UL1TR001427: Efficient Modeling of Individualized COVID-19 Mortality Risk

Jason Cory Brunson, PhD

Project Summary

Artificial intelligence methods using large healthcare data sets frequently involve trade-offs, most notedly between accuracy and interpretability. Recent work on individualized models, in the tradition of case-based reasoning, has achieved both, but at a cost to efficiency. The COVID-19 pandemic has highlighted the need for real-time modeling with EHR-derived data that satisfies all three criteria. This proposal leverages recent advances in scientific modeling software to more efficiently manage ensembles of individualized models. The primary application is to predict outcomes for COVID-19 patients in Florida, and to augment these predictions with individualized risk assessments. The team brings clinical, mathematical, and computer science expertise toward the open-source implementation of this new modeling approach.

Relevance Statement

There is a widely-recognized need for point-of-care decision support that is both personalized and evidence-based: specific to the present environment, the local population or providing institution, or to the individual patient, and conducted with the same rigor as clinical trials and retrospective studies. This need has become especially apparent during the COVID-19 pandemic, during which the dominant viral strain, evidence-based prevention and treatment guidelines, public health policy, and individual behavior have varied widely across time and space. While such resources will necessarily make use of EHR-derived data, EHR-based predictive models illustrate a common trade-off in artificial intelligence: a loss in interpretability in exchange for a gain in predictive accuracy. A recent approach, which we call individualized modeling, builds on the promise of case-based reasoning by fitting a classical (and interpretable) statistical model to a cohort of past cases retrieved for its similarity to a new case. However, while this approach achieves both accuracy and interpretability, it has to date been limited by its memory and runtime costs. Our goal with this project is to make progress toward what might be called the “triple-aim” of EHR-based decision support: models that achieve the accuracy of advanced machine learning methods, the interpretability of classical statistical models, and the efficiency of both.