Data Science for Improved Outcomes in Interventional Healthcare

我在外科数据科学方面的长期研究目标是开发用于术中手术数据定量分析的机器学习方法，以改善患者的预后。由于疾病和手术技术的多样性，机器学习方法必须包含领域知识。我的工作将集中在：(1)实时决策支持和(2)绩效评估和指导。在这个资助期内，我建议研究两个主要应用：即时超声和眼科手术。这些徒手干预有许多自由度，需要对环境进行精确的感知。本研究致力于机器学习在护理点超声中进行全面的绩效评估和指导。我们将探索基于高精度跟踪系统和低成本惯性测量单元的传感器数据的神经网络架构在多方面技能评估中的应用。我们还将建立干预工作流程的模型，并利用几个数据源（例如姿势跟踪器，超声视频，网络摄像头视频）来识别正在进行的步骤和使用状态转换建模的最佳反馈。此外，我们将研究在培训课程中不同的x-reality模式的附加价值。这项工作促进了世界范围内医学教育向以能力为基础的模式转变（在这种模式下，认证完全基于是否达到了能力）。这将减轻对专家时间的限制，并通过提供自动评估和反馈来提高客观性。本研究将探讨实际白内障手术的质量保证方法，目的是直接改善患者的预后。我们将研究结合卷积神经网络和递归神经网络的手术视频相位检测方法。随后，使用因果推理，我们可以确定将优化患者结果的行动方案：向实时决策支持迈出了一步。我们还将通过视频分类网络调查手术每个阶段的质量措施，以及在此评估中低成本光学工具跟踪器的附加价值。这项研究通过提供质量和决策支持的实时测量来改善患者的结果。这些信息可以集成到手术导航软件中，使用数据驱动的方法优化患者的结果。

The long-term goal of my research in surgical data science is to develop machine learning methods for quantitative analysis of intra-operative surgical data to improve patient outcomes. Due to the diversity in disease and surgical techniques, it is imperative machine learning methods incorporate domain knowledge. My work will focus on: (1) real-time decision support and (2) performance assessment and coaching. For this funding period, I propose research in two primary applications: point-of-care ultrasound and ophthalmological surgery. These freehand interventions have many degrees of freedom and require precise sensorization of the environment. Performance Assessment and Coaching in Point-of-care Ultrasound This research works towards machine learning for comprehensive performance assessment and coaching in point-of-care ultrasound. We will explore the use of neural network architectures for multi-faceted skills assessment based on sensor data from both high accuracy tracking systems as well as low-cost inertial measurement units. We will also build models of the intervention's workflow and exploit several data sources (e.g. pose trackers, ultrasound video, webcam video) to identify the ongoing step and optimal feedback using state transition modelling. Furthermore, we will examine the added value if different x-reality modalities within training curricula. This work facilitates the worldwide shift in medical education to a competency-based model (one where accreditation is based solely upon whether competencies are achieved). This will alleviate constraints on expert time and increase objectivity by offering automated assessment and feedback. Quality Assurance in Ophthalmological Surgery This work will investigate methods for quality assurance in real cataract surgeries with the goal of directly improving patient outcomes. We will investigate methods for phase detection from surgical videos by combining convolutional and recurrent neural networks. Subsequently, using causal inference, we can determine the course of action that will optimize patient outcomes: a step towards real-time decision support. We will also investigate measures of quality in each phase of surgery via video classification networks, and the added value of low-cost optical tool trackers in this assessment. This research works towards improved patient outcomes by providing real-time measures of quality and decision support. Such information can be integrated into surgical navigation software to optimize patient outcomes using a data-driven approach.