Collaborative Research: PPoSS: LARGE: Principles and Infrastructure of Extreme Scale Edge Learning for Computational Screening and Surveillance for Health Care

该项目研究了一个全新的跨学科概念“计算筛选和监测（CSS）”，利用边缘学习来检测疾病的早期指标，并监测个人和群体的健康变化。CSS分析和解释人类受试者的连续和异构的物理和生理传感数据流，以产生有关其健康状况的实时信息、知识和见解。该项目的新颖之处在于，它是一种数据驱动的范式，彻底改变了对急性/感染性、慢性生理和心理疾病的理解、预测、干预、治疗和管理。该项目的影响是为个人、组织和医疗保健系统带来巨大的社会和经济效益：早期发现、先发制人的干预和管理可以大大提高医疗质量，并为每年花费数千亿美元的多种疾病节省大量资金。研究人员设计、开发和评估通过极端规模边缘学习实现的CSS原则和解决方案，涵盖四个维度：数据模式、健康状况和数据模式、人工智能/机器学习（AI/ML）算法和模型，以及个人/群体。该设计遵循四个原则：利用规模和异质性，为不确定性设计，作为一等公民的隐私，以及作为规范的错误和攻击。研究人员将1)设计AI/ML算法，用于在极端尺度下学习个人和群体不同健康状况的数据模式和相关性；2)量化安全、隐私保护和学习准确性之间权衡的理论界限，以防止边缘和云上的数据和模型受到各种攻击；3)开发编程抽象，用于在不确定性下自动探索相互竞争的AI/ML方法，以及保护流处理完整性免受敏感数据泄露和错误/恶意分析的系统机制；4)设计神经架构和加速器，以提高约束边缘的计算效率、使用有限训练集的数据效率和使用AutoML的人工效率。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 project investigates a completely new cross-disciplinary concept of “Computational Screening and Surveillance (CSS)” that utilizes edge learning to detect early indicators of diseases, and monitor health changes in both individuals and populations. CSS analyzes and interprets continuous and heterogeneous physical and physiologic sensing-data streams of human subjects to produce real-time information, knowledge, and insights about their health status. The project’s novelty is a data-driven paradigm that revolutionizes the understanding, prediction, intervention, treatment, and management of acute/infectious, chronic physical and psychological diseases. The project’s impacts are enormous social and economic benefits to individuals, organizations, and the healthcare system: early detection, preemptive intervention and management can lead to greatly improved quality of care, and huge savings for multiple diseases each costing hundreds of billions of dollars every year.The investigators design, develop and evaluate principles and solutions for CSS enabled by extreme-scale edge learning spanning four dimensions: data modalities, health conditions and data patterns, Artificial Intelligence/Machine Learning (AI/ML) algorithms and models, and individuals/populations. The design is guided by four principles: exploit scale and heterogeneity, design for uncertainty, privacy as a first-class citizen, and faults and attacks as a norm. The investigators will 1) design AI/ML algorithms for learning data patterns and correlations for diverse health conditions in both individuals and populations at extreme scales; 2) quantify theoretical bounds on the tradeoffs between security, privacy protection, and learning accuracy in order to protect against various attacks on data and models at both the edge and cloud; 3) develop programming abstractions for automated exploration of competing AI/ML methods under uncertainty, and system mechanisms to protect stream processing integrity against sensitive data disclosure and faulty/malicious analytics; and 4) devise neural architectures and accelerators for computation efficiency at the constrained edge, data efficiency using limited training sets, and human efficiency utilizing AutoML.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.