The ability to predict the clinical course of intensive care unit (ICU) patients in a timely manner can guide decision making and accelerate research into therapeutic efforts and the economics of care. To achieve this, a variety of ICU-based risk scoring systems have been proposed to risk stratify patients using "snapshots" of variables at specific points in time. These systems are not designed to provide a continuous assessment of patient status in real-time. In recent years, there has been a growing interest in intelligent patient monitoring (IPM) systems that can use information in continuously recorded ICU signals to recognize changes suggestive of dangerous pathophysiologieis.
The aim of our research is to augment the knowledge base for these IPM systems by exploring the feasibility of pattern discovery methods to identify novel markers of future risk from large volumes of continuous ICU data. We focus, in particular, on sophisticated methods from data mining and machine learning that are computationally efficient, robust, and able to identify multi-parameter time-series trends useful for risk stratification. To facilitate these efforts, we will utilize the high volume acuity environment of the Surgical ICU (SICU) at the Henry Ford Hospital in Detroit. We will study the use of pattern discovery methods to identify high and low risk patterns in historical data from over 5,000 patients admitted to the SICU, and then prospectively evaluate our findings on real-time SICU data at the Henry Ford from over 3,700 patients.
The specific aims of this proposal are: (1) to develop novel decision support tools based on high and low risk patterns discovered from large historical time-series ICU datasets. Using fully-automated and computationally efficient algorithms, we will first identify characteristic units of temporal activity that explain the multi-parameter physiological time-series for ICU patients, and then discover approximate sequences of these characteristic units associated with adverse events in patients with known outcomes. The patterns discovered using our approach will be integrated into real-time decision support tools to signal alerts when high risk patterns are observed;and (2) to prospectively validate decision support tools based on high and low risk patterns in ICU time-series on real-time ICU data. We will conduct a pilot study validating these tools by studying the association between the predictions for these tools and actual adverse outcomes observed in the SICU. Each tool will be evaluated based on the following metrics: sensitivity, specificity, positive predictive value, negative predictive value, area under the receiver operating characteristic curve, and clinical time-frame of prediction. We will further compare the improvement provided by these tools to existing risk scoring systems.
Successful management of critical care patients depends on the ability of clinicians to estimate the risk of future adverse events, and ideally, to anticipate and prevent those events. We propose to explore low-cost, and widely applicable methods that would improve the ability of physicians to identify patients at high risk for adverse events in real time in ICU settings. The key concept for this proposal is that time-series data collected in the ICU contain valuable prognostic information that can be extracted using novel, sophisticated techniques drawn from the domains of machine learning and data mining. This information can be used to develop real- time clinical tools to improve management of critically ill patients.