VAM: A Parallel Cross-Modal Hybrid Network for Accurate and Interpretable Vascular Age Estimation from PPG.

Area of Science:

• Biomedical Engineering
• Cardiovascular Research
• Artificial Intelligence in Medicine

Background:

• Vascular age is a critical clinical marker for assessing vascular health.
• Traditional vascular age assessment methods (e.g., PWV) are complex.
• Photoplethysmography (PPG)-based methods offer convenience but lack generalization.

Purpose of the Study:

• To develop an accurate and adaptable Vascular Age Estimation Model (VAM).
• To integrate multi-source data using a hybrid deep learning architecture.
• To improve upon existing vascular age estimation techniques.

Main Methods:

• Proposed a parallel cross-modal hybrid network (VAM) integrating CNN and Transformer architectures.
• Utilized multi-source datasets for training and validation.
• Conducted experiments on the VitalDB dataset and an external test set.

Main Results:

• VAM achieved a Mean Absolute Error (MAE) of 7.66±0.39 years and R² of 0.55±0.04 on the VitalDB dataset.
• The model demonstrated at least a 20% reduction in prediction errors compared to baseline models.
• VAM showed strong performance in younger populations (MAE=5.85 years, R²=0.63) with minimal bias from gender and BMI.

Conclusions:

• The proposed VAM offers an efficient and interpretable solution for vascular age assessment.
• The model's accuracy and adaptability highlight its clinical application potential.
• Interpretability analysis identified the diastolic peak as a key feature for vascular age estimation.