59 test set, the F1 score was 79.3 ± 2.0%, sensitivity 70.5 ± 3.1%, and specificity 84.3 ± 4.8%. On the Vortal PPG test set, the F1 score was 72.8 ± 2.5%, sensitivity 86.9 ± 5.9%, and specificity 79.0 ± 2.0%. Performance improved with larger test sets and broader age ranges, suggesting robust model generalisation, with higher sensitivity indicating greater accuracy in classifying older subjects.

The authors observed morphological differences in pulse waveforms between age groups: younger individuals display a distinct secondary peak, resulting in attractors with looped edges and closed centres, while older subjects exhibit attenuated or absent secondary peaks, producing more open attractors with reduced looping. The presence of noise in tonometry recordings did not impact the quality of the resulting attractor images compared to those derived from PPG, indicating that the SPAR method is inherently robust to signal noise.

The main limitation is the small size of the training set, mitigated through 10-fold cross-validation. Additionally, detailed selection criteria applied to the Asklepios dataset could not be applied to the Vortal dataset due to missing metadata, so the Vortal dataset may include individuals with high blood pressure or BMI; however, the large age difference between the younger and older Vortal groups supports its suitability as a proof-of-concept PPG test set.

In summary, this study demonstrates that a SPAR-based CNN approach can classify individuals into two closely spaced age ranges within a CVD-free population using both tonometry and PPG signals. The findings support further research into the use of SPAR-transformed PPG signals from community or wearable devices for the early detection and stratification of cardiovascular risk, given the infrastructure demands of gold-standard vascular ageing assessments such as pulse wave velocity.

Improvements to AI Systems:

1. Robust Time-Series-to-Image Encoding for Noisy Physiological Data

• Integrate the Symmetric Projection Attractor Reconstruction (SPAR) method as a preprocessing layer in AI systems that handle raw, noisy time-series signals (e.g., wearable PPG, ECG, or arterial waveforms). This improves the model’s resilience to sensor noise without requiring heavy denoising, as demonstrated by the comparable performance on tonometry and PPG data.

2. Cross-Domain Generalization via Morphological Feature Learning

• Train convolutional neural networks (CNNs) on SPAR images to learn age-related morphological features (e.g., secondary peak attenuation) that are transferable across different sensor types (tonometry vs. PPG). This enables an AI system to classify biological age or cardiovascular risk using diverse input devices, reducing the need for device-specific retraining.

3. Stratified Age Classification with Confidence Calibration

• Use the model’s sensitivity/specificity trade-off (e.g., higher sensitivity for older groups) to build a calibrated classifier that outputs age-group probabilities with confidence intervals. This can be extended to multi-class age bins or continuous age regression, improving risk stratification in preventive healthcare.

4. Data-Efficient Training with Cross-Validation for Small Cohorts

• Adopt the 10-fold cross-validation strategy and the simplified TinyVGG architecture to train robust models on small, well-curated datasets (e.g., <2,500 subjects). This makes the AI system feasible for clinical studies with limited labeled data, while maintaining F1 scores >70%.

5. Automated Cardiovascular Risk Screening from Wearable Signals

• Deploy the SPAR-CNN pipeline as a low-cost, non-invasive screening tool that uses short (20-second) PPG recordings from smartwatches or fingertip sensors to flag individuals with accelerated vascular aging—without requiring expensive pulse wave velocity equipment. The system can output a binary risk score (e.g., “older” vs. “younger” vascular phenotype) for early intervention.

6. Explainable Morphological Biomarker Extraction

• Enhance the AI system with post-hoc interpretability tools that visualize which regions of the SPAR attractor (e.g., looped edges vs. open centers) drive the classification. This allows clinicians to understand the physiological basis (e.g., loss of arterial elasticity) behind the AI’s decision, improving trust and clinical adoption.

7. Adaptive Thresholding for Imbalanced or Heterogeneous Populations

• Incorporate the observed performance improvement with broader age ranges to dynamically adjust classification thresholds based on the target population’s age distribution. This ensures the system remains accurate when applied to community cohorts with mixed health statuses (e.g., including hypertensive or obese individuals, as in the Vortal dataset).

What the Improved AI System Can Do:

• Screen for early cardiovascular aging using just 20 seconds of fingertip PPG or tonometry data, with >70% F1 score and >79% specificity, even in noisy, real-world recordings.

• Transfer across devices (e.g., from clinical tonometry to consumer wearables) without retraining, due to SPAR’s noise robustness and morphological feature invariance.

• Provide explainable risk assessments by highlighting specific waveform morphology changes (e.g., secondary peak disappearance) that correlate with age-related vascular stiffening.

• Operate with minimal computational resources (TinyVGG architecture) on edge devices, enabling real-time, on-device health monitoring.

• Adapt to different age cohorts (e.g., 30–59 years) and health statuses, making it suitable for large-scale community screening or longitudinal studies where gold-standard vascular assessments are impractical.

Arterial pulse waveform morphology evolves with age, reflecting structural and functional changes in the cardiovascular system. Thus, vascular age is a valuable surrogate marker of cardiovascular health, and premature vascular ageing can indicate increased disease risk. Pulse wave analysis could support risk stratification in otherwise asymptomatic adults. We transformed pulse wave time-series data from photoplethysmography (PPG) and arterial tonometry into images, using the Symmetric Projection Attractor Reconstruction (SPAR) method. These SPAR images were used to train a convolutional neural network to classify healthy subjects into two closely spaced age groups (35-40 and 50-55 years). The model demonstrated consistent classification performance across internal and external test sets, achieving F1 scores above 70% for both PPG and tonometry signals. These results suggest that SPAR-derived pulse wave images contain discriminative morphological features even among healthy adults close in age. This proof-of-concept lays the groundwork for future research into the use of SPAR for early risk detection using smart wearables.