Deep Learning and Streaming Analytics for Prediction of Adverse Events in the ICU

点击翻译按钮获取中文摘要

 DESCRIPTION (provided by applicant) Critical care medicine in the United States costs over 80 billion dollars annually. Over the past decade the rate of intensive care unit (ICU) use has been increasing, with a recent study reporting almost one in three Medicare beneficiaries experiencing an ICU visit during the last month of their lives. Every year, sepsis, a medical condition characterized by whole-body inflammation, strikes between 800,000 and 3.1 million Americans, killing approximately one in four patients affected. There is currently no definite treatment for sepsis in spite of many clinicl trials. However, early detection of sepsis and timely initiation of interventions are widely considered as important determinants of patient survival. However, basic care tasks (such as microbiological sampling and antibiotic delivery within 1 h, fluid resuscitation, and risk stratification using serum lactate or alternative), which are known to benefit most patients, are not performed in a timely manner. Previous literature suggests that high-resolution vital signs (such as heart rate, blood pressure, respiratory rate, etc.), and other sequential measurements within the electronic medical records (EMRs), can be dynamically integrated using Machine Learning techniques to help with early detection of sepsis. With the ubiquity of inexpensive high- capacity storage and high-bandwidth streaming technology it is now possible to monitor patients' vital signs continuously (for instance, the research application developed by the Emory hospital ICU uses IBM's streaming analytics platform to transmit over 100,000 real-time data points per 100 beds, per second). Despite this continuous feed of data, commonly used acuity scores, such as APACHE and SAPS, are based on snapshot values of these vital signs (typically the worst values during a 24 hours period). This limitation is partially due to unavailability of computationally efficient and robust algorithms capable of finding predictive features in multivariate, nonlinear and nonstationary sequential data, which may reveal inter- organ communication and disintegration of causal couplings with critical illnesses such as sepsis. We have recently developed a novel Machine Learning algorithm to discover automatically a collection of predictive multivariate dynamical patterns in a database of patient time-series, which can be used to classify patients or to monitor progression of disease in a given patient. The primary goal of this proposal is to apply our method to assess the predictive power of high- resolution multivariate time-series of vital signs and sequentially recorded EMR data in the ICU for early detection of sepsis and risk stratification of septic patient. To accomplish this, we aim to benchmark our technique on a large ICU cohort (the MIMIC II database with over 60,000 patients), as well as simulated data from a multiscale mathematical model of influence of inflammatory mediators on dynamical patterns of vital signs. Next, the technique will be externally validated on two separate ICU sepsis cohorts (the Emory Sepsis dataset and the Mayo Clinic Metric dataset). Finally, we will provide a real-time implementation of the proposed algorithm in an streaming environment (such as the IBM streaming analytics), in order to address the Big Data challenge of harnessing real-time, streaming sensor data from bedside monitors within the ICU, while enabling advanced pattern recognition and real-time forecasting. Ultimately we believe these methods can change the current standard of care through faster recognition and initiation of basic care, as well as guiding interventional strategis based on severity of illness and mechanisms underlying physiological deterioration.