Human-AI Collaborations to Improve Accuracy and Mitigate Bias in Acute Dyspnea Diagnosis

项目摘要 急性呼吸困难（呼吸短促）是急诊科就诊的最常见原因之一， 每年住院。心力衰竭、肺炎、慢性阻塞性肺疾病最多 常见病因，代表2017年美国250万例住院治疗。确定确切的原因 急性呼吸困难是非常重要的，但具有挑战性，因为表现出的症状，实验室检查， 成像结果可能难以解释，特别是在老年人和合并症患者中， 严重的疾病。高达30%的患者可能发生诊断错误和不适当的治疗， 与更差的患者结局相关。人工智能（AI）工具已经被提出来增强 提供者在诊断过程中，并处于有利地位，以支持急性 呼吸困难然而，不准确的AI工具也会使临床医生的表现恶化。因此，只要保持 临床医生的“在环”并不能保证对表现不佳的模型的支持。该提案寻求 实现有效的临床医生-AI协作，以提高急性呼吸困难的诊断准确性。我们建议： 1)评估计算策略，以提高用于支持临床医生的AI工具的鲁棒性， 急性呼吸困难的诊断，2）增强临床医生和AI工具之间协作的测试策略，3） 在临床环境中前瞻性评估急性呼吸困难AI工具，同时评估收集 临床医生反馈，以实现持续的模型改进。我们的多学科团队由专家组成， 临床医学、计算机视觉、机器学习和人机交互处于有利地位， 应对这些重大挑战。成功完成本提案将产生一个强有力的、可推广的 急性呼吸困难AI工具，用于增强医生对急性呼吸困难的诊断评估。更广泛的 该提案将产生可推广的知识，以支持更安全的开发和集成人工智能工具， 医疗保健设置。

PROJECT SUMMARY Acute dyspnea (shortness of breath) is one of the most common reasons for emergency department visits and hospitalizations each year. Heart failure, pneumonia, and chronic obstructive pulmonary disease are the most common etiologies, representing 2.5 million hospitalizations in the US in 2017. Determining the precise cause of acute dyspnea is critically important but challenging, as presenting symptoms, laboratory testing, and imaging results may be difficult to interpret, particularly in the elderly and patients with comorbid disease or severe illness. Diagnostic errors and inappropriate treatment may occur in up to 30% of patients, which is associated with worse patient outcomes. Artificial Intelligence (AI) tools have been proposed to augment providers in the diagnostic process and are well-positioned to support the diagnostic evaluation of acute dyspnea. However, inaccurate AI tools can also worsen clinician performance. Therefore, simply keeping clinicians “in-the-loop” is not a guaranteed back-stop against a poorly performing model. This proposal seeks to enable effective Clinician-AI collaborations to improve diagnostic accuracy in acute dyspnea. We propose to: 1) evaluate computational strategies to improve the robustness of an AI tool used to support clinicians in the diagnosis of acute dyspnea, 2) test strategies to enhance collaborations between clinicians and AI tools, 3) prospectively evaluate an acute dyspnea AI tool in a clinical environment while evaluating strategies to collect clinician feedback to enable ongoing model improvement. Our multidisciplinary team consisting of experts in clinical medicine, computer vision, machine learning, and human-computer interaction are well positioned to tackle these important challenges. Successful completion of this proposal will result in a robust, generalizable acute dyspnea AI tool to augment physicians in the diagnostic evaluation of acute dyspnea. More broadly, the proposal will lead to generalizable knowledge to support safer development and integration of AI tools across healthcare settings.