Multi class photoplethysmography-based deep model for cardiovascular disease classification.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiovascular Diagnostics

Background:

• Cardiovascular disease is a leading global cause of mortality.
• Photoplethysmography (PPG) from wearables offers scalable diagnostic potential.
• Existing PPG methods face limitations in feature extraction, classification scope, and generalizability.

Purpose of the Study:

• To develop a deep learning framework for multi-class cardiovascular diagnostics using raw PPG signals.
• To overcome limitations of handcrafted features and binary classification in prior PPG-based approaches.
• To create a computationally efficient and generalizable model for real-time applications.

Main Methods:

• A deep hierarchical convolutional neural network (CNN) was designed to extract morphological and rhythmic features from PPG signals.
• A dual-stage normalization strategy (Z-score then Min-Max) was employed for signal stability and training efficiency.
• The model was trained and validated on a multi-source dataset of 612 patients across six diagnostic classes (AF, HF, ACS, CVA, DVT, NSR).

Main Results:

• The framework achieved 93.48% overall accuracy, a 0.9386 macro-F1 score, and 0.8968 Cohen's Kappa.
• High performance was observed for atrial fibrillation (AF) and heart failure (HF) with perfect precision and recall.
• The model demonstrated inference speeds under 5ms per segment on consumer hardware, indicating real-time feasibility.

Conclusions:

• The proposed CNN framework offers a balance of representational depth and computational efficiency for cardiovascular diagnostics.
• This study provides a foundation for multi-class PPG-based diagnostics, enhancing clinical decision support.
• Future work should focus on multi-center validation and integrating explainability for broader clinical adoption.