How can surrogate modelling optimise medical device design?

11 Jun 2025 14min read

Design optimisation is a fundamental step in developing innovative and efficient medical devices.

Before committing to manufacturing prototypes, simulations and models are developed to indicate the expected performance of the medical device design. They offer great precision however, they are computationally expensive and time-consuming to develop, especially for iterative processes like design exploration and optimisation.

As design challenges become more complex, involving numerous constraints and sub-systems, the computational burden of high-fidelity simulations becomes a bottleneck. This approach limits the potential to perform sensitivity analysis – exploring how different designs will affect device performance and the ability to find the best design that balances multiple objectives. Ultimately it will delay progress and increases costs.

Is there a way to reduce the computational cost of medical device design optimisation while maintaining the accuracy needed to make reliable decisions? Surrogate modelling offers a solution.

What is design optimisation for medical devices?

Design optimisation is the process of determining the ideal configuration of a product. It balances constraints, essential features for functionality and costs. In medical device development, this involves iteratively refining key design parameters to meet specific objectives, with the aid of surrogate models.

Creating a medical device is a complex and multifaceted process involving regulatory, usability and mechanical design considerations to name a few. So, how do we innovate in the right direction while managing these demands?

The answer lies in using an engineering design process. This provides a framework to tackle problems and bring medical devices to life, without losing sight of the overall aims.

Figure 1: Diagram showing the engineering design process

Progressing through the engineering design process, the focus shifts to refining solutions. Each option must then be evaluated against the design requirements for performance, regulatory adherence, usability and beyond.

What is surrogate modelling?

Surrogate modelling, also known as metamodeling, is a technique that uses simplified models to mimic the behaviour of more complex and computationally expensive simulations. By acting as surrogates for high-fidelity simulations, these models enable faster evaluations and make large-scale optimisation feasible.

At its core, a surrogate model is a statistical technique that can link complex input and output relationships at a reduced computational cost through black box modelling.

Figure 2: The black box modelling process

Once the surrogate model is made, it can be used as a tool for design optimisation, sensitivity analysis and design space analysis – the defined range of possible design variables.

Benefits of using surrogate models in medical device design

1. Surrogate modelling is particularly valuable for medical device design where competing objectives need to be balanced.

Conflicting objectives may involve minimising size while maximising strength or ensuring durability without compromising biocompatibility for example. By efficiently mapping the design space, surrogate models make it possible to identify optimal trade-offs. This supports multi-objective optimisation, ensuring design decisions are effectively managed and aligned with both clinical and commercial goals thus reducing the number of prototypes manufactured.

Image of a woman with an engineering model on a computer

2. Incorporating surrogate modelling into the medical device design process promotes sustainability and scalability.

By enabling the exploration of environmentally friendly materials and manufacturing processes, engineers can create eco-conscious device designs with minimal waste. Additionally, surrogate models facilitate scalable innovation across production lines by adapting to optimise new designs or variations of existing devices.

3. The ability to predict performance across diverse scenarios makes surrogate modelling a powerful tool for reducing costs and development time.

By minimising the number of physical prototypes and costly simulations required, engineers can streamline the development process while maintaining a focus on meeting requirements. This is particularly beneficial in medical device design, where iterative refinement is crucial to meet regulatory standards alongside end user expectations.

4. Surrogate modelling integrates sensitivity analysis which helps developers to choose the most effective device design.

Sensitivity analysis demonstrates how design variables influence device performance. It allows engineers to identify the most critical factors driving a device’s behaviour, enabling a targeted optimisation approach. This is important in medical devices, where small variations in parameters such as material properties, dimensions or operating conditions can have substantial effects on safety and efficacy.

By leveraging a surrogate model, engineers can conduct sensitivity analyses across a wide range of variables without requiring extensive physical testing or more computationally expensive simulations like Finite Element Analysis (FEA) or Computational Fluid Dynamics (CFD).

In stent development, sensitivity analysis can reveal how changes in strut thickness or material composition affect flexibility and, importantly, the risk of restenosis which is the recurrence of a stenosis or narrowing of a blood vessel, leading to restricted blood flow. The analysis guides refinements to the stent to enhance overall performance.

Similarly, in prosthetic design, it can help determine how adjustments to geometry or material stiffness impact user comfort and durability. The ability to pinpoint critical medical device design drivers reduces development time, enhances robustness and ensures that the final product performs reliably.

Evaluating different surrogate models

There are numerous surrogate modelling techniques, each suited to specific problems. Here, we focus on three commonly used models: Polynomial Response Surfaces (PRSs), Kriging Models and Artificial Neural Networks (ANNs).

Figure 3: Example of surrogate modelling

Figure 3 shows how simple design variables such as material thickness and width can result in demand for a high number of FEA simulations to determine the geometry that is able to withstand the highest load due to the vast design space. Within a simple scenario such as Figure 3, experienced engineers can apply intuition to their designs but multi objective optimisation is still a challenge within complex devices.

1. Polynomial Response Surfaces

Polynomial Response Surfaces approximate a system’s behaviour using polynomial equations. Typically, these models are used for problems with relatively smooth responses, where relationships between input variables (like design parameters) and outputs (like performance measures) can be expressed as low-order polynomials.

Applications of PRS models include:

Predicting stress-strain relationships in materials
Approximation of system dynamics in a mechanical design.

Polynomial Response Surface models are popular due to their simplicity, interpretability and low computational cost. They are well-suited for problems with low to moderate nonlinearity, small design spaces and where insight into variable relationships is valuable. PRS models are easy to implement using standard regression techniques and provide smooth, differentiable approximations, making them ideal for optimisation tasks.

However, PRS models have limitations as they struggle with highly nonlinear or complex problems, are prone to overfitting with high-order polynomials and become impractical for high-dimensional input spaces due to the curse of dimensionality. PRS models also assume smoothness in the response surface (visual representation of the system response to a change in one or more input), making them unsuitable for problems with abrupt changes or discontinuities. Additionally, they struggle to accurately extrapolate outside of the data range they were trained on and are sensitive to input scaling.

2. Kriging Models

Kriging, also known as Gaussian Process Regression (GPR), creates a probabilistic model that predicts system behaviour and provides uncertainty estimates. It is particularly effective for systems with limited data and nonlinearity.

Applications of Kriging Models include:

Optimising stent geometry
Sensitivity analysis in structural mechanics.

Kriging excels in capturing complex, nonlinear relationships in design problems as it can efficiently model underlying function with minimal assumptions. One of its key strengths is its ability to provide not only predictions but also a measure of uncertainty for each output, making it ideal for tasks that require confidence in the model’s reliability, such as optimising geometry for patient-specific solutions. Kriging models are highly flexible, capable of handling small to medium-sized datasets with varying degrees of smoothness and noise.

With stronger ability to handle complex design problems, Kriging brings an increased computational burden. The computational cost rises with the number of input dimensions or data points, making it less suitable for large-scale problems. Kriging requires careful selection of kernel functions (functions that quantify similarity between a pair of data points) and hyperparameters (parameters that control how a model learns) to ensure accurate modelling. It can struggle with very high-dimensional spaces due to the resulting computational burden.

3. Artificial Neural Networks

Artificial Neural Networks are inspired by the human brain and fundamentally consist of layers of interconnected nodes. These models excel in approximating highly nonlinear systems with large datasets.

Applications of ANNs include:

Optimising fluid flow in medical devices
Predicting biological responses in pharmaceutical design.

Artificial Neural Networks make powerful, highly flexible surrogate models. Their primary strength is their ability to capture complex, nonlinear relationships in large and high-dimensional datasets by modelling intricate patterns and interactions. ANNs can also generalise well with sufficient training data, making them suitable for a wide range of applications, especially for large-scale simulations.

The significant amount of data required to effectively train ANNs however, can be computationally intensive, especially with large networks or datasets. They also tend to be less interpretable to understand how input variables influence the output, compared to simpler models like PRSs or Kriging. Like PRS models, ANNs are prone to overfitting if not properly regularised (a form of regression) or if the training data is noisy or insufficient.

When deciding on the best surrogate model type to integrate into the medical device design process, developers should consider their key requirements. ANNs are a valuable tool for dealing with complex systems where traditional models fall short. Whereas PRS models remain a practical choice for early-stage design exploration or simpler engineering challenges. Kriging is useful for engineering optimisation, providing both accuracy and uncertainty quantification in a wide range of applications.

Image of a man with an engineering model on a computer

Integrating surrogate modelling in the medical device design process

Surrogate modelling is a key tool in engineering for the development of medical devices. Ultimately, the complexity and type of problem will dictate the type of surrogate model to be used, and it is worthwhile for developers to explore how surrogate models could enhance their device design process. Surrogate models allow engineers to innovate rapidly while reducing costs, meeting regulatory challenges and maintaining the highest standards of safety and efficacy for end users.