可解释性新冠死亡风险和治愈时间预后分析模型研究

分析疫情期间积累的大量异构多模态临床数据，进而理解患者死亡风险及治愈时间的深层次关系，对新冠患者的治疗具有至关重要的意义。现有的临床数据存在结构差异大、患者结局分布不均以及时变非线性关系难以解释等问题，对建模分析带来巨大挑战。现有建模方法普遍依赖于线性时不变假设，难以适应上述复杂场景。本项目旨在提出一种新型可解释性深度学习预后建模方法，有效利用上述复杂数据构建模型，以实现提前预测患者死亡风险和治愈时间之目的。重点研究以下内容：1.建立异构多模态数据特征提取和表征模型，并通过引入可解释性机制，解决模型“黑匣子”问题；2.设计合理的多事件损失函数，利用不同结局患者数据训练模型，构建鲁棒的时变非线性关系；3.在此基础上，揭示与死亡风险和治愈时间强相关的临床因素。本项目研究在理论上有助于发展可解释性人工智能预后新技术，同时对指导临床医生制定合理的患者检查项目和个性化治疗方案具有重要的应用价值。

It is clinically significant to analyze the large number of accumulated heterogeneous multimodal clinical data of COVID-19 patients, and to explore the potential relationships between the data and patients’ mortality risk and recovery time. However, modeling analysis of these data is a challenging task due to 1) wide variation of data structures, 2) imbalance of data distribution of patients with various outcomes (i.e., death, recovery, and censored), and 3) complicated time-variant non-linear relationships between the data and the outcomes. Prior modeling methods generally rely on the strong assumption of linear and time-invariant relationship between the data and the outcomes, which is not robust enough to address the above-mentioned data issues. This project aims to propose a novel interpretable deep learning-based modeling method, which can effectively use the above-mentioned data to construct the prognostic model for early prediction of both mortality risk and recovery time of COVID-19 patients at admission. The study will focus on the following three main contents: 1) Developing suitable model architecture for feature extraction and characterization of the heterogeneous multimodal data, and solving the “black box” issue by introducing novel interpretability mechanism; 2) Designing reasonable multi-event loss functions to optimize the model, which can aid to establish a robust time-variant non-linear relationship between the data and the outcomes; 3) Revealing the significant clinical factors that are highly related to the mortality risk and recovery time of the COVID-19 patients. The successful implementation of the project will promote the development of new interpretable prognosis technology of artificial intelligence. Furthermore, this project has great value in clinical application such as guiding clinicians to arrange reasonable patient examinations and personalized treatment plans for the patients.