Multi class photoplethysmography-based deep model for cardiovascular disease classification

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: Overall accuracy 93.48%

• Primary Outcome: Multi-class cardiovascular disease classification accuracy

• Nnt Or Sensitivity: AF and HF: perfect precision and recall (1.000); ACS sensitivity: 0.964

• Confidence Interval: Not reported

Clinical Application

High technical feasibility with <5ms processing time, but clinical integration pathway unclear Potential integration with TGA-approved wearable devices; alignment with RACGP cardiovascular risk assessment guidelines needed; cost-effectiveness analysis required for PBS consideration Patients requiring cardiovascular disease screening using wearable devices, particularly those with suspected arrhythmias or heart failure

Abstract

BACKGROUND: cardiovascular disease is the leading global cause of mortality. Photoplethysmography (PPG), widely embedded in consumer wearables, offers a scalable diagnostic modality. However, prior approaches are often constrained by handcrafted features, binary classification, and poor generalizability, limiting their clinical impact. METHODS: A deep hierarchical convolutional neural network (CNN) was designed to extract both morphological and rhythmic characteristics directly from raw photoplethysmography (PPG) signals. The architecture employs progressively structured convolutional filter hierarchies to capture multi-scale signal features. To enhance signal stability and training efficiency, a dual-stage normalization strategy was implemented, consisting of Z-score standardization followed by Min-Max scaling. In addition, batch normalization and dropout regularization were incorporated to improve model generalization and reduce the risk of overfitting. The proposed framework was trained and evaluated on a multi-source dataset comprising 612 patients and 2,448 annotated PPG segments distributed across six diagnostic classes: atrial fibrillation (AF), heart failure (HF), acute coronary syndrome (ACS), cerebral vascular accident (CVA), deep vein thrombosis (DVT), and normal sinus rhythm (NSR). RESULTS: The model achieved an overall accuracy of 93.48%, a macro-average F1-score of 0.9386, and a Cohen's Kappa of 0.8968, indicating "almost perfect" agreement. AF and HF were detected with flawless precision and recall (1.000), while ACS achieved high sensitivity (recall 0.964). Errors were primarily confined to physiologically related conditions (e.g., ACS vs. CVA). Inference efficiency was demonstrated with <5 ms per segment on consumer-grade hardware, confirming feasibility for real-time applications. CONCLUSION: The proposed framework advances beyond lightweight but underpowered or overly complex models by combining representational depth with computational efficiency. Limitations include the need for external, multi-center validation and explainability integration. This study establishes a robust foundation for PPG-based, multi-class cardiovascular diagnostics, supporting clinical decision support and next-generation wearable health technologies.