R. C. Cockrell1, G. An1 1University Of Chicago,Surgery,Chicago, IL, USA
Introduction: Personalized medicine requires the right interventions for the right patient at the right time. This necessitates parsing individual patient trajectories at a mechanistically relevant temporal resolution, a task for which existing biomedical data sets are inadequate. High-performance computational modeling and simulation can help dynamically contextualize multi-dimensional data arising from complex systems; however, knowledge of the mechanics of a complex system does not directly lead to the understanding of how to alter these mechanics to a specific end. The combination of high performance simulation and machine learning methods provides a means to identify sets of putative interventions tailored to specific system trajectories. In this study, we examine and assess the efficacy of two methods of developing control strategies for a stochastic dynamical immune system: evolutionary algorithms and artificial neural networks.
Methods: All experiments were performed using a previously validated agent-based model of systemic inflammatory response syndrome. Complete chemokine data for neural net forecasting was collected every seven simulated minutes for approximately 150,000 in silico patients. This data was used to train a convolutional “deep-learning” neural network to determine patient state at a future point in time using a five-point scale. In order to develop interventions using evolutionary algorithms, each intervention was represented as a numerical vector of protein-synthesis augmentations and inhibitions. The genetic algorithm would run until the starting population of interventions converged to a single or small number of near-optimally fit interventions. Intermediate data was saved and used to train a convolutional neural network that produces a near-optimal intervention vector as output.
Results:We show that neural network forecasting models can be used to accurately predict in silico patient disease progression with sufficient data sampling when considering the immune system to be a stochastic dynamical system. Additionally, we show that both evolutionary algorithms and artificial neural networks can be used develop treatment strategies for in silico patients suffering from sepsis. Evolutionary algorithms are directed primarily by characteristics if disease (invasiveness, toxigenesis, virulence, etc.) while artificial neural networks operate on patient state (serum levels of chemokine molecules and their associated rates of change). While both methods show the ability to improve patient state and probability of a positive outcome, this study confirms the conventional wisdom that optimal outcomes can be most widely achieved with personalized solutions.
Conclusion:Machine learning combined with computational simulation of disease provides a viable path towards personalized medicine. Artificial neural networks represent the most efficient strategy to develop interventions that guide the immune system to a desired state.