CRII:SCH: Interactive Explainable Deep Survival Analysis

每年，美国在医疗保健方面的支出几乎占国内生产总值（GDP）的20%，随着老龄化人口进入医疗保险，这一增长将继续得到推动。虽然费用巨大，但许多患者未能得到及时有效的药物治疗。准确的诊断对临床决策至关重要。然而，“预防胜于治疗”，因为预防和早期干预将防止老年人遭受更多的疾病和/或更广泛的治疗。而且，当观察到大量患者处于严重损害其健康的进展性疾病的晚期时，建立预测模型为时已晚。同时，为了供医疗保健提供者使用，预测模型需要具有可解释性和可信性。此外，人类利益相关者（例如，开发人员、领域专家和/或最终用户）之间的有效交互以及清晰的模型解释不仅可以提高模型性能，还可以增强人类的信任。拟议的研究项目旨在开发算法和方法，支持为预防和早期干预实施可信和高效的数据驱动决策。本项目提出的主要方法是交互式可解释的深度生存分析。生存分析的目的是预测发生感兴趣事件的时间，这在模拟疾病进展、确定预后因素、评估健康风险的医疗保健中非常有益。本项目将构建健康老龄化和精准医疗领域的深度生存分析模型，以支持临床决策，特别是在大量患者尚未遭受疾病侵袭的疾病早期。深度生存分析是一种“黑箱”模型，利益相关者无法知道模型是如何运作的，以及它是如何做出决策的，因此限制了它在实践中的使用。本项目将开发利用领域知识编码和专家反馈实现深度生存分析模型透明度和可信度的方法，以获得更好的预测性能。更具体地说，该项目将提出一个时间相关的反事实梯度积分，以解释是什么使模型输出与每个时间间隔的反事实生存状态区分开来。本项目还将在深度生存分析模型的训练过程中加入特征归因先验，以提高解释的一致性以及模型的性能和可信度。受human-in-the-loop的启发，本项目将进一步研究有效的方案，以数学方式表达医生的定性反馈，并通过强大的感知用户界面将其交互地纳入模型的学习过程中，从而有效地编码来自医生的各种类型的反馈。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Annually, the United States spends almost 20% of gross domestic product (GDP) in healthcare with growth continued to be boosted by a greying population aging into Medicare. Although the cost is huge, numerous patients fail to get timely and effective medication cure. Accurate diagnosis is critical in clinical decision making. However, “prevention is better than cure” as prevention and early intervention will prevent the aging people from suffering more diseases and/or more extensive treatments. Also, it is too late to build the prediction model when a lot of patients have been observed in the late stage of a progressive disease, which severely damages their health. Meanwhile, in order to be usable by healthcare providers, the prediction model needs to be interpretable and trustable. Also, efficient interaction between human stakeholders (e.g., developers, domain experts and/or end-users) and clear model interpretation not only improve the model performance but also enhance human trust. The proposed research project aims at developing algorithms and methods that support implementation of trustworthy and time-efficient data-driven decision making for prevention and early intervention.The main approach proposed in this project is interactive explainable deep survival analysis. Survival analysis aims at predicting the time to event of interest, which is extremely beneficial in healthcare for modeling disease progression, identifying prognostic factors, assessing risk of health. This project will build deep survival analysis models in healthy aging and precision medicine to support clinical decision making, especially in the early stage of a progressive disease before a lot of patients have been suffered from that disease. Deep survival analysis is a kind of “black box” model that stakeholders cannot tell how the model operates and how it comes to its decisions and hence limits its usage in practice. This project will develop methods to achieve both transparency and trustworthiness in deep survival analysis models with encoding of domain knowledge and expert feedback to achieve better prediction performance. More specifically, this project will propose a time-dependent counterfactual gradient integration to interpret what makes the model output differentiate from the counterfactual survival status at each time interval. This project will also incorporate feature attribution priors into the training process of deep survival analysis model to improve consistency of the explanation as well as the performance and trustworthiness of the model. Inspired by human-in-the-loop, this project will further investigate efficient schemes to mathematically formulate physicians' qualitative feedback, and interactively incorporate them in the learning process of the model with powerful perceptual user interface to efficiently encode diverse types of feedback from physicians.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.