SBIR Phase I: Driving Timely Point-of-Care Treatment in Hospitals with a High Precision Bayesian Machine Learning Platform

这一小型企业创新研究（SBIR）I期项目的更广泛影响/商业潜力是提供临床决策支持软件，帮助住院提供者改善对可预防的急性住院伤害的护理，从而降低死亡率和发病率。该资助开发了一个基于云的平台，该平台将机器学习（ML）算法实时应用于从电子健康记录（EHR）和附在患者身上的生理监测设备中提取的数据。所采用的ML工具估计每个数据元素的可靠性程度，因为它们被收集并整合这些信号，以提供随着时间的推移对患者健康的准确，个性化的风险估计，以便最好地指导患者治疗和医院资源的分配。我们最初的目标是败血症，这是医院中最昂贵和最致命的疾病之一。该赠款开发了一个端到端系统，以提供风险评估和及时治疗的实施。对于商业潜力，底层核心技术可以扩展到其他临床场景。拟议的项目能够扩展高精度最先进的贝叶斯机器学习技术，预测急性恶化的机会。这包括解决扩展这个机器学习系统以在许多护理提供者，患者和医院中发挥作用的挑战。为了实现这些目标，该项目将开发在云计算环境中以分布式方式运行机器学习算法的新方法，特别是在区分多台机器需要协调的地方，以及可以说更重要的是，它们可以避免在数据训练中协调。此外，该项目还开发了软件，向提供者提供信息，以便能够进行可以改变患者轨迹的干预。在这里，该软件将包括如何最好地使用由此产生的推论指导护理。

The broader impact/commercial potential of this Small Business Innovation Research (SBIR) Phase I project is to provide clinical decision support software that assists inpatient providers in improving care for preventable acute inpatient harms, and thereby reduce mortality and morbidity. This grant develops a cloud-based platform that applies machine learning (ML) algorithms in real-time on data extracted from Electronic Health Records (EHRs) and physiologic monitoring devices attached to a patient. The ML tools employed estimate the degree of reliability for each of the data elements as they are collected and integrates these signals to provide an accurate, individualized risk estimate of patient health over time in order to best guide patient treatment and allocation of hospital resources. Our initial target condition is sepsis, one of the most costly and most deadly diseases in hospitals. This grant develops an end-to-end system to provide risk assessment and implementation of timely treatment. For commercial potential, the underlying core technology can be extended to other clinical scenarios. The proposed project enables scaling of high-precision state-of-the-art Bayesian machine learning techniques that forecast the chance of acute deterioration. This includes tackling the challenges in scaling this machine learning system to function across many care providers, patients, and hospitals. To achieve these goals, this project will develop new methods for running machine learning algorithms in a distributed fashion in cloud computing settings, especially in distinguishing where multiple machines need to coordinate, and arguably more importantly, where they can avoid coordinating in training on data. Further, the project develops software to provide information back to providers so as to enable interventions that can alter patient trajectory. Here the software will encompass how to best use the resulting inferences in guiding care.