Artificial Intelligence in Critical Care

Abstract

The modern father of artificial intelligence (AI), Professor Andrew Ng (Stanford University), once described AI as the “new electricity.” Over the last few years, research and innovation in the application of AI in health care has occurred primarily because of increasing availability of data from electronic health records (EHRs) and ever-expanding computational power. Critical care has been at the forefront, evaluating applications in preventive analytics, imaging, smart devices, and administration. Although still evolving, the future holds the promise of innovative solutions assisting clinicians in smarter real-time decision-making, delivery of advanced diagnostics, and cost-effective therapeutics. Concerns around AI interpretability, generalization, scalability, and data privacy are of utmost importance. Through collaboration and thoughtful investigation, clinicians and AI scientists can foster research, development, and implementation of AI. The opportunity is available within critical care, one of the most time-sensitive, data-rich, and complex patient care environments in health care. The official introduction of AI occurred during the 1956 Dartmouth Artificial Intelligence Conference. Since that time, widespread implementation, particularly in health care, has gained traction.1,2 Adoption of EHR systems and the associated increase in accessible health care data have made it possible to use various AI techniques. Novel techniques have evolved for more efficient data extraction and preprocessing, allowing “big data” to be used in a scalable, efficient, and effective manner for machine learning (ML).3 Although initial research has focused on the use of retrospective health care data to develop advanced AI algorithms, pioneering clinical trials are evolving into Food and Drug Administration-approved algorithms and paving the way for the implementation of AI in clinical settings. Several key technical developments have allowed the rapid rise in health care AI. These include the adaptability of deep learning models to analyze heterogenous data sets, the rapid development of large medical data sets, availability of open-source AI algorithms, and ability to improve performance within larger data sets.3 There is no greater place for generation of health care data than in the intensive care unit (ICU). It is a data-driven, dynamic environment that is perfect for AI to thrive and aid clinicians in their care of health care’s most complex patients. The implementation of AI in the ICU has long been postulated. In 1995, Drs Hanson and Marshall4 had suggested that AI could reduce the cost of care and improve ICU patient outcomes. Considerable health care research has since been carried out using various AI techniques and data available from ICUs. One of the most used open-source data sets using natural language processing (NLP) and ML is the “Medical Information Mart for Intensive Care” (MIMIC-III), a deidentified data set that contains ∼58,000 hospital admissions of 47,000 patients who stayed in the ICUs of the Beth Israel Deaconess Medical Center between 2001 and 2012.5 This data set not only includes laboratory data, imaging reports, and vital signs, but most importantly clinical notes from physicians and nursing. The interest of the scientific community in this area is evident by the fact that the number of publications in AI and ML in critical care has evolved at a steady pace from the year 2000, and has tripled in the last 10 years6 (Fig. 1). AI is defined as the theory and development of computer systems able to perform tasks that normally require human intelligence, such as visual perception, speech recognition, decision-making, and translation between languages. ML is a subset of AI that allows for statistical analysis models to be developed on available data sets using computational sciences.