A new hybrid modeling framework combining biophysics and deep learning to predict and optimize peripheral neuromodulation outcomes in lower urinary tract disease

点击翻译按钮获取中文摘要

Project Summary There is huge potential benefit for peripheral neuromodulation to treat lower urinary tract (LUT) dysfunction through highly targeted interventions. But development and optimization of therapies have been slow, we believe, because we lack the ability to predict the system level, functional response of the LUT to different types and parameterizations of nerve stimulation. Without such an ability, the only recourse is to explore the vast space of possible neuromodulation therapies in animal models, which is slow and expensive. The goal of this project is to invent a predictive model that can assess orders of magnitude more parameterizations through computer simulation, so we can then focus costly experimental efforts on the most promising computationally identified candidates. To achieve this, we will create a framework that unites two powerful modeling approaches: first-principal biophysics models and data-driven deep learning. The biophysics models let us precisely and powerfully represent all the physiology that we understand quantitatively in a way that is both generalizable and understandable. The problem with only using this approach, however, are the many parts of the LUT that we do not understand with this level of confidence and detail. We will insert deep neural networks into the model structure to statistically approximate the less well-understood LUT physiology. We will integrate both approaches together in a single unified hybrid model, and train (tune parameter weights) the entire hybrid model at once with data from cystometry experiments. In this way, we retain the power of biophysics-based models while simultaneously reducing the size (and therefore data requirements) of the neural networks that need to be trained. The neural networks will also be constrained by our LUT physiology knowledge, because they are linked directly with biophysics-based models during simulation and training. We call the framework biomechanistic learning augmentation of deep differential equation representations, or BLADDER. In this project we will first design and validate the BLADDER modeling framework using existing biophysics- based models of LUT organs and training the neural network approximations on data from physiologically nominal cystometry studies. We will then expand the hybrid model’s generalizability and robustness by manipulating the biophysics-based models to allow us to train on data from a wide array of experimental contexts. Finally, we will use the expanded-context model to make predictions about the contributing physiological factors and optimal neuromodulation therapies for underactive bladder syndrome, a highly prevalent LUT dysfunction without adequate treatment options. Our project goal is to develop and validate the BLADDER framework, then use it to make clinically useful predictions for underactive bladder treatment. Our long term goals are to apply the BLADDER approach to many LUT dysfunctions that could benefit from neuromodulation treatments, as well as to other physiological systems.