SBIR Phase I: Artificial intelligence platform for secure, collaborative learning across medical institutions

小型企业创新研究(SBIR)第一阶段项目的更广泛影响/商业潜力是开发一个协作学习平台，通过安全访问患者记录生成准确的人口健康模型。拟议的系统将克服隐私方面的担忧，使人工智能方法能够安全地访问美国多个机构的数亿患者记录，以学习高性能的预测模型。从这个平台学到的模型是不同的，因为它们的培训数据和帮助付款人、供应商和制药公司受益于对可能仍未诊断的患者的早期诊断和治疗。这一系统将改善患者的预后和医疗保健系统的性能。这个小型企业创新研究(SBIR)第一阶段项目将解决可能阻碍医疗机构共享患者数据以支持学习临床级模型的基本限制。与联合学习(FL)不同，既不共享局部数据，也不共享局部模型参数；相反，局部分类器预测未标记的全局数据集的标签。在FL中共享模型参数会违反隐私要求，并暴露用于本地培训的患者数据。该平台对FL的“白盒”攻击具有免疫力，其成员关系推理的主要隐私风险显著降低。该平台将准确地包含来自所有子群的信息，并将支持多个ML算法的协作学习，包括无法使用FL学习的人类可解释的算法。准确性将通过敏感度和特异度进行评估，通过会员漏洞进行隐私评估。比较的方法包括联合学习和差异隐私。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

The broader impact / commercial potential of this Small Business Innovation Research (SBIR) Phase I project is to develop a collaborative-learning platform that generates accurate population health models from secure access to patient records. The proposed system will overcome privacy concerns to enable AI methods to securely access hundreds of millions of patient records from multiple institutions across the US to learn high-performing predictive models. Models learned from this platform are differentiated due to their training data and help payors, providers, and pharma companies that benefit from early diagnosis and treatment of patients that may have remained undiagnosed. This system will improve patient outcomes and health care system performance. This Small Business Innovation Research (SBIR) Phase I project will address fundamental limitations that can deter medical institutions from sharing patient data to support learning clinical-grade models. Unlike federated learning (FL), neither local data nor local model parameters are shared; rather, local classifiers predict labels for an unlabeled global dataset. Sharing model parameters in FL can violate privacy requirements and expose patient data used for local training. This novel platform is designed to be immune to the “white box” attacks of FL, and its main privacy risk for membership inference is significantly lower. This platform will accurately include information from all subpopulations and will support collaborative learning for multiple ML algorithms, including human-interpretable algorithms that cannot be learned with FL. Accuracy will be evaluated via sensitivity and specificity, privacy via membership vulnerability. Methods compared include federated learning and differential privacy.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.