Artificial intelligence in the intensive care unit

Given the complexity and heterogeneity of ICU patients, scoring systems have been generated to record severity of illness and predict probability of mortality. They can assist in clinical decision-making, comparisons of quality of care, and stratification for clinical trials. However, they do not incorporate variations between departments, regions and countries and perform better on entire ICU populations than on individuals or subsets [3].

AI is well suited for developing algorithms which overcome these limitations and also increase prediction accuracy. The artificial neural network of Dybowski et al. could be re-trained in individual ICUs, tailoring predictions to that unit [4]. Pirracchio et al. [5] and Aczon et al. [6] used multiple machine learning (ML) methods to achieve superior areas under the curve (AUCs) of 0.94 and 0.93 respectively.

Early detection and prediction of sepsis enable earlier treatment and better outcomes, yet sepsis is often unclear until late stages. Existing tools have poor predictive accuracy and often rely on time-consuming laboratory results. Desautels et al. found that, in 22,853 ICU stays, systemic inflammatory response syndrome (SIRS), Simplified Acute Physiology Score II (SAPS II) and sequential organ failure assessment (SOFA) had AUCs of 0.609, 0.700 and 0.725 respectively for identifying sepsis at the time of onset [7].

The AI model by Nemati et al. predicted sepsis 12 h before onset with an AUC of 0.83 [8]. Kamaleswaran et al. used multiple ML techniques to identify novel predictive markers [9]. They found that variations in vital signs, processed by AI, could identify children who would develop severe sepsis [9]. Without the wait for laboratory results, earlier treatment is enabled.

Mechanical ventilation is one of the most common interventions in ICU patients. Appropriate levels of sedation and analgesia are important but complicated by significant inter-patient variability. Timing of ventilator removal is also important, as both premature extubation and prolonged ventilation are associated with higher mortality rates. However, a wide discrepancy of practices is seen and accurate prediction is challenging.

An AI tool may enable more personalised sedation and analgesia to reduce inter-clinician variability. The algorithm of Prasad et al. outperformed clinical practice, as measured by regulation of vital signs [10]. Using AI to guide extubation timing is challenging. Such algorithms are trained using outcome data, such as the timing of removal and whether it was successful. However, successful extubation only indicates that it was ready at that point, and doesn't rule out being ready at an earlier stage. This is also true in the reverse for premature removal. Despite this, the algorithm of Parreco et al. predicted the need for prolonged ventilation with an AUC of 0.82 [11]. The AI algorithm of Prasad et al., used to advise when to wean, outperformed clinical practice in terms of number of re-intubations [10].