Artificial intelligence in the intensive care unit
Area under the curve
Intensive care unit
The use of artificial intelligence (AI) in healthcare is receiving increasing interest, driven by a surge in scientific research and funding. AI has shown ophthalmologist-level performance at detecting retinal pathology  and can provide individualised treatment decisions for sepsis that could improve patient outcomes . There are many potential applications in the intensive care unit (ICU), particularly given the large amounts of data collected routinely. However, there are some important considerations for ensuring successful implementation.
Why artificial intelligence?
There is a paucity of positive multi-centre prospective randomised control trials in ICU settings. This reflects the challenge of running studies in such environments, where multiple treatments are given simultaneously to individuals who respond in variable ways based on their individual physiology. The resultant absence of strong guidelines means that clinician decision-making is driven largely by experience and instinct, resulting in significant variability amongst clinicians.
AI could reduce this inter-clinician variability and offer other benefits. AI excels at finding complex relationships in large volumes of data and can simultaneously and rapidly analyse many variables to predict outcomes of interest, such as sepsis or mortality. The modern ICU environment is data-rich, providing fertile soil for the development of more accurate predictive models, better decision support tools, and greater personalisation of care.
Predictive models and decision support tools
Severity scoring and mortality prediction
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 .
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 . Pirracchio et al.  and Aczon et al.  used multiple machine learning (ML) methods to achieve superior areas under the curve (AUCs) of 0.94 and 0.93 respectively.
Prediction of sepsis
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 .
The AI model by Nemati et al. predicted sepsis 12 h before onset with an AUC of 0.83 . Kamaleswaran et al. used multiple ML techniques to identify novel predictive markers . They found that variations in vital signs, processed by AI, could identify children who would develop severe sepsis . Without the wait for laboratory results, earlier treatment is enabled.
Decision support in mechanical ventilation
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 . 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 . The AI algorithm of Prasad et al., used to advise when to wean, outperformed clinical practice in terms of number of re-intubations .
Improving data with new technology
Novel types of data, collected by new methods, enable improvement of existing models and development of new tools. Davoudi et al. explored the use of wearable sensors, light and sound sensors, and a camera to collect data on ICU patients and their environments . The authors found that delirious patients were more sensitive to light but not to noise. Such data analysis could improve detection of delirium and enable real-time interventions to improve sleep hygiene . Pickham et al. detected patient positioning with wearables to identify when patients should be turned and thus reduced hospital-acquired pressure injuries . A variety of AI methods allow the processing of these novel types of data, such as convolutional neural networks which excel at analysing visual information.
AI is also taking advantage of the move towards higher-resolution continuous data capture. For example, deep learning analysis of electrocardiogram data, measured continuously in ICU patients, can detect ST-changes  and a broad range of arrhythmias .
While AI may enable the development of accurate tools, their introduction must follow careful consideration of real-world clinical utility, efficiency and existing workflows. The use of AI should be appropriately weighted alongside other sources of information and should be validated by well-designed prospective studies before widespread implementation.
Training AI algorithms requires integrated, well-structured data. Many ICUs use a combination of paper and electronic data and do not electronically link together data collected from multiple sources. Data are often incomplete or incorrectly entered. It is also heterogeneous, and many different tests are measured at different times. However, many AI models have achieved accuracy despite these factors.
Healthcare data are sensitive so data security and patient privacy are important considerations. Appropriate consent must also be obtained for data collection, yet many ICU patients lack sufficient capacity until recovery.
ICU doctors are often required to analyse large volumes of complex, heterogeneous data to make life-critical decisions. AI, if used effectively, could reduce this burden by transforming data into more actionable information. We can use AI to predict adverse outcomes before they happen, better manage highly complex situations, and ultimately allow clinicians to spend less time analysing data and more time harnessing their experience and human touch in delivering care.
Availability of data and materials
Data sharing is not applicable to this article as no datasets were generated or analysed during the current study.
CAL, VB, and MM all conceptualised and wrote the paper. All authors approved the final version. CAL acts as guarantor of the study.
Ethics approval and consent to participate
Approval and consent were not required. No primary research was undertaken, and there was no involvement of human subjects, human material or human data.
Consent for publication
This article contains no details, images or videos relating to an individual person and so no specific consent was required.
CAL and VB are employees of Cera Care. MM is an investor and employee of Cera Care. Cera Care is a domiciliary care provider conducting research into how artificial intelligence can be used to improve the care delivered to elderly people living at home.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
- 6.Aczon M, Ledbetter D, Ho L, Gunny A, Flynn A, Williams J, et al. Dynamic Mortality Risk Predictions in Pediatric Critical Care Using Recurrent Neural Networks. ArXiv170106675 Cs Math Q-Bio Stat. 2017 Jan 23; Available from: http://arxiv.org/abs/1701.06675. Accessed 4 Jan 2019.
- 10.Prasad N, Cheng L-F, Chivers C, Draugelis M, Engelhardt BE. A Reinforcement Learning Approach to Weaning of Mechanical Ventilation in Intensive Care Units. ArXiv170406300 Cs. 2017 Apr 20; Available from: http://arxiv.org/abs/1704.06300. Accessed 4 Jan 2019.
- 12.Davoudi A, Malhotra KR, Shickel B, Siegel S, Williams S, Ruppert M, et al. The Intelligent ICU Pilot Study: Using Artificial Intelligence Technology for Autonomous Patient Monitoring. ArXiv180410201 Cs Eess. 2018 Apr 25; Available from: http://arxiv.org/abs/1804.10201. Accessed 4 Jan 2019.
- 15.Rajpurkar P, Hannun AY, Haghpanahi M, Bourn C, Ng AY. Cardiologist-Level Arrhythmia Detection with Convolutional Neural Networks. ArXiv170701836 Cs. 2017 Jul 6; Available from: http://arxiv.org/abs/1707.01836. Accessed 4 Jan 2019.
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.