Non-invasive ICP monitoring using deep learning on cerebral blood flow signals from near-infrared photonic sensors.

ProValve Design
ML-driven generative design framework for prosthetic heart valves
Highlights
The Kemerovo Cardiology Center needed a faster, data-driven method to prototype prosthetic heart valves, as traditional CAD/FEM approaches were time-intensive.
Build an ML-optimization pipeline to streamline valve design and explore a vast design space efficiently.
Trained ML models on 11,500+ FEM-evaluated geometries and integrated six optimization algorithms to identify high-performance designs.
Achieved 95% design efficacy, >96% model accuracy (R²), and significantly reduced prototyping time for clinically viable valves.
Core Team
Overview
Valvular heart disease affects millions globally, and prosthetic heart valves (PHVs) are a primary intervention. Traditional methods for PHV design, such as computer-aided design (CAD) and finite element analysis (FEM), are time-intensive and limited in exploring large design spaces.
This project introduces a generative design framework that integrates machine learning models and optimization algorithms to accelerate and improve PHV design. The framework achieves 95% design efficacy, significantly reducing error rates and enabling the generation of new valve geometries tailored to clinical and manufacturing constraints.
Data
The dataset comprised 11,565 prosthetic heart valve designs, each generated using a parametric model with six critical parameters:
- •Parameters: Height (HGT), diameter (DIA), leaflet thickness (THK), curvature (CVT), elevation angle (ANG), and material stiffness (ELM).
- •Simulations: FEM simulations were performed for each design to evaluate mechanical performance under physiological conditions.
- •Output Metrics: Leaflet opening area (LMN) and peak stress (STS), critical indicators of PHV functionality and durability.
The valve structure and parametric model are illustrated in Figure 1.
Methods
The methodology integrates machine learning and optimization algorithms to enable efficient exploration of the design space:
- Parametric Modeling and Simulation: PHV geometries were generated algorithmically using a MATLAB-based parametric model, then analyzed using FEM to simulate mechanical performance under physiological conditions.
- Machine Learning Model Training: Using the dataset of 11,565 FEM-evaluated designs, we trained regression models including Random Forest, LightGBM, and ensemble methods. Models predicted LMN and STS with R² > 96%.
- Optimization Algorithms: Six state-of-the-art algorithms were deployed: Random Search (RS), Tree-structured Parzen Estimator (TPE), CMA-ES (CMA), NSGA-II, MOTPE, and Quasi-Monte Carlo (QMC).
- Hybrid Approach: By combining ML predictions with optimization algorithms, we accelerated the search for optimal designs while reducing reliance on computationally expensive FEM simulations.
The complete generative design workflow integrating parametric modeling, ML predictions, and optimization is shown in Figure 2.
Results
The proposed framework demonstrated significant improvements in PHV design efficiency:
- Model Accuracy: Trained ML models achieved R² values exceeding 96% with MAPE of 11.8% and 10.2% for opening area and peak stress.
- Design Efficacy: TPE and NSGA-II achieved efficacy scores exceeding 95%, identifying superior designs in fewer iterations.
- Parameter Trends: Algorithms favored thinner leaflets and larger diameters, contributing to higher opening areas while maintaining acceptable stress levels.
Examples of optimized valve designs in the open state are shown in Figure 3, showcasing improvements over the initial geometry.
The videos below demonstrate the valve opening behavior for designs generated by the following optimization algorithms:
Conclusion
This generative design framework demonstrates the transformative potential of combining machine learning and optimization for medical device development. It enables faster, more efficient design iterations, significantly reducing the time and cost of prototyping.
Future work will expand this approach to incorporate advanced material modeling and fluid-structure interactions, further refining prosthetic heart valve designs for improved patient outcomes.
Keep Exploring
Deep learning solution for surgical tool segmentation in 3D ultrasound, achieving 93.6% Dice score for real-time catheter localization.






