Project R-14507

Using Machine Learning to Accelerate Computer Simulations for Burn Care (Research)

Background: Adequate burn-care aims at preventing hypertrophic scars and severe contractions that characterize serious burn injuries. To optimize burn-care it is necessary to develop quantitative insight in the underlying biological mechanisms occurring during post-burn skin evolution. In earlier projects, a mathematical foundation for modelling post-burn skin evolution has been laid. This foundation is based on nonlinearly coupled partial differential equations (PDEs), solved using finite elements. Since many input parameters are patient-specific or badly documented, the simulations contain uncertainty. Hence simulation outcomes are only useful in a probabilistic (Bayesian) sense, where one predicts probability distributions over various scenarios. Hence multiple (many) finite element runs are required, which makes the finite element method unattractive in clinical settings. Therefore, we will develop an efficient computational framework that reproduces the expensive finite element simulations. This tool helps clinicians predict probabilities of success for various scenarios and treatments. Hypothesis: Mathematical models can reproduce clinical observations. Machine learning can be used to reproduce finite element simulations, and hence machine learning can be used as a computational tool to predict probabilities of various clinical scenarios (e.g. hypertrophy and contracture) regarding post-burn skin evolution. The principle of machine learning can be extended to various geometries of burn injuries. Probabilities of various scenarios in post-burn skin evolution on the basis of geometrical and patient-specific physiological uncertainties can be predicted. Objective(s): The key objective is the development of a reliable and computationally fast simulator that helps clinicians predict probabilities of success under various treatments and pathological circumstances. Research question(s): - How can we use artificial intelligence to provide the clinic with extremely fast access to finite element simulations without expensive computational platforms? - What is the optimal design for a neural network to replace our PDE-based framework? - How can we directly translate scans of a patients' wounds into initial wound intensity fields and geometries? - Can current clinical datasets be used to train neural networks so that they offer reliable predictions? - To what extent do patient-specific data, e.g. age, determine the intensiveness of contraction, hypertrophy and discomfort? - Which adjustments in the PDE-based model are needed to improve simulation results significantly, without increasing simulation times significantly? - How could the immune system be included in the current formalism? Methods: We use neural networks to reproduce finite element-based simulations and clinical observations, and to reproduce patient-observed data (POSAS), by feeding (training) neural networks with finite element results and additional clinical datasets from anonymized patients. We will study the use of Feed Forward Neural networks, Physics-Informed Networks, Convolutional Neural Networks, and Long Short-Term Memory Networks. Subsequently we will create an app that can be used on a small computing device to quickly estimate the probability of occurrence of certain skin characteristics after trauma. Anticipated results: The research will deliver a fast computing environment to estimate probabilities of various scenarios in post-burn skin under clinical intervention. Further, the translation from scanning data to model input will be made. This will result in an application for a small computing device.

01 January 2023 - 31 December 2026