Predicting clinical outcomes for patients admitted to intensive care unit: developing and validating a Canadian data based ICU prognostic and planning system

当受伤或疾病严重损害重要身体功能时，在称为重症监护病房（ICU）的专科医院单位提供维持生命的干预措施。ICU的患者管理是复杂的，医生通过分析快速变化的患者数据来做出至关重要的决定。目前的ICU预后工具仅提供医院死亡率的预测，因此对预测患者在ICU住院期间的临床结果没有帮助。由于现有预后工具的限制，ICU临床决策导致患者护理不规范，ICU资源利用效率低下。该项目与(a)新斯科舍省卫生局、重症监护部、(b)新斯科舍省卫生和保健部以及(c)新斯科舍省医生合作，旨在根据加拿大ICU数据开发新的ICU预后模型，帮助ICU医生在临床重要时间点做出有效的护理决定，以解释患者在整个ICU住院期间病情的变化。该项目将研究深度学习（DL）方法来开发ICU预后预测模型，以预测5个临床重要时间点的临床结果。入院时、24小时、48小时和72小时后以及出院前24小时。我们建议解决两个关键挑战，即：(a)逐步更新预测模型，以确保预测与新的ICU数据相关；(b)患者在ICU住院期间不同时间点的临床结果的时间概率预测。我们将整合来自新斯科舍省卫生局中心区（NSHA-CZ）哈利法克斯icu的多个患者数据来源。ICU临床资料、病理资料及影像学资料。我们计划应用可解释的人工智能方法，在属性和概念层面上为预测结果提供临床有意义的解释。预测模型将基于NSHA-CZ的新ICU病例进行为期1年的前瞻性评估，以衡量模型预测的临床结果。该项目将提供一个原型ICU预测和规划系统（ICU- pps），该系统包含了为NSHA-CZ ICU开发的预测模型。

When injury or illness severely compromises vital bodily functions, life-supporting interventions are provided in specialized hospital units called Intensive Care Unit (ICU). Patient management in ICU is complex and physicians make life-critical decisions by analyzing rapidly changing patient data. Current ICU prognostic tools only provide a prediction of hospital mortality, hence are not useful to predict clinical outcomes during the patient's ICU stay. Due to the limitations of current prognostic tools, ICU clinical making leads to non-standardized patient care and inefficient ICU resource utilization. In partnership with (a) Nova Scotia Health Authority, Department of Critical Care, (b) Nova Scotia (NS) Department of Health and Wellness, and (c) Doctors Nova Scotia, this project aims to develop novel ICU prognostic models, based on Canadian ICU data, to help ICU physicians make effective care decisions at clinically important time points to account for changes in the patient's condition during the patient's entire ICU stay. The project will investigate Deep Learning (DL) methods to develop ICU outcome prediction models to predict clinical outcomes at 5 clinically important time points-i.e. at time of admission, after 24, 48 and 72 hours, and 24 hours before discharge. We propose to address two key challenges-i.e. (a) progressive update of the prediction models to ensure predictive relevance to new ICU data; and (b) temporal probabilistic prediction of clinical outcomes across different time points during the patient's ICU stay. We will integrate multiple patient data sources from Nova Scotia health Authority Central Zone (NSHA-CZ) ICUs in Halifax-i.e. ICU clinical data, pathology data and radiologic data. We plan to apply explainable AI methods to provide clinically meaningful explanations of the predicted outcome at the attribute and conceptual levels. The prediction models will be prospectively evaluated over a 1-year period based on new ICU cases at NSHA-CZ to measure the clinical consequences of the model's predictions. The project's will deliver a prototype ICU Prognostic and Planning System (ICU-PPS) incorporating the prediction models developed for NSHA-CZ ICUs.