Machine-learning algorithm to predict ICU readmissions in the cardiothoracic surgical population

The Problem

The concept of failure to rescue (FTR) embodies the idea that, although not every complication of medical care is preventable, health care systems should be able to rapidly identify and treat complications when they occur. A resilient health-care team is able to identify changes in a patient’s condition quickly and to act on those changes in a manner that benefits the patient’s outcomes. A growing body of evidence implicates FTR after surgery (ie, patient death after a postoperative complication) as a significant contributor to the variability seen in surgical outcomes. As a safety and quality measure, FTR has been defined as the inability to prevent death after the development of a postoperative complication.1 CMS and AHRQ have introduced FTR measures as part of the Patient Safety Indicators (PSIs). PSI-04 (death rate among surgical inpatients with serious treatable conditions) has continually been updated (to take advantage of changes in coding practices, conditions present on admission, and the ICD-10 coding system). This PSI is publicly reported by CMS, and a risk-adjusted average national rate of 16.2% is presently reported on Hospital Compare. For the most recent reporting period (Oct 2015-June 2017), the overall rate at DUH (across all surgical platforms) was 19.2%, with a performance indicator rating of “worse than national average”.2 Many specialty surgical services, including adult cardiac surgery, have developed context-specific approaches to measure FTR. Reports from these studies using secondary analyses of large data sets indicate significant variability in FTR rates between hospitals (10.9-13.3%), and concluded that promoting early identification of clinical deterioration and timely treatment of postoperative complications (rescue) is a critical component of quality improvement.3

Delayed escalation of care has been associated with increased FTR rates and adverse surgical outcomes, and systematic reviews have reported delayed identification of patient deterioration as a key contributor.4 Critical care clinicians need predictive analytic tools to assist in reduction of near misses, failure to recognize, failure to rescue, and ICU readmission events, by: 1) gathering and pooling together high-resolution physiologic patient data from disparate devices and clinical information systems; 2) automatically analyzing the data in a timely manner to extract clinically relevant information; and3) efficiently interfacing with the data and the predictive analytics to make quick accurate clinical decisions. The overall long-term objective of our team is to develop a framework for concurrent multi-patient, multi-diagnosis and multi-stream (ie, multidimensional) temporal analysis of intensive care unit (ICU) and step-down unit (SDU) data in real-time for clinical management and historically for clinical research. The real-time component for clinical decision making would utilize new stream processing approaches while the clinical research component would utilize new approaches to data mining (ie, machine-learning) more suited to the analysis of physiological stream behaviors and within the healthcare context. In doing so, our investigative team seeks to address several components of the knowledge gap for understanding at the patient-level where and why failures to rescue occur, while acknowledging that a) the delay in diagnosing and treating complications may be related to processes, structure, or both; and b) that context (type/acuity of surgical population) may influence the performance of predictive modelling and the effectiveness of interventions to reduce FTR. The objective of this proposal, which is part of our long-term goal, is to combine physiologic, laboratory, and biometric data from cardiothoracic surgical (CTS) patients into an analytical platform to identify associations and make predictions at a key transition of care –namely between postoperative ICU care and SDU care -that can inform improvements in CTS quality, outcomes, and value of care in this resource-intensive patient population.

Limitations of static analyses vs real-time streaming analytics for CTS patients: The main clinical registry data for CTS (the STS registry) involves systematic data collection and capture from EHR (with standardized data elements and definitions), and is used to measure quality of care, provide quality benchmarks, and conduct clinical research –but it not available in a timely manner (data is submitted to the registry after the episode of care) and therefore not actionable. Conversely, individual patient data available in real-time, and subjected to innovative data science analytical methodology to identify correlation and patterns amid complex data and data associated user-friendly visualization tools, could potentially overcome the limitations of perioperative clinical and administrative risk scores, by providing actionable information to improve health-care delivery (guiding clinical decisions and prioritization of care) and outcomes in CTS patients. In summary, the promise of the proposed predictive algorithm stems from inclusion of larger, more diverse data sets than traditional risk models, with potentially higher accuracy or risk prediction, and available in real-time at the point of care.

Our Solution

Clinicians participating in the postoperative care of CTS patients in the ICU and SDUs deal with large volumes of rapidly changing data from multiple devices that constantly monitor vital organs. We have reached a critical crossroad where the ability to gather this information has outpaced our ability to aggregate and interpret it in a clinically meaningful way, and predict accurately and timely patient instability or catastrophic conditions. These issues are amplified during transitions of care, typically between a highly monitored environment (ICU) anda less monitored one (SDU). The current proposal seeks to develop predictive models to identify rapidly and accurately CTS patients at high-risk (and high-cost) for postoperative clinical deterioration necessitating readmission from the SDU back to the ICU.The technology proposed herein addresses both the paralyzing weight of data overload as well as the “alarms race” by integrating vital sign, laboratory, and biometric information into an actionable index that predicts patient deterioration.

References