From simulations to surrogates: Neural networks enhancing burn wound healing predictions

Burn injuries trigger substantial inflammation, complicating wound healing and potentially leading to severe systemic complications. Understanding the immune response to burns is crucial for improving treatment. Although agent-based models (ABMs) are valuable for studying these interactions, they are computationally demanding. This paper explores the integration of neural networks (NNs) as surrogate models to approximate and forecast ABM simulation results in predicting cytokine concentrations over time and space. We present the development of a baseline ABM using the CompuCell3D software, simulating the innate immune response and generating extensive cytokine concentration data. This data is processed and prepared for neural network training, involving data cleaning, transformation into suitable formats, and a time-series-aware train-test split. We then implement and assess various neural network architectures. Each model is designed to capture the temporal and spatial dynamics of cytokine concentrations, with adjusted model architectures (kernels, number of layers, neurons per layer) to better suit this problem. The models are evaluated using Mean Squared Error, R-squared, and Mean Absolute Percentage Error. In this paper, we assess how different NN architectures (convolutional neural networks (CNNs), long short-term memory (LSTM) neural networks, attention mechanisms, and physics-informed neural networks (PINNs)) predict the concentration of cytokines in this biological system. We find that STA-LSTM generally performs best across statistical metrics.