CRII: SCH: Domain-guided Machine Learning for Clinical Decision Support in Epilepsy

尽管全国范围内神经科医生短缺，但当今的神经科护理在很大程度上依赖于训练有素的工作人员对患者数据进行耗时的目视检查。这在癫痫学领域尤其重要，癫痫学家花费大量时间来目视检查和解释脑电活动的漫长多通道时间序列，称为脑电图 (EEG)。这种负担不仅导致癫痫专家的倦怠升级，而且还引入了审稿人的偏见和临床决策中的潜在错误。该提案的目标是开发一个基于机器学习 (ML) 的决策支持框架，与癫痫学家合作，并将他们的注意力集中在可操作的信息上。我们将利用伊利诺伊州的计算专业知识和梅奥诊所合作者的临床领域专业知识，并在整个数据科学生命周期中展示重大创新，以实现上述目标。本研究中使用的数据和方法将作为高级跨学科课程和培训医疗保健专业人员的范例。我们还相信，医疗保健应用程序的天然吸引力将激发本科生和代表性不足的少数群体的兴趣。这项研究将开发一套新颖的领域引导分析方法来处理时间序列脑电图数据，提取可操作的信息并为诊断癫痫提供临床决策支持。拟议研究的智力价值在于通过开发由临床领域专业知识指导的新型可解释机器学习架构来解决癫痫学领域未满足的需求。我们提出的工作包括：a）利用基于深度学习的方法的廉价推理能力，开发全自动且高效的脑电图预处理流程； b) 在领域专业知识的指导下设计新颖的机器学习模型，捕获脑电图数据的时空动态； c) 模型预测的解释和预测不确定性的量化以支持临床决策； d) 通过开发强大的分析工具来增强脑电图专家审查并提高癫痫诊断的敏感性，在现实世界中展示该框架。该奖项反映了 NSF 的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

Despite the nationwide shortage of neurologists, present-day neurological care relies heavily on time-consuming visual review of patient data by trained staff. This is particularly emphasized in the field of epileptology where epileptologists spend a substantial amount of their time on visually reviewing and interpreting lengthy multi-channel time series of brain electrical activity, called electroencephalography (EEG). This burden not only contributes to the escalation of epileptologist burnout, but also introduces reviewer bias and potential errors in clinical decisions. The goal of this proposal is to develop a machine-learning (ML)-based decision support framework that works together with epileptologists and focuses their attention to actionable information. We will leverage the computing expertise of Illinois and the clinical domain expertise of our collaborators at the Mayo Clinic and demonstrate significant innovations across the data-science lifecycle to achieve the aforementioned goal. The data and the methods utilized in this research will serve as examples in advanced interdisciplinary classes and training healthcare professionals. We also believe that the natural appeal of healthcare applications will stimulate the interest of undergraduates and underrepresented minorities.This research will develop a set of novel domain-guided analytical methods to process time-series EEG data, extract actionable information and provide clinical decision support for diagnosing epilepsy. The intellectual merit of the proposed research is in addressing an unmet need in the field of epileptology through the development of novel explainable machine learning architectures guided by clinical domain expertise. Our proposed work includes, a) development of a fully automated and efficient EEG preprocessing pipeline by leveraging the cheap inference capability of deep learning-based approaches; b) designing novel ML models, guided by domain expertise, that capture the spatio-temporal dynamics of EEG data; c) interpretation of model predictions and quantification of prediction uncertainty for clinical decision support; and d) demonstration of the framework in the real world by developing a robust analytical tool to augment expert review of EEGs and improve the sensitivity of epilepsy diagnosis.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.