Aligning Patient Acuity with Resource Intensity after Major Surgery

点击翻译按钮获取中文摘要

Project Summary The broad, long-term objective of this application is to generate an efficient, effective decision-support system to augment postoperative triage, transfer, and discharge decisions that affect more than 15 million patients in the United States annually. Evidence from single-institution studies suggests that postoperative overtriage of low acuity patients to intensive care units (ICUs) is associated with low value of care (outcomes/costs) compared with general ward admission, and that undertriage of high acuity patients to general wards is associated with increased mortality. These associations require validation externally and prospectively. In addition, further investigation is needed to determine whether there are similar, identifiable misalignments between patient acuity and resource intensity occurring throughout postoperative hospital admission and at the time of hospital discharge. Our central hypothesis is that aligning automated, data-driven patient acuity assessments with postoperative resource intensity using explainable, fair, uncertainty-aware deep learning models will be associated with decreased mortality and increased value of care. We will test our central hypothesis by performing three sets of related but independent experiments. First, we will externally validate an interoperable version of our postoperative triage classification system, initially using retrospective data at 42 hospitals across four institutions, then performing similar analyses with retrospective data on a federated learning platform, and finally using prospective data from 15 hospitals at two institutions. Second, we will generate continuous postoperative patient acuity assessments with novel DL architectures using multicenter, multimodal (including clinical notes), retrospective EHR data at three hospitals within a single institution. Third, we will critically evaluate and optimize model certainty and fairness using retrospective data at 43 hospitals across four institutions, generate an EHR-embedded decision-support system, and perform prospective decision support usability testing and optimization at two institutions. The proposed research is intended to produce a validated, interoperable postoperative triage classification system, foundational evidence for generating continuous streams of postoperative transfer and discharge recommendations, a postoperative triage decision support system ready for clinical implementation, and open-source software for optimizing deep learning certainty and fairness. Achieving these outcomes would increase the probability of success for automated, real-time postoperative triage decision-support in subsequent clinical trials, and the ultimate goal of augmenting personalized, patient-centered decision making in surgery.