Improved Glaucoma Monitoring Using Artificial-Intelligence Enabled Dashboard

检测青光眼引起的功能和结构损失对于做出治疗决定至关重要， 保护视力和维持生活质量。然而，大多数青光眼评估方法 通过视野（VF）或光学相干断层扫描（OCT）测量具有几个局限性， 对它们的临床应用提出了严峻的挑战。 从VF或OCT数据序列中识别脑卒中引起的变化具有挑战性， 处于疾病的早期阶段，具有微妙的结构和功能体征，或者如果患者 在疾病的后期阶段，具有显著的VF变异性和OCT地板效应。的主要限制 目前的青光眼监测技术是它们产生青光眼是否 恶化与否而当前的高吞吐量数据（例如，OCT）比二元结果具有更多的信息。 这些方法的另一个主要缺点是它们依赖于传统的发展模式 检测，如线性回归。然而，昏迷进展的速率可能是非线性的和快速的， 特别是在疾病的后期阶段。另一个限制是，采用特别规则来界定 青光眼进展，而需要客观标准来定义进展的阈值。最后，一个主要 这些方法中的大多数的不足之处在于它们缺乏先进的可视化和解释。 我们建议通过开发支持人工智能（AI）的可视化工具来解决这些限制， 有效监测青光眼患者的功能和结构损失。这种方法提供 监测1）整体视觉功能和结构恶化的定性和定量方法，2） 半野的损失，和3）先进的2-D可视化工具上的功能和结构损失的局部模式。到 为了实现这些目标，我们组建了一个跨学科专家团队，可以访问大型临床 注释的青光眼数据。 该提案的中心假设是，先进的可解释机器学习应用于一个完整的 所有测试位置的VF分布（例如，24 - 2系统中的54个）和OCT衍生的视网膜神经测量 纤维层（RNFL）（例如，768 A扫描周围的视盘和7个全球部门区域）可以客观和 自动学习和量化最重要的特征，从而产生更具体和敏感的方法， 监测青光眼恶化比目前主观指定或诊断确定的方法。 我们还假设机器学习可以提供具有几层青光眼的可解释模型 这些知识可能为当前青光眼评估测试提供有希望的补充。 我们提出的研究可能会提供实质性的改善预后和青光眼的管理， 有效利用分析和可视化来改善青光眼管理， 治疗方案。

Detecting functional and structural loss due to glaucoma is critical to making treatment decisions with the goal of preserving vision and maintaining quality of life. However, most of the approaches for glaucoma assessment through visual fields (VFs) or optical coherence tomography (OCT) measurements have several limitations that poses critical challenge to their clinical utility. Identifying glaucoma-induced changes from a sequence of VF or OCT data is challenging either if the patients is in the early stages of the disease with subtle manifested structural and functional signs or if the patients are in the later stages of the disease with significant VF variability and OCT flooring effect. A major limitation of the current glaucoma monitoring techniques is that they generate a binary outcome of whether the glaucoma is worsening or not while current high-throughput data (e.g., OCT) has more information than a binary outcome. Another major drawback of some of these approaches is that they rely on traditional paradigms for progression detection such as linear regression. However, rates of glaucomatous progression may be non-linear and rapid, particularly during the later stages of the disease. Another limitation is that ad-hoc rules are adopted to define glaucoma progression while objective criteria are required to define thresholds for progression. Finally, a major deficiency of most of these methods is that they lack advanced visualization and interpretation. We propose to address these limitations by developing artificial intelligence (AI)-enabled visualization tools for effectively monitoring the functional and structural loss in patients with glaucoma. This approach provides qualitative and quantitative means to monitor 1) global visual functional and structural worsening, 2) extent of loss in hemifields, and 3) local patterns of functional and structural loss on advanced 2-D visualization tools. To achieve these objectives, we have assembled a team of interdisciplinary experts with access to large clinically annotated glaucoma data. The central hypothesis of this proposal is that advanced interpretable machine learning applied to a complete profile of VFs in all test locations (e.g., 54 in 24-2 system) and OCT-derived measurements of retinal nerve fiber layer (RNFL) (e.g., 768 A-scans around the optic disc and 7 global sectoral regions) can objectively and automatically learn and quantify the most important features, yielding a more specific and sensitive means for monitoring of glaucoma worsening than current subjectively-specified or statistically-identified approaches. We also hypothesize that machine learning can provide interpretable models with several layers of glaucoma knowledge that may provide a promising complement to current glaucoma assessment tests. Our proposed studies may offer substantial improvements in prognosis and management of glaucoma through effective use of analysis and visualization to improve glaucoma management and making more informed treatment options.