Quantitative Estimation of Sensitivity of Lipolysis to Insulin

肥胖是一个持续存在的问题。了解全身性炎症与胰岛素抵抗之间的关系及其在胰岛素调节游离脂肪酸和葡萄糖中的作用是必要的。

Obesity is a continuing problem. Understanding the relationship between systemic inflammation and insulin resistance and its role in the regulation of FFA and glucose by insulin is essential. Background: Quantitative evaluation of insulin regulation on plasma glucose and free fatty acid (FFA) in response to external glucose challenge is clinically important to assess the development of insulin resistance. Mathematical minimal models (MMs) based on insulin modified frequently-sampled intravenous glucose tolerance tests (IM-FSIGT) are widely applied to ascertain an insulin sensitivity index. Our previous work has extended MMs to include the dynamics of FFA. This FFA MM provides a useful index for the sensitivity of lipolysis to insulin, but the dynamics of FFA much after the insulin bolus are not well-accounted for because of the model's minimality. Objective: We have two objectives. (1) To develop a deep-learning methodology for determining parameters in physiological ranges for MMs with parameters that enter in a nonlinear manner; and (2) To develop an FFA MM that adequately fits data during long duration (>4 hour FSIGT) protocols. Methods: We are training deep learning convolutional neural networks trained on hypothetical model predictions made with parameters in known physiological ranges. We use Gaussian process regression to select very large numbers of suitable joint hypothetical datasets with FFA, insulin, glucose. We evaluate features in the datasets that improve neural network learning performance in learning the map from data to parameters. Results: Thus far we have found that it is necessary for this machine learning approach to physiological model parameter determination to carefully select appropriate features constructed from the original dataset. We found that neural network training requires parameters that are well-determined by the selected feature set constructed from the data, even though some of these parameters are not directly evident in the model description. Outlook: Work continues in constructing a general modeling framework for how to train a deep learning neural network to go from a hypothetical model to direct parameter determination without involving optimization algorithms. We continue work on developing an FFA MM that addresses long-time behavior of FFA dynamics in an FSIGT as a test case of our deep learning approach.