A New Paradigm for Systems Physiology Modeling: Biomechanistic Learning Augmentation with Deep Differential Equation Representations (BLADDER)

点击翻译按钮获取中文摘要

Many promising peripheral neuromodulation techniques have been proposed to treat lower urinary tract (LUT) dysfunction, but our lack of predictive models has forced the community (including the PI’s lab) to explore the vast parameter space of nerve targets, stimulation parameterizations, and electrode designs empirically in animal experiments by trial and error. This type of exploratory experimentation is the only current method of optimizing, personalizing, or discovering novel LUT neuromodulation techniques. Motivated by this clinical need, our long-term goal for this work is to predict the effects of neuromodulation on the LUT. To move toward this goal, we propose to develop a new modeling framework that integrates disparate biophysics models through machine learning, thereby emulating an entire organ system through a process we call Biomechanistic Learning Augmentation of Deep Differential Equation Representations (BLADDER). We will develop and use the general BLADDER framework to create an organ-level model of the normal healthy LUT throughout its filling and voiding cycles, including non-volitional neural reflex control over the bladder and urethra. Our focus on neural reflex control and organ-level scales ensures that, if successful, the BLADDER LUT model will be poised to predict effects of neuromodulation using computational studies, which so far has been impossible due to the complexity of the LUT. The BLADDER framework unites multiple individual mechanistic models (each accounting for a component function of an organ system) by using deep recurrent neural networks (RNN) to learn the appropriate coupling dynamics linking each component model. The combination of mechanistic and machine learning models under a single framework allows us to harness the advantages of both: mechanistic models excel at interpretability but suffer from a lack of scalability (becoming intractable at the level of organ systems), while machine learning models are excellent at scale but lack generalizability and insights for hypothesis generation. The BLADDER framework will scale up mechanistic models to the level of systems physiology by linking tractable model components together using a supervisory RNN, allowing the BLADDER framework to deliver both interpretability and scale. We will draw on existing SPARC datasets in the cat (e.g., Bruns and Gaunt), existing publicly available data in rat, and generate new data in the rat to construct a training dataset for the supervisory RNN. We will further draw from already published small-scale mechanistic models, validated on human and animal data, for the mechanistic components of the BLADDER LUT model. The formal process of identifying these models and datasets, and checking their validity and robustness, will clearly reveal the deficits and strengths in our theoretical and experimental understanding of the LUT in a straightforward and rational way. We will use the 10 Simple Rules to vet mechanistic models for inclusion in the BLADDER LUT model and compile a public inventory for the neurourology community. Major task 1 (Q1-2): Identify available datasets and candidate mechanistic models from published literature. Major deliverables are a public database and a whitepaper detailing the state of the field and prospects for modeling and experimental work. Major Task 2 (Q1-3): Demonstrate proof of concept of BLADDER framework. Major deliverables are a publicly available code linking two LUT component models via supervisory RNN and a report on suitable RNN architectures based on fully described dynamical systems. Major Task 3 (Q3-6): Create a multi-component BLADDER model. Major deliverables are code used to link separate mechanistic LUT models via the supervisory RNN, and an in vivo rat dataset to fill in critical measurables for the machine learning training set. Major Task 4 (Q6-8): Deploy the fully operational BLADDER model of the LUT, including autonomously predicted neural reflex control. Major deliverables are publicly available codes and datasets, and a hypothesis-driven computational experiment to predict simple interventions.