In the United States, intensive care units (ICUs) costs exceed 4% of national health costs, and ICU mortality rates can be as high as 29%. Precise assessment and prediction of patient status in the ICU can enable early interventions, and can result in improved patient outcomes. However, today's ICUs still face many barriers for assessing and predicting patient status. First, essential information such as pain and functional status are not captured automatically, but rather are repetitively measured by overburdened ICU nurses, with new assessments added each year. Second, existing methods for predicting patient status have limited accuracy and are used infrequently (e.g. every 24 hours as opposed to in real-time). This leads to missing opportunities for early interventions. Finally, existing models cannot automatically incorporate family caregiver feedback for improved patient status prediction. Together, these challenges point to the critical need for developing several fundamental intelligent building blocks of future ICUs. These building blocks should address: (a) how to learn new patient status prediction models without compromising performance on previous prediction models, (b) how to handle the complex clinical data for precise prediction of patient outcome, and (c) how to incorporate family caregiver feedback into the prediction models.
This project will pursue three specific research objectives that will address these issues: (1) Lifelong Multi-Task Learning: Novel multi-task deep learning models will be developed for recognizing clinical expressions and functional activities related to pain and functional status assessments. These models will be customized in an innovative manner to maximize information sharing among related tasks. (2) Multi-Scale and Dynamic Learning: Novel multi-scale recurrent neural networks will be developed to predict precise patient outcomes while handling multiple temporal scales and implicit input changes over time. (3) Continual Opportunistic Learning: Novel active deep learning models will be developed to query the labels of the most informative data points for improving the models over time, with minimum burden on users. The proposed project will bring together novel elements of machine learning algorithms and critical care medicine. This will represent the first attempt to autonomously assess pain and functional status in the ICU, to predict precise patient trajectory from high-resolution data, and to improve predictive clinical models through user feedback. In addition, the research will be impactful because what is learned here, will contribute to a broader understanding of future design considerations for the next generation of lifelong learning systems and intelligent hospitals. The PI will also provide a highly-integrated research and educational program for Florida high school teachers and students, and University of Florida (UF) undergraduate students in the context of the intelligent ICU. The PI proposes to: (1) sponsor summer internships for math teachers, (2) organize an Intelligent Machines workshop on coding and machine intelligence for the high school students, and (3) develop focused research and training activities for undergraduate students. These outreach and training programs will be used to promote interest in science, technology, engineering, and mathematics (STEM) fields among Florida high school students and UF undergraduate students.
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.
