Mechanistic Machine Learning

项目摘要/摘要 该项目的目标是将经验数据与机械生理学知识相结合，以产生 个性化、量化的预测可以带来更好的治疗。在正常情况下，医生 从一般生理学原理进行类比推理，但这项技术的存在甚至是为了开发不完善的东西 生理学模型使治疗个性化和量化地以生理学为基础，并改进 从经验数据中学习。我们将应用数据同化(DA)、机械数学建模、 机器学习和控制理论，它们彻底改变了太空旅行、天气预报、 运输和飞行，以及制造业。数据同化和控制理论的应用非常有限。 在医学上，通常应用于数据丰富的环境，如连续血糖监测或包装机。 我们以前的工作演示了使用数据同化与葡萄糖-胰岛素模型来预测血糖在 门诊2型糖尿病的设置。我们将扩展数据同化和控制理论，例如， 约束集成卡尔曼滤波和离线马尔可夫链蒙特卡罗算法，以更好地处理 针对快速变化的患者的稀疏、简短的训练集，我们将其应用于血糖设置 重症监护病房(ICU)的管理。此外，我们将通过以下方式开发用于表型鉴定应用的DA 利用DA的参数估计能力。数据同化可以用来估计可测量的 和不可测量的生理状态和参数，我们将使用这些估计来创造更高的 定义表型。虽然我们专注于ICU中的血糖管理，但我们将开发出 很可能推广，开始努力在更广泛的医疗保健背景下发展发展议程。这项工作 我们提出，朝着能够使用机制驱动的DA进行测试、验证和优化的方向迈出了必要的一步 个性化的短期治疗策略、长期健康预测和机械生理学 理解。 我们将实现以下目标：目标1--预测--扩展发展议程方法以实现预测， ICU背景下的个性化、模型评估和模型选择，将治疗输入与 生理结果；目标2-表型-将DA框架扩展到状态和参数估计，以 考虑到基于机制的表型，仔细的不确定性量化，以及对困难或 无法测量的生理学；目标3-控制-扩展DA以包括以 预期的临床结果，例如血糖范围，并估计需要的投入，例如胰岛素或营养，以 实现成果。

PROJECT SUMMARY / ABSTRACT The goal of this project is to combine empirical data with mechanistic physiologic knowledge to produce personalized, quantitative predictions that can lead to improved treatments. In normal practice, physicians reason by analogy from generic physiologic principles, but the technology exists to exploit even imperfect physiologic models make treatment personalized and quantitatively grounded in physiology, and to improve learning from empirical data. We will apply data assimilation (DA), mechanistic mathematical modeling, machine learning, and control theory, which have revolutionized space travel, weather forecasting, transportation and flight, and manufacturing. Data assimilation and control theory have seen very limited use in medicine, usually applied in data-rich circumstances like continuous glucose monitoring or packemakers. Our previous work demonstrated use of data assimilation with glucose-insulin models to predict glucose in the outpatient type 2 diabetes setting. We will extend data assimilation and control theory using, for example, a constrained ensemble Kalman filter and an offline Markov Chain Monte Carlo algorithm, to better handle sparse, short training sets on rapidly changing patients, and we will apply it in the setting of glucose management in the intensive care unit (ICU). Moreover, we will develop DA for phenotyping applications by exploiting the parameter estimation capabilities of DA. Data assimilation can be used to estimate measureable and unmeasureable physiologic states and parameters, and we will use these estimates to create higher definition phenotypes. While we are focusing on glucose management in the ICU, we will develop methods that are likely to generalize, beginning the effort to develop DA in the context of healthcare more broadly. The work we propose is a necessary step toward being able to use mechanism-driven DA to test, validate and optimize personalized short-term treatment strategies, long-term health forecasts, and mechanistic physiologic understanding. We will carry out the following aims: AIM 1—forecast—extend the DA methodology to allow forecasting, personalization, model evaluation, and model selection in the ICU context, relating treatment input to physiologic outcome; AIM 2—phenotype—extend the DA framework to state and parameter estimation to allow for mechanism-based phenotyping, careful uncertainty quantification, and inference of difficult or impossible-to-measure physiology; AIM 3—control—extend the DA to include a controller that begins with desired clinical outcomes, e.g., glucose range, and estimates the inputs, e.g., insulin or nutrition, required to achieve the outcomes.