NSF/FDA SIR: A Modeling Tool for Assessment of Radiological Workflow Prioritization Based on Computer-assisted Diagnosis

机器学习（ML）算法的进步支持在医学图像采集的几分钟内诊断和检测各种疾病。这些算法的一个新兴应用是辅助确定具有异常状况的患者的（人）图像审查的优先级。自2018年以来，FDA已经批准了其中几种ML辅助分流设备。这些算法的患者/病例级输出不会取代放射科医师的解释，但它可以通知放射科医师对病例审查的优先级排序，从而通知放射科医师的工作流程。一方面，这种ML算法可以通过增加早期诊断和治疗严重和时间敏感性疾病的可能性来改善患者的预后。另一方面，放射科医生对工作的重新优先级排序可能会延迟对错误分析和遗漏的病例或不在ML算法范围内的病例的审查。虽然这些算法通常在几分钟内处理图像，但由于临床医生工作流程的复杂性，很难辨别整体的、风险调整的、节省患者等待时间的益处。为了提高对ML辅助分诊和它所输入的工作流之间相互作用的理解，本项目开发了灵活的分析和模拟模型，这些模型可以适应临床阅读中的各种ML辅助分诊算法和工作流管理规则。该研究的目的是为监管机构提供一个定量的原则性方法来评估，在工作流程相关的方式，这些分流算法的优先级性能。该项目的结果将使开发人员能够确定最佳的优先级策略，并让用户获得有关这些设备的基于科学的信息，以便他们能够做出明智的医疗保健决策。该奖项反映了NSF的法定使命，并被认为值得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来支持。

Advances in machine learning (ML) algorithms support the diagnosis and detection of various disease conditions within minutes of medical image acquisition. One emerging application of these algorithms is as an aid in determining the priority for (human) image review of a patient with an abnormal condition. Since 2018, the FDA has approved several of these ML-aided triage devices. The patient/case-level output of these algorithms does not replace the interpretation by a radiologist but it can inform the radiologist's prioritization of case reviews and hence the radiologist's workflow. On one hand, such ML algorithms can lead to better patient outcomes by increasing the likelihood for earlier diagnosis and treatment of severe and time-sensitive conditions. On the other hand, the radiologist’s re-prioritization of work can delay the review of cases that are incorrectly analyzed and missed or those not in the scope of the ML algorithm. While the algorithms typically process the images in minutes, the overall, risk-adjusted, patient waiting time-saving benefits are difficult to discern due to the complexity of the clinician's workflow. To improve the understanding of the interplay between the ML-aided triage and the workflow into which it feeds, this project develops flexible analytical and simulation models which accommodate various ML-aided triage algorithms and workflow management rules in clinical reading. The aim of the research is to provide regulators with a quantitative principled approach to evaluate, in a workflow-relevant way, the prioritization performance of these triage algorithms. The results of this project will enable developers to identify optimal prioritization strategies, and let users have science-based information about these devices so that they can make informed health care decisions.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.