Abstract
The cardiac intensive care unit (ICU) is an analytic environment, one in which insight about a patient’s condition and evolving trajectory is gleaned from data and information from monitoring systems, electronic health records, and other data repositories. That data, both structured and unstructured, presents rapidly to clinicians and must be effectively managed in a timely manner to meet goals for patient care. With advances in computer and information technology, the evolution of the fields of data science and biomedical informatics, and examples of high-impact experiences with advanced data analytics approaches inside and outside of health care, the ability to effectively process and utilize clinical data, particularly high-frequency data, is on the rise. In the ICU, real-time data analytics, including streaming analytics that employ refined and validated predictive algorithms, have the potential to improve our accuracy in assessing a patient’s condition, to support better assessment of the patient’s course and response to treatment, and eventually to guide decisions and actions toward achieving the best possible outcomes. In this chapter an introduction to streaming analytics is provided, and experience with a recently developed predictive analytics software platform is discussed as an example of the current state of data analytics applications in the pediatric cardiac care environment.
Keywords
information technology, stream processing, streaming analytics, data analytics, predictive analytics, biomedical informatics, machine learning
Current-state computer and information technology allows extraction and aggregation of large volumes of diverse data types to support medical decision making in hospitalized patients. With abundant and readily available data there has been growing interest and accumulating experience with advanced data analytics techniques such that there now exists both the demand and an opportunity for intensivists, among other clinicians, to pursue innovative information technology applications to benefit our fragile patients. But will we soon reach the “peak of inflated expectations” followed by a rapid descent to the “trough of disillusionment” before realizing the promises of precision medicine, big data analytics, and machine learning applications as some have suggested? It is the author’s belief that leveraging information technology and blending expertise from multiple disciplines such as medicine, engineering, mathematics, computer science, and even behavioral science will drive greater medical intelligence and more accurate and efficient health care across the range of clinical and nonclinical scenarios. The formal establishment of the field of biomedical informatics and growing experience with data science applications to support medical decision making and patient care delivery processes support this perspective.
The purpose of this chapter is to describe basic principles of high-frequency data analytics and, specifically, how the streaming analytics approach may be applied in the pediatric cardiac intensive care unit (CICU) to enhance our understanding and management of the critically ill patient with cardiac disease.
Overview of Real-Time Data Analytics
Real-time data analytics involves the capture, processing, and analysis of data as data is entering the system. In the data-rich intensive care unit (ICU) environment, information and networking technology, the ready access to clinical information, and the application of mathematical modeling methods are beginning to support our ability to better manage multiparameter data streams in our assessment and treatment of patients. Depending on the analytics platform and database structure, stored genetic and other patient-specific, structured, and unstructured historical data may also be included, thereby adding context for the population or circumstance of interest. At the same time, this would allow for growth of captured volumes of information for later use in research, quality, and outcomes analysis. Currently there is enthusiasm that robust data management capabilities will foster the development of predictive algorithms that may eventually support enhanced real-time decision making. For example, in patients with borderline left heart structures being considered for or undergoing biventricular repair, preoperative and postoperative hemodynamic, imaging, laboratory, and historical data may be incorporated in a real-time analytics approach to support such decisions as (1) determining whether or not biventricular repair is feasible, (2) which targeted interventions may improve likelihood of success in attaining biventricular physiology, and (3) whether the intraoperative and postoperative data indicate a successful physiology or predict successful short- and long-term outcomes in this patient population. Population-based and patient-specific historical and real-time data are available in such scenarios and, ideally, would inform the decision-making approach during each management step over time. As one can see, such a scenario exemplifies a range of data analytics approaches that may be described to include the following major categories: descriptive analytics, predictive analytics, and prescriptive analytics. These categories may be viewed along a continuum that considers what has happened, what could happen, and what should be done, respectively. Descriptive analytics are familiar to all of us and involve the use of summary statistics that give insight into the past and are common in health care and non–health care industries. Predictive analytics employ the use of mathematical models to forecast the future or estimate data that are not yet in existence. Accurate prediction requires that certain conditions remain constant over time and that the correct or relevant data elements are included in the models. There are a growing number of health care applications and numerous non–health care applications that are currently in existence as examples of predictive methods. The extent to which any of these applications is successful is highly dependent on its level of validation during algorithm development and testing, as well as a robust system of application evaluation and reengineering as needed. Prescriptive analytics incorporate the features of predictive modeling derived from historical and real-time data to achieve accurate near- or real-time prediction for the expressed purpose of driving actions. At this time, practical applications using this type of analytics in health care are relatively rare, at least in the live clinical setting because of limitations in extracting real-time data and a paucity of successfully validated predictive models. On the other hand, there are numerous non–health care applications that are successfully driven by prescriptive analytics methodology. Some examples of these include credit card fraud protection, cyber security monitoring, GPS location and guidance, and market data management to name a few. Finally, from the capabilities afforded by predictive and prescriptive methods comes streaming analytics, or event stream processing, the basic foundation of which will be reviewed in the next section.
Streaming Analytics
Optimal care of critically ill cardiac patients requires thoughtful and accurate assessment of their physiologic state using a variety of data elements not just at a single point in time, but repetitively and in response to expected and unexpected events. Hence any care strategy must seek to achieve accurate diagnoses and to discern the patient’s overall trajectory pattern with the paradigm that deviations from expected patterns or valid early warnings will be heeded and appropriate adjustments in care will be made.
Digital data stream analysis is not new, and, historically, high-frequency, high-volume data have been processed in batches and at periodic intervals (daily, weekly, or at longer intervals). Much of the impact of this work has been realized through automation of clerical tasks and standardization of routine decisions and actions. In commercial industries, stream analytics has resulted in improved business efficiency and customer satisfaction owing to data access, the Internet, and ubiquity of personal computers and mobile devices. Unfortunately, many data management systems remain fragmented and struggle with the management of large volumes of high-variety data that may present rapidly, a major consequence of which is long latency periods for access to the right data to make timely and correct decisions. In contrast to batch processing, event stream processing involves the continuous analysis of flowing time series data with the primary goals of providing real-time insight about a situation and to allow early detection of anomalies. Importantly, there is an inverse relationship between response time latency (i.e., the time from event or data element capture to delivery of action) and tactical value, thus emphasizing the importance of timely data capture, analysis, and action within the streaming analytics process. For such a system to be functional, it requires a complex architecture ( Fig. 3.1 ) with the capability to support the swift collection and aggregation of highly variable data types, the ability to analyze that data in real time, and the ability to directly influence end users (or even automated computerized systems) to make and execute appropriate decisions and actions. Ideally, such a system would include measurable outcomes and similar events as input data elements from the data stream feedback loop. This enables a rapid-cycle process to assess system performance and to support system learning and refinement as needed, as would occur in a machine learning paradigm. An additional feature worth noting is the integration of historical data or low-frequency data elements that support this approach for patient-specific or population-based applications.
Development of a Real-Time Analytics Platform in the Cardiac Intensive Care Unit
In 2010 a team from the CICU at Boston Children’s Hospital, led by P. Laussen and M. Almodovar began development of the T3 system to capture, display, store, and analyze physiologic data streams from bedside monitoring devices. The term “T3” arose from the primary functional elements of the system as initially conceived to include the ability to track relevant physiologic data, enable assessment of a patient’s trajectory throughout the patient’s course in the CICU, and to support the ability to trigger appropriate responses or actions. The T3 data collection and analytics platform was designed as a Web-based, vendor agnostic, and scalable software system with three main features: an interactive data visualization user interface, a robust data analytics engine, and high-volume data storage capability. Fig. 3.2 describes an overview of the T3 platform beginning with data stream capture and ending with the multiple interfaces between its main features and data use applications. Fig. 3.3 shows a simplified version of the system architecture, noting its relatively simple structure and the ability to add third-party algorithms plus access to cumulatively stored data. The primary goals of the technology were to improve visualization of data trends for CICU patients, to develop algorithms and predictive models driven by the real-time physiologic data to guide decision making, and to create a hosting platform for the development and testing of algorithms using large volumes of stored data.
The T3 visualization user interface, as originally designed, is shown in Fig. 3.4 and demonstrates how multiple data streams are viewed simultaneously with trend patterns easily viewed in context with one another. The interactive interface allows selected parameters to be placed on the screen where desired and the ability to zoom in for higher resolution (down to 5-second interval averages between data points). The trend pattern can be examined along the timeline using a sliding, zoomable window for as long as the patient has been connected to a bedside monitor. The dashboard view in the upper right shows a sparkline summary of the recent data trends, the current parameter value, and user-determined boundary limits for each parameter. The user interface also allows for note entry or event annotation on the timeline (bottom) using free text entry or menu-driven options. This allows for decisions or other entries to be captured for easy querying and reporting at a later time.
