Abstract
The current study employs state-of-the-art optimisation methods for estimation of unknown parameters in a mathematical model of highly non-linear partial differential equations describing drug delivery from a drug-eluding stent. A classical optimisation scheme entails enormous run times due to the need to numerically solve the computationally expensive equations a large number of times to obtain the objective (black-box) function. We address this issue by employing an efficient global optimisation scheme, i.e. Bayesian optimisation (BO). This scheme aims to find the optimum of the black-box function by using an emulator of the original objective function to select the next query point (while balancing exploration and exploitation), and sequentially refining the emulator. Additionally, the proposed optimisation scheme is adapted to scenarios where there are hidden constraints in parameter space by incorporating a classifier that learns the infeasible parameter domains. We demonstrate that given a fixed number of expensive mathematical model evaluations, the proposed BO scheme outperforms state-of-the-art classical optimisation methods in terms of accuracy.
Introduction: The Stent Design Challenge
Drug-eluting stents (DES) are critical in treating heart disease, releasing medication directly into arterial walls to prevent re-blockage. However, designing the perfect stent is a "Goldilocks" problem: release too little drug, and the treatment fails (low efficacy); release too much or too fast, and the tissue becomes damaged (toxicity).
Engineers use complex mathematical models—specifically nonlinear partial differential equations (PDEs)—to simulate this process. The primary challenge is computational cost: a single simulation can take 1.5 hours to run. Traditional optimization methods, such as Sequential Quadratic Programming (SQP), require hundreds of these simulations to find an ideal design, making the process prohibitively slow. Furthermore, some designs are "hidden" failures—the toxic threshold might be crossed in a way that is only detectable after the expensive simulation has concluded.
Modeling and Problem Definition
The study utilizes a 2D-axisymmetric model representing a stent strut embedded in an arterial wall. The optimization focuses on three key design variables: initial drug loading ($m_{drug}$), coating diffusion coefficient ($D_{coat}$), and coating thickness ($h_{coat}$).
As shown in the figure above, the model tracks receptor saturation ($Rs2$) and drug concentration ($DC$). The goal is to maximize receptor saturation—ensuring the drug actually binds to target cells—while maintaining drug concentrations within safe, non-toxic limits.
Methodology: Bayesian Optimization with Hidden Constraints
To overcome the computational bottleneck, the researchers propose a Bayesian Optimization (BO) framework. Unlike classical methods that rely on local gradient estimations, BO builds a global understanding of the parameter space.
Gaussian Process Emulators
Instead of running the full physics simulation for every potential design, the framework constructs a "surrogate model" using Gaussian Process (GP) regression. This emulator learns to predict the simulation's output based on previous evaluations, providing a computationally "cheap" way to explore the design space.
Handling Safety Constraints
To manage toxicity, a second GP is trained as a classifier. This classifier learns the boundaries of the "safe zone" where drug levels remain below the toxic threshold. The researchers then employed a modified acquisition function, $u_{HCW}$, which weights the search for high performance by the probability that a design is actually safe. This ensures the optimizer avoids wasting time in high-performance but toxic regions.
Key Results and Insights
The BO scheme was benchmarked against classical SQP with a strict budget of only 90 model evaluations.
- Efficiency and Accuracy: BO reached significantly lower (better) objective function values faster than the classical baseline. While SQP often plateaued or became trapped in local minima due to the high cost of numerical derivative approximation, BO consistently improved.
- Convergence: Across five separate runs starting from different initial designs, the BO framework converged to a similar, high-performing optimum (objective value ~0.046).
- Optimal Design Characteristics: The results favored a slow-release coating with moderate drug loading. This finding aligns with clinical literature suggesting that sustained, steady release is superior for long-term arterial healing compared to rapid "burst" releases.
Conclusion and Future Directions
By replacing slow, classical optimization with an intelligent Bayesian framework, this research demonstrates that complex medical device design can be significantly accelerated. The ability to learn "no-go" zones on the fly allows the system to navigate safety constraints more efficiently than simple penalty-based methods.
Future work will look toward full Bayesian uncertainty quantification to map how sensitive designs are to manufacturing variations. Additionally, moving from 2D-axisymmetric models to 3D multi-layer arterial models will further improve the clinical relevance of these optimized designs.
Summary updated