Bringing the Promise of Artificial Intelligence to Critical Care: What the Experience With Sepsis Analytics Can Teach Us

Abstract

In 1985, development of a computer system called "Deep Thought" began at Carnegie Mellon University with the lofty objective of developing an autonomous system capable of outperforming the world's top chess grandmasters. Later renamed "Deep Blue, " this chess-playing expert system defeated world champion Gary Kasparov in 1997 in a six-game match. However, it was not until 2017 that a deep artificial neural network algorithm known as "AlphaZero" achieved super-human performance in several challenging games, including Chess, Shogi, and Go (1). Such triumphs in computer-based technologies are common today as artificial intelligence (AI) applications, such as ChatGPT and DALL-E, are mimicking human capabilities, even passing medical board examinations (2). The term AI is used to describe the general ability of computers to emulate various characteristics of human intelligence, including pattern recognition, inference, and sequential decision-making, among others. Machine learning (ML) is a subset of AI that can learn the complex interactions or temporal relationships among multivariate risk factors without the need to hand-craft such features via expert knowledge (3). Retrospective studies have demonstrated ML applications are particularly useful for their diagnostic and prognostic capabilities leveraging vast quantity of data available in the ICU (4, 5). Certain ML algorithms have approached human performance at narrow tasks such as predicting resuscitation strategies in sepsis (6), need for mechanical ventilation (7), mortality in critically ill patients (8), and ICU length of stay (9).Sepsis is an attractive target for ML approaches as it is an inherently complex, common, costly, and deadly condition. Prediction of sepsis is the most common ML application described, although recent advances include approaches to optimize therapeutics and resuscitation strategies (6, 10). Given the potential to improve patient-centered outcomes and excitement about newer analytic approaches, it is no surprise that the number of ML algorithms aimed to improve sepsis care is increasing at a rapid rate. However, errors in sepsis prediction are often highlighted both in anecdotal and health system-wide failures that can be traced to poor implementation approaches, rudimentary ML algorithms, application of algorithms outside their intended use, or without proper maintenance. Noting these criticisms, what can be done at this point to demonstrate value of these predictive models? We believe that a revised focus on data enrichment, proper implementation, and rigorous testing is required to bring the promise of AI to the ICU.

Impact of a deep learning sepsis prediction model on quality of care and survival

“…Despite major interest in strategies to relieve the morbidity and mortality of sepsis, novel therapeutics have failed to translate into meaningful patient-centered outcomes. The potential to improve care through the use of artificial intelligence is attractive, particularly with advances in machine learning in the past decade 22 , 23 . Unfortunately, the majority of algorithms designed to predict sepsis never make it to the bedside 24 .…”

Clinician Prediction of Survival vs Calculated Prediction Scores in Patients Requiring Extracorporeal Membrane Oxygenation

“…Future research is indicated on whether artificial intelligence deep learning models can be used to enhance prediction tools for both the purposes of patient selection and survival, as has begun in other areas of clinical practice. 16…”

“…Our results suggest that combining the SAVE/RESP score with a comprehensive evaluation by an experienced multidisciplinary ECMO team would be best for decision making on determining ECMO candidacy and long-term prognosis until better prediction models are developed. Future research is indicated on whether artificial intelligence deep learning models can be used to enhance prediction tools for both the purposes of patient selection and survival, as has begun in other areas of clinical practice 16 …”

A Delphi Process to Identify Relevant Outcomes That May Be Associated With a Predictive Analytic Tool to Detect Hemodynamic Deterioration in the Intensive Care Unit

“…However, a recent systematic review of 494 AI studies using data obtained during an ICU stay found none that reported on patient outcomes following AI model integration [ 14 ]. Implementing AI tools has been described as an “afterthought” compared with investment in model development [ 16 ]. It could be speculated that these tools are not being studied and developed in a manner relevant to bedside clinicians, which could be a barrier to the clinical adoption of a particular tool.…”