We also have X and podcasts
Windkessel Physics-Informed Neural Networks Enable Personalized Cardiovascular Digital Twins
A new approach combining physics-based cardiovascular modeling with neural networks delivers accurate, personalized blood pressure predictions from wearable bioimpedance sensors—with minimal training data required.
Background
Accurate, non-invasive blood pressure monitoring remains a clinical challenge. Current wearable technologies struggle to provide personalized accuracy. Windkessel physics-informed neural networks (WPINNs) embed cardiovascular physiology directly into machine learning models, creating a hybrid approach that learns from both data and physical laws.
Key Findings
- WPINNs improved prediction accuracy by 12–25% over conventional data-driven models, achieving 6.49 mmHg error for systolic and 3.99 mmHg for diastolic blood pressure
- Accurately estimated personalized cardiovascular parameters (arterial compliance, peripheral resistance) with errors of 0.77–6.07%
- Physics-driven uncertainty quantification identified unreliable predictions, with statistically significant correlation to actual errors (p<0.001)
- Effective in data-scarce scenarios: required only ~5 samples for diastolic and ~15 samples for systolic BP prediction
Why It Matters
The ability to extract personalized cardiovascular parameters while maintaining high predictive accuracy opens doors to precision medicine applications. WPINNs provide both accurate predictions and clinically interpretable outputs—crucial for clinical decision support. The framework’s efficiency with minimal training data makes it practical for diverse patient populations.
Limitations
The study used relatively small sample sizes (6 and 23 participants) and bioimpedance sensors on specific body locations. Generalization to other wearable modalities and broader populations requires further validation. Testing occurred during controlled physiological maneuvers rather than free-living conditions.
Original paper: Cardiovascular digital twins using a Windkessel physics informed neural network. — NPJ digital medicine. 10.1038/s41746-026-02610-9
