++
Over the last five decades, critical care researchers and clinicians have amassed a great body of pathophysiologic and clinical knowledge that has advanced the care of critically ill patients through the use of data. These data have informed the development of severity-of-illness scores designed to risk stratify the outcomes of critically ill patients.
The most well-known severity-of-illness scores are the Acute Physiology and Chronic Health Evaluation (APACHE), the Mortality Probability Models (MPM), Simplified Acute Physiology Score (SAPS), and Sequential Organ Failure Assessment (SOFA). Most of these have been updated over time and evaluated in diverse populations.
The primary uses of these scores have been for clinical investigation (to compare severity of illness between patient cohorts) and ICU performance assessment (to compare process and outcome metrics over time or between health care settings).
An emerging frontier in the care of critically ill patients seeks to use real-time risk prediction tools to inform bedside clinical decisions that can mitigate further patient deterioration, identify patients at risk for organ failure, and inform longer-term outcomes following intensive care.
The use of innovative statistical, machine learning, and artificial intelligence algorithms that leverage detailed data from electronic health records and physiologic monitors is expected to continue to produce significant advances in clinicians’ capabilities to risk stratify, phenotype, and intervene in critical illness.
However, a number of key challenges related to risk prediction accuracy and bias; clinician training and workflow integration; and patient engagement remain largely unsolved. Thus, as of today, only a paucity of these predictive models have produced robust evidence of benefit in critical care outcomes.
++
The use of clinical and physiologic data to risk stratify and predict patient outcomes has a rich history in critical care. Starting 50 years ago, severity-of-illness scoring systems were developed to evaluate the delivery of care and predict outcomes among critically ill patients admitted to intensive care units (ICUs). Recent advances in data availability and algorithm development have built off of this history and seek to bring real-time prediction tools and artificial intelligence/machine learning (AI/ML) to bedside intensive care. The purpose of this chapter is to review the scientific basis for severity-of-illness scoring systems, describe emerging advances in ML and predictive modeling in critical care, and discuss the promise and pitfalls that accompany the use of these severity-of-illness and risk prediction tools at the bedside.
+++
ROLE OF SEVERITY-OF-ILLNESS SCORING SYSTEMS
++
Traditionally, three major purposes of severity-of-illness scoring systems have been identified.1–5 First, scoring systems have been used in research studies to account and adjust for cohort severity of illness. For example, randomized controlled trials (RCTs) use scores to assess the balance of treatment and control group severity of illness at baseline and over time.6–30 Scoring systems are also used in observational clinical research which use statistical techniques (eg, matching, propensity scoring, and causal inference methods) to identify potentially causal effects ...