Human-centered design of clinical AI to support the diagnosis of pediatric suprasellar tumors

项目摘要/摘要 子专科的临床决策，如儿科神经肿瘤学，正在变得越来越数据化- 干劲十足且复杂。人工智能(AI)是一个强大的工具，可以帮助提取这些不断扩大的数据集 在正确的时间向临床医生提供正确的信息。人工智能在临床应用中收效甚微 到目前为止，新的方法(如OpenAI的Dall-E或GPT3模型)现在进入了公众的视线，清楚地展示了 总的来说，这项技术的力量。将这项技术有效地转化为临床环境需要 全面了解具体的临床环境，如人员/角色、使用的数据/技术、 临床目标和工作流程。以人为中心的设计(HCD)是一个强调需求的解决方案框架 执行特定任务的人员，并且非常适合于促进临床人工智能的设计。然而，HCD 在这个领域执行是具有挑战性的，因为组建一个专家团队既困难又昂贵 横跨临床医学、人工智能、可视化和社会科学。学术界和工业界都有 建立了多学科的HCD/AI团队，但正在寻找解决方案来填补跨学科的领导角色 在这些队伍中。 我之前发表了一个深度学习模型，用于将儿童鞍上肿瘤从 术前核磁共振检查。我的模型在相同的数据集(86%)上的表现与人类专家一样好，这也是 与之前关于人类专家在这项任务上表现的研究是一致的。此外，儿科鞍上区 肿瘤几乎总是通过手术病理诊断，大约8%的患者是通过放射学检查确诊的。 诊断出来了。初步数据(目标2)表明，我们可以将深度学习模型的性能提高到 95%，通过结合贝叶斯方法来估计模型的不确定性。其他初步数据(AIM 3)表明将我的模型嵌入到谷歌的假设分析工具(WIT)中可以帮助临床医生进行放射学检查 诊断这些肿瘤的难度更小，信心更强。因此，这项提案 中心假设是可解释的人工智能解决方案可以改善人类专家的儿童鞍上肿瘤 放射诊断超出了当前的性能水平。此外，使用HCD方法，谷歌的 假设工具(WIT)可以调整到临床医生的工作流程中，从而导致采用该工具 辅助技术。我将调查我的目标，专门为提供这些方面的实用知识而设计的 使我能够有效地领导由主题专家组成的多学科团队来开发可靠的 临床人工智能工具。我由一个专家导师团队指导，他们代表儿科神经肿瘤学，神经外科， AI、可视化和HCD。完成这项提议将对我的职业目标做出重大贡献：成为一名 应用HCD开发支持儿童神经肿瘤患者的临床人工智能技术的领先者 关心。

PROJECT SUMMARY/ABSTRACT Clinical decision-making in subspecialties, like pediatric neuro-oncology, is becoming increasingly data- driven and complex. Artificial intelligence (AI) is a powerful tool that can help distill this expanding dataset to present the clinician with the right information at the right time. AI has had little success in clinical applications so far, but new methods (like OpenAI's DALL-E or GPT3 models) are now in the public eye, clearly demonstrating the power of the technology generally. Effective translation of that technology into the clinical setting requires a comprehensive understanding of the specific clinical setting, such as personnel/roles, data/technology used, clinical goals and workflows. Human-Centered Design (HCD) is a solution framework that emphasizes the needs of the people who perform a specific task and is well suited to facilitate the design of clinical AI. However, HCD is challenging to execute in this space because it is difficult and expensive to assemble a team of experts that spans clinical medicine, artificial intelligence, visualization, and social sciences. Academia and industry have established multidisciplinary HCD/AI teams but are seeking solutions for filling interdisciplinary leadership roles in these teams. I previously published a deep learning model for classifying pediatric suprasellar tumors from preoperative MRI. My model performed as well as human experts on the same dataset (86%), which was also congruent with previous studies on human expert performance on the task. In addition, pediatric suprasellar tumors are almost always diagnosed via surgical pathology, with roughly 8% of patients being radiographically diagnosed. Preliminary data (Aim 2) suggests that we can improve the deep learning model performance up to 95% by incorporating a Bayesian methodology to estimate model uncertainty. Additional preliminary data (Aim 3) indicates that embedding my model into Google's What-If Tool (WIT) can help clinicians radiographically diagnose these tumors with less perceived difficulty and greater perceived confidence. Therefore, this proposal's central hypothesis is that explainable AI solutions can improve human experts' pediatric suprasellar tumor radiographic diagnosis beyond the current performance levels. Moreover, using an HCD approach, Google's What-If Tool (WIT) can be adapted into the clinician's workflow in a manner that will result in adoption of this assistive technology. I will investigate my aims, specifically designed to provide functional knowledge in these topics to enable me to effectively lead a multidisciplinary team of subject-matter experts in developing robust clinical AI tools. I am guided by an expert team of mentors representing pediatric neuro-oncology, neurosurgery, AI, visualization, and HCD. Completion of this proposal will significantly contribute to my career goal: to be a leader in the application of HCD to develop clinical AI technology that supports pediatric neuro-oncology patient care.