Video Analysis of Neurosurgical Technical Performance and Adverse Events

项目总结 拟议的研究和职业发展计划旨在为应聘者提供知识， 成为一名独立的神经外科医生兼科学家所需的经验和资源，其研究 通过机器的设计和实施减少神经外科手术后的中风和神经性残疾 为外科医生提供反馈以改进外科手术的学习(ML)和计算机视觉(CV)系统 性能。经过正式培训后，执业神经外科医生收到的反馈很少，而是通过 在手术过程中积累了经验。然而，大多数都没有积累足够的案例数量来实现最佳 每一次手术的结果。最终，16000名患者会受到可预防的神经外科错误的伤害 每年，导致中风和神经系统残疾的比例高达70%，死亡的比例高达16% 病人。不幸的是，由于缺乏数据集，对手术期间有害不良事件的研究受到阻碍 包含导致不良事件或结果的外科手术的。这位候选人提议克服 这一限制是通过使用高级CV和ML方法来分析以前未研究过的多模式数据集实现的 结合脑下垂体手术录像和临床资料。美国每年要做10,000次脑下垂体手术。 美国，可以记录下来进行分析，其步骤、错误和不良事件最近在 一份国际共识声明。具体地说，候选人将测试核心假设，即 可视化外科医生技能因素与可视化特征的交互作用，包括患者解剖和疾病 病理学，产生可识别的特定步骤的手术错误，导致术后中风、神经学 残疾，以及其他不良事件。具体目标：1)使用CV识别特定步骤的错误(由工具定义 用法、步进和阶段特征)之前的不良事件；2)训练ML模型以预测 来自外科医生技能和外科领域当前视觉特征的先前指标的即将到来的不良事件。 识别高危神经外科手术，预测外科医生即将发生的不良事件的方法 运动，并回顾强调与不良事件相关的关键时间点和视觉特征 是合理设计和实施干预措施以减少中风、神经残疾和 其他不良事件。候选人的杰出表现将有助于这项工作的可行性和成功 指导团队，包括一名具有指导基于简历的手术性能的经验的外科医生兼科学家 使用医学图像和临床数据的ML程序视觉数据和专家的评估，生物医学 深度学习在复杂预测模型和多机构垂体外科研究中的应用。在……里面 在奖项的最后一年，候选人将申请R01奖项，以期实现概括性 预测模型(在AIM 2中开发)使用来自几个手术程序的更大数据集和 开发从手术视频中实时收集数据并对其采取行动的方法。

PROJECT SUMMARY The proposed research and career development plan aim to provide the candidate with the knowledge, experience, and resources necessary to become an independent neurosurgeon-scientist whose research reduces stroke and neurologic disability after neurosurgery through the design and implementation of machine learning (ML) and computer vision (CV) systems that provide surgeons with feedback to improve surgical performance. After formal training, practicing neurosurgeons receive little feedback and instead learn by experience accrued during procedures. However, most do not accrue sufficient case volume to achieve optimal outcomes in every procedure. Ultimately, >16,000 patients are harmed by preventable neurosurgical errors each year, resulting in stroke and neurologic disability in up to 70% and death in up to 16% of affected patients. Unfortunately, the study of harmful adverse events during surgery is obstructed by a lack of datasets containing surgical actions leading up to adverse events or outcomes. The candidate proposes to overcome this limitation by using advanced CV and ML methods to analyze a previously unstudied, multimodal dataset combining pituitary surgical video and clinical data. Pituitary surgery is performed >10,000 times annually in the U.S., can be recorded for analysis, and its steps, errors, and adverse events were recently standardized in an international consensus statement. Specifically, the candidate will test the central hypothesis that the interaction of visible surgeon skill factors with visualized features, including patient anatomy and disease pathology, produces identifiable step-specific surgical errors that result in postoperative stroke, neurologic disability, and other adverse events. Specific Aims: 1) Use CV to identify step-specific errors (defined by tool usage, step progression, and phase characteristics) preceding adverse events; 2) Train ML models to predict upcoming adverse events from prior metrics of surgeon skill and current visual features of the surgical field. Methods to identify high-risk neurosurgical actions, predict upcoming adverse events from a surgeon’s movements, and retrospectively highlight critical timepoints and visual features associated with adverse events are necessary to rationally design and implement interventions to reduce stroke, neurologic disability, and other adverse events. The feasibility and success of this work will be facilitated by the candidate’s outstanding mentoring team, including a surgeon-scientist with experience conducting CV-based surgical performance assessments from procedural visual data and experts in ML using medical image and clinical data, biomedical applications of deep learning in complex prediction models, and multi-institutional pituitary surgical research. In the final year of the award, the candidate will apply for an R01 award to prospectively implement generalizable predictive models (developed in Aim 2) using a larger dataset of videos from several surgical procedures and to develop methods to collect and act upon data from operative video in real-time.