CAREER:Real-Time Nonparametric Machine Learning for Healthcare with Guarantees

这个NSF CAREER项目是为医疗保健开发新的实时机器学习模型。一个关键的重点是预测个体患者水平的持续时间，例如，到死亡的时间，到疾病恢复的时间，以及住院时间。所开发的模型不仅将利用深度学习的最新进展，而且还将具有严格的准确性保证，说明何时，为什么以及如何工作。这些模型是“非参数的”，因为它们保证在对底层数据的很少假设下工作。此外，这些模型将用于辅助医疗运输和临床检查室调度系统，这些系统依赖于预测患者在临床就诊的不同阶段花费的时间。该研究具有使患者护理更加个性化和提高医院资源利用效率的潜力。在进行这项研究时，大部分工作不仅是将机器学习和医疗保健社区聚集在一起，而且还为这些社区提供教育资源（研讨会、教程和一本新书），这些资源将更好地塑造机器学习和医疗保健交叉的相对较新的研究领域的数学和统计基础。为了在这个交叉点发展社区，PI计划教授一门关于医疗保健机器学习的新课程，并继续指导本科生和研究生的学生。所吸取的经验教训将为PI在高中和大学水平的推广工作提供信息，将计算机科学和概率概念带给不同的受众。尽管近年来机器学习方法取得了许多进步，但在临床环境中部署的机器学习模型很少。少数几家公司往往有几十年的历史，或者利用深度学习在成像应用中的成功案例。然而，许多医疗保健预测任务仍然具有挑战性，其中最先进的模型（基于深度学习或非深度学习）难以产生准确的预测。例如，许多生存分析问题（预测时间到事件的结果，如死亡时间、出院时间等）很难。更复杂的是，在许多这类问题中，收集数据用于对真实患者进行培训或验证可能成本高昂，并且需要进行长期研究（例如，为了预测癌症患者的死亡时间，这些持续时间可能以年为时间尺度）。为了帮助从业者决定使用哪个模型，以预测模型的统计保证形式提供的可靠性保证将非常有价值。此外，为了使这些模型能够协助时间敏感的高风险决策，它们必须扩展到实时临床数据流。该项目旨在为医疗保健开发一系列实时机器学习模型，这些模型不仅利用了深度神经网络和树集成中最先进的机器学习进展，而且还具有严格的准确性和不确定性保证。即使模型没有对数据的分布做出参数假设，这些保证仍然成立。此外，开发的模型将在临床环境中进行测试，以调度医院资源的使用，最终目标是在医院系统中部署。重点是生存分析问题，这在医疗保健中非常常见，但在机器学习社区中鲜为人知。拟议的研究与在机器学习和医疗保健交叉领域教授关键概念的教育计划紧密结合。这包括通过研讨会和教程使机器学习算法的统计保证更容易被从业者访问和使用，通过一本新书向普通机器学习受众介绍生存分析，并通过开发新课程向本科和研究生层次的学生教授医疗保健机器学习的基本概念。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

This NSF CAREER project is on developing new real-time machine learning models for healthcare. A key emphasis is on predicting durations at the individual patient level such as, time until death, time until recovering from a disease, and hospital length of stay. The models developed will not only take advantage of recent advances in deep learning but will also come with rigorous accuracy guarantees that say when, why, and how well they work. These models are "nonparametric" in that they are guaranteed to work under very few assumptions on the underlying data. Moreover, these models will be used to assist medical transport and clinical exam room scheduling systems that rely on predictions of how long patients spend in different stages of their clinical visits. This research has the potential to make patient care more personalized and to improve the efficiency of hospitals' resource use. In conducting this research, much of the work is in not only bringing together machine learning and healthcare communities, but also providing these communities with educational resources (workshops, tutorials, and a new book) that will better shape the mathematical and statistical foundations of the relatively new field of study at the intersection of machine learning and healthcare. To grow the community at this intersection, the PI plans to teach a new course on machine learning for healthcare and continue mentoring students at both undergraduate and graduate levels. Lessons learned will inform the PI's outreach efforts at the high school and college levels that bring computer science and probability concepts to a diverse audience.Despite numerous machine learning methodological advances in recent years, few machine learning models are deployed in clinical settings. The few that are, tend to be many-decades-old or capitalize on deep learning success stories in imaging applications. However, many healthcare prediction tasks remain challenging, where state-of-the-art models (deep-learning-based or not) struggle to produce accurate predictions. For example, many survival analysis problems (predicting time-to-event outcomes such as time until death, time until hospital discharge, etc) are hard. To complicate matters, in many such problems, collecting data for training or validation on real patients could be costly and require long-term studies (e.g., for predicting time until death for cancer patients, these durations could be on the time scale of years). To help practitioners decide on which models to use, reliability assurances in the form of statistical guarantees on prediction models would be extremely valuable. Moreover, for these models to assist in time-sensitive high-stakes decisions, they must scale to real-time clinical data streams. This project aims to develop a family of real-time machine learning models for healthcare that not only take advantage of state-of-the-art machine learning advances in deep neural networks and tree ensembles, but also come with rigorous accuracy and uncertainty guarantees. These guarantees hold even though the models do not make parametric assumptions on the distribution of the data. Moreover, the models developed will be tested in clinical settings for scheduling hospital resource use with the ultimate goal of deployment in hospital systems. A heavy emphasis is on survival analysis problems, which are extremely common in healthcare but less well-known within the machine learning community. The proposed research tightly integrates with an education plan of teaching key concepts at the intersection of machine learning and healthcare. This includes making statistical guarantees of machine learning algorithms more accessible and usable by practitioners through workshops and tutorials, introducing survival analysis to a general machine learning audience with a new book, and teaching students at the undergraduate and graduate level fundamental concepts of machine learning for healthcare by developing a new course.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.