Machine Learning to Determine Dynamically Evolving New-Onset Venous Thromboembolic (VTE) Event Risk in Hospitalized Patients

点击翻译按钮获取中文摘要

Failure to rescue (FTR), a nurse-sensitive national metric of health care quality, refers to death of a hospitalized patient from a treatable complication, and is potentiated by failure to recognize and appropriately respond to early signs of complications. There is a paucity of research examining patient features predictive of FTR complications. Such information could shift the current paradigm of nursing surveillance to earlier recognition, prevention and treatment of FTR complications, thereby saving lives. New-onset venous thromboembolism (VTE), an FTR complication occurring as either a deep vein thrombosis (DVT) or a pulmonary embolism (PE), is the leading cause of preventable hospital death, carrying a high risk of mortality and a national cost burden of $7 billion annually. VTE is a complex disease process involving interactions between clinical risk factors and acquired and/or inherited susceptibilities to thrombosis. Although biomarkers and clinical factors associated with VTE have been identified, clinical manifestations are subtle, presenting gradually over hours to days. Current VTE risk assessment models (RAM), the cornerstone of prevention, have limited utility due to their complexity and lack of reliability, generalizability and external validation. A critical gap in VTE risk modeling research is that while new-onset VTE pathology evolves over the course of hospitalization, no current models incorporate the progressive accrual of dynamic patient data and pattern evolution over time in their modeling approaches. The totality of routinely collected electronic health record (EHR) data is massive in terms of volume, variety, and production at a rapid velocity in real-time. Such big data could be used in machine learning (ML) analytic approaches to process time series clinical data to identify subtle, evolving feature patterns predictive of new-onset VTE and address this gap. This study proposes to assemble a large scale, multi-source, multi-dimensional VTE study dataset, and in tandem, systematically define the EHR data elements associated with a new-onset VTE diagnosis for computable phenotype algorithm development. We will then apply machine learning analytic approaches to baseline and accruing intensive time series clinical data in the curated dataset to develop models identifying data patterns and features predictive of dynamically evolving new-onset VTE in adult hospitalized patients. This proposal aligns with NINR’s strategic vision for nurse scientists to employ new strategies for collecting and analyzing complex big data sets to permit better understanding of the biological underpinnings of health, and improve ways nurses prevent and manage illness. This innovative study and individualized training plan under a strong and well- established team, represents initial steps in the applicant’s research trajectory focused on data science approaches to predict FTR complication risk, and develop, implement and test dynamic RAMs to inform targeted prevention and treatment decisions. Discovering new knowledge informing real-time decision making, nursing surveillance practices and care delivery systems can improve nurse sensitive patient outcomes.