VOLUME 23

2023

ISSUE 4

Artificial
Intelligence
in the ICU
Towards Artificial Intelligence as a Decision Support Individualisation of Mechanical Ventilation in
Tool to Combat AMR in the ICU, TD Corte, J Verhaeghe, Obstructive Lung Disease: Not All Ventilated
F Ongenae, SV Hoecke, JJ De Waele Patients Have ARDS, AMA Álvarez, IT Pérez,
SP Martínez, FG Vidal
Why Artificial Intelligence is Not Fixing the Problem
of Sepsis in the Hospital, V Herasevich, D Diedrich, The Mechanical Power as a Guide for Protective
B Pickering Mechanical Ventilation in Patients With and Without
ARDS, D Chiumello, G Catozzi, S Coppola
Artificial Intelligence in Anaesthesia and Critical
Care - Temptations and Pitfalls, F Michard, Seven Myths of Mechanical Ventilation in Paediatric
FA Gonzalez, P Schoettker and Neonatal Patients, DP Rojas-Flores, JN Soriano-
Martínez, R Toledo-Pérez, G Montesinos-Gómez
Transforming a PICU in the Digital Age, AF Villalobos,
EE Torné, FG Cuyàs, RC Llandrich, FXE Elías, P Garcia-
Canadilla, FL Ruiz, MIN Martínez, GD Carrillo, IJ Garcia,
FJC Lasaosa

INTENSIVE CARE I EMERGENCY MEDICINE I ANAESTHESIOLOGY icu-management.org @ICU_Management

 ARTIFICIAL INTELLIGENCE IN THE ICU 159

Frederic Michard
MiCo
Vallamand, Switzerland
[email protected]
Artificial Intelligence in
Anaesthesia and Critical Care -
@MichardFrederic
michardconsulting.com

Temptations and Pitfalls

Filipe A Gonzalez
Intensive Care Unit
Big data, Artificial Intelligence (AI) and machine learning are buzzwords. In this
Hospital Garcia de Orta, Almada
University of Lisboa article, we briefly discuss what they mean for anaesthesiologists and intensivists,
Portugal
[email protected]
focusing on existing clinical applications.

in the field of imaging (radiology and autoVTI algorithm may help trainees to be
pathology). It is indeed relatively easy to as efficient as echocardiography experts in
train an algorithm with a large database estimating VTI, stroke volume (SV ~ VTI
of images so that it becomes capable of x Pi) and cardiac output using ultrasounds
Patrick Schoettker detecting abnormalities that could be (Gonzalez et al. 2022). Several ML algo-
Department of Anesthesiology
CHUV missed by a medical trainee or a seasoned rithms have also been developed for the
University Hospital Lausanne but distracted clinician. In this respect, automatic estimation of left ventricular
Switzerland
[email protected]
many ML algorithms have been designed ejection fraction (LVEF). Comparison
to analyse chest x-rays and CT scans and studies suggest they may enable novices
to suggest a diagnosis (e.g., tracheal tube to measure LVEF as accurately and with
not correctly positioned on the chest x-ray better reproducibility than experts taking
of a mechanically ventilated patient or manual measurements (Varudo et al. 2022).
CT scan images suggestive of COVID-19 Other ultrasound algorithms have been
in a patient with ARDS). Recently, ML designed to predict fluid responsiveness
We are today able to collect and store a algorithms have also been implemented in mechanically ventilated patients from
considerable amount of patient-related into ultrasound machines to facilitate and the automatic quantification of the inferior
data. These “big” data are typically part automate point-of-care echocardiographic vena cava respiratory variation or to detect
of Electronic Medical Record (EMR) evaluations (Nabi et al. 2019). pulmonary oedema from the automatic
systems and usually combine demographic, quantification of lung B lines.
clinical and biological information. They
may also contain images (e.g. ultrasound AI and Point-of-Care Echocar- In summary, the value of ML algorithms
diography to help novices perform point-of-care
cardiac images) and physiologic wave-
echocardiographic evaluations has been
forms. These data can be analysed with Several ML algorithms have been trained documented in several clinical studies.
simple descriptive methods to report to recognise heart images and guide users However, given the fact that the propor-
basic information regarding patient char- to hold and position their transthoracic tion of intensivists trained to perform
acteristics and outcomes such as hospital probe correctly. Such algorithms are also echocardiography is increasing sharply,
mortality, morbidity, and length of stay. able to grade image quality and label whether AI innovations are necessary
This approach, useful for benchmarking heart structures. An example is displayed to increase the number and quality of
and research, does not require artificial in Figure 1. Some ML algorithms can ultrasound haemodynamic evaluations
intelligence (AI). take echocardiographic measurements remains to be established.
A step further in the data analysis process automatically. For instance, the autoVTI
consists of using machine learning (ML) algorithm can recognise a 5-chamber apical
algorithms (a subfield of AI), which have view of the heart, automatically position AI and Continuous Blood Pressure
been trained to detect specific patterns the pulse wave Doppler caliper in the left Monitoring
of disease states or adverse events. As of ventricular outflow tract and record the In the search for cuffless and continuous
today, most ML innovations approved sub-aortic velocity time integral (VTI) blood pressure monitoring techniques, ML
for medical use have been developed over a short time window (Figure 1). A algorithms have been proposed to estimate
recent clinical evaluation suggests that the blood pressure and its changes from the

ICU Management & Practice 4 - 2023

160 ARTIFICIAL INTELLIGENCE IN THE ICU

the oscillometric brachial cuff method).

In these settings, the reference method
would be used not only to confirm changes
in blood pressure but also to recalibrate
the algorithm.

AI to Forecast Clinical Deterioration

As mentioned above, ML algorithms can
detect specific patterns of overt disease
states. They can also be trained to detect
patterns associated with pre-disease states
or patterns observed before the occurrence
of specific adverse events.
For instance, multiple ML algorithms
have been developed to create scores (e.g.,
eCART or HAVEN scores) predicting severe
adverse events in patients hospitalised in
regular hospital wards. Several studies have
shown these AI-derived scores are able to
predict ICU admission, cardiac arrest, and
death with an area under the curve (AUC)
around 0.8-0.9 (as a reminder, a random
guess would be associated with an AUC of
0.5 and a perfect prediction with an AUC
of 1.0). However, their predictive value is
frequently only slightly higher when not
simply comparable to what is possible to
achieve with existing scores such as the
modified early warning score (MEWS) or
the national early warning score (NEWS)
- both scores which are easy to calculate
Figure 1. Examples of machine learning algorithms designed for point-of-care echocardio- from vital sign spot-checks (Bartkowiak
graphy. Top - Auto guiding to obtain an apical 4-chamber view of the heart, auto-grading to et al. 2019).
ensure optimal image quality, and auto-labelling. From EchoNous (Redmond, WA, USA), with
permission. Bottom - Automatic detection of the apical 5-chamber view of the heart, automa- Multiple attempts have been made to
tic positioning of the Doppler caliper in the left ventricular outflow tract (LVOT) and automatic
measurements of subaortic velocity-time integral (VTI). From GE Healthcare (Chicago, IL, detect sepsis at an early stage, fasten thera-
USA) with permission. peutic management and improve patient
outcomes. As of today, the results of sepsis
“sniffer” implementation programmes
have been conflicting, with some report-
analysis of photoplethysmographic (PPG) dipping nor therapeutic changes in blood ing a decrease in time-to-antibiotic and
waveforms. Historically, PPG waveforms pressure (Tan et al. 2023). As a matter in-hospital mortality (Shimabukuro et
were recorded by medical-grade pulse of fact, these devices require frequent al. 2017), whereas others, including the
oximeters, but they are today frequently recalibrations and carry the potential to recent evaluation of the EPIC system
obtained from smartwatches, adhesive track changes in blood pressure over short (widely used in the US), reported poor
patches, optical bracelets, rings or smart- time periods rather than measure absolute discrimination (AUC 0.63) and calibration
phone cameras (Festo et al. 2023). A few numbers (Ghamri et al. 2020). Interestingly, in predicting the onset of sepsis (Wong et
of these devices, mainly designed for the this would not be an obstacle to their use al. 2021). Another potential ML applica-
detection or follow-up of patients with during surgery, in ICU patients or even tion is known as reinforcement learning.
chronic hypertension, have been cleared for in hospital wards to detect hypotensive It enables the development of algorithms
medical use. Recent independent clinical and hypertensive episodes and trigger designed to provide dynamic therapeutic
evaluations suggest they may not always be intermittent blood pressure spot-checks recommendations, which have been shown
able to detect the physiologic night-time with a reference clinical method (e.g., to be associated with improved organ func-

ICU Management & Practice 4 - 2023

 ARTIFICIAL INTELLIGENCE IN THE ICU 161

tion and/or survival (Komorowski 2018). fact that HPI is the mere reflection of the project, it is likely that brainwave data simply
Whether such prescriptive algorithms may mean arterial pressure (MAP) and, as a did not contain the predictive information
be accepted by clinicians (particularly result, that its predictive value may not they were looking for. In addition, some
by experts in sepsis management) and be superior to MAP monitoring (Mulder events are unpredictable by nature. As an
may improve clinical outcomes remains et al. 2023). example, which algorithm could predict
unknown. In summary, the predictive value of hypotension related to surgical injury (e.g.,
Machine learning algorithms have also machine learning algorithms is hardly vena cava injury during liver surgery) or
been developed and proposed to predict disputable. However, the superiority over the decision to deepen anaesthesia or
postoperative morbidity and mortality, existing and simpler methods often remains sedation with a propofol bolus? During
with reported AUCs that may exceed to be determined, and the complexity/ surgery and in ICUs, multiple external
0.9. However, this predictive value does benefit and cost/benefit ratio may therefore factors are susceptible to modify clinical
not always overcome what is possible to be questioned. trajectories in one direction or the other.
achieve with simple scores such as the When steady states do not exist, it becomes
SORT score (Wong et al. 2018). Of note, challenging to predict short-term clinical
The Pitfalls of Predictive Analytics trajectories (Michard and Teboul 2019).
the subjective prediction made by clinicians
has been shown to be associated with an Predictive analytics is associated with at Secondly, poor data quality is one of
AUC of 0.89! (Wong et al. 2018). Therefore, least four main limitations and/or pitfalls, the main factors holding up the big data
whether there is a need for complex ML which are summarised in Figure 2. revolution in healthcare (Dhindsa et al.
scores to predict postoperative outcomes The first one is to believe that every- 2018). This limitation is often summarised
remains debatable. thing is predictable. As highlighted by as “garbage in, garbage out”. Indeed, one
Machine learning algorithms have recent- Chen and Asch in a famous New Engl. J. may use the best predictive algorithm, but if
ly been proposed to predict haemodynamic Med. editorial (Chen and Asch 2017), “no we feed it with wrong data, artefacts and/or
instability and, more specifically, systemic amount of algorithmic finesse or computer damped physiologic waveforms, one may
hypotension. The hypotension prediction power can squeeze out information that logically end up with wrong predictions.
index (HPI) is a commercially available is not present”. Google X, an Alphabet Thirdly, it is paramount to understand
ML-derived score calculated from the subsidiary, reported that its initiative to that predicting does not necessarily mean
analysis of the arterial pressure waveform. discover a biomarker for depression and preventing. When the prediction is not
It has been shown to forecast intraoperative anxiety in brainwave data fell short of its followed by one or more appropriate actions
hypotension 5-15 minutes ahead with an goal. Given the fact that they had almost susceptible to modify the clinical trajec-
AUC ranging between 0.75-0.95. However, unlimited resources and an army of top- tory, logically, nothing can be prevented.
recent publications have highlighted the level computer scientists working on this In the largest HPI randomised controlled

Figure 2. The four main pitfalls of predictive analytics

ICU Management & Practice 4 - 2023

162 ARTIFICIAL INTELLIGENCE IN THE ICU

trial published so far (Maheshwari et al. harm to the patient. In contrast, giving ing and simpler methods. In the digital
2020), anaesthesiologists who were alerted antibiotics to a probability of sepsis or medicine era, whereas many medical
about the risk of hypotension failed to administering vasopressors to a prob- students are eager to work on AI projects
prevent hypotensive events. Interestingly, ability of hypotension might be risky and and to participate in datathons, it might be
it appeared that most of them did not feel is, therefore, questionable (Michard and useful to remind them that “the immedi-
the need and/or the right to give fluid, Futier 2023). Who would accept receiv- ate challenge to improving quality of care
vasopressors, or inotropes to patients ing treatment with known side effects is not discovering new knowledge, but
who were still haemodynamically stable for a predicted disease or adverse event rather how to integrate what we already
and only had a probability of becoming that may never occur? And who would know into practice” (Urbach and Baxter
hypotensive. This finding is an excellent be responsible in case of complications? 2005). Therefore, although we should
illustration of the reluctance of clinicians keep our eyes and ears wide open for AI
to trust and follow AI recommendations innovations, we should also continue to
Conclusion
(Gaube et al. 2021). focus on basic initiatives (more nurses and
Big data, AI, and, more specifically, machine doctors, better training with simulation,
Fourthly, there are risks associated with
learning algorithms are hot topics for better compliance to existing guidelines,
the treatment of probabilities. Therefore,
medical journals and scientific events. and better use of existing monitoring
one may hardly envision being proactive
For start-ups, they are also very useful tools) that are known to improve patient
from a therapeutic standpoint. One may
keywords to raise funds. However, one may outcomes and satisfaction.
be proactive by performing bacteriologi-
acknowledge that, as of today, and from a
cal samples when predicting sepsis or by
practical standpoint, the AI elephant gave
upgrading surveillance when predicting
birth to a mouse in the field of anaesthe- Conflict of Interest
clinical deterioration (e.g., by offering
siology and intensive care. Prospective FM is the founder and managing director
continuous monitoring and/or ICU admis-
clinical trials are indispensable not only of MiCo, a consulting and research firm
sion). There is no harm in doing so. There
to assess the safety of AI innovations but based in Switzerland. FAG and PS have
might be economic consequences, but no
also to demonstrate superiority over exist- nothing to disclose.

References artificial intelligence-enabled tool in critically ill patients. Tan I, Gnanenthiran SR, Chan J et al. (2023) Evaluation of the
Br J Anaesth. 129:e116-e119 ability of a commercially available cuffless wearable device to
Bartkowiak B, Snyder AM, Benjamin A et al. (2019) Validating track blood pressure changes. J Hypertension. 41:1003-10.
Komorowski M, Celi LA, Badawi O et al. (2018) The Artificial
the electronic cardiac arrest risk triage (eCART) score for
Intelligence Clinician learns optimal treatment strategies for Urbach DR, Baxter NN (2005) Reducing variation in surgical
risk stratification of surgical inpatients in the postoperative
sepsis in intensive care. Nat Med. 24:1716-20. care. BMJ. 330:1401-2.
setting: retrospective cohort study. Ann Surg. 269:1059-63.
Maheshwari K, Shimada T, Yang D et al. (2020) Hypotension Varudo R, Gonzalez FA, Leote J et al. (2022) Machine learning
Chen JH, Asch SM (2017) Machine learning and prediction
prediction index for prevention of hypotension during moderate- for the real-time assessment of left ventricular ejection fraction
in medicine - Beyond the peak of inflated expectations. N
to high-risk non-cardiac surgery. Anesthesiology. 133:1214-1222. in critically ill patients: a bedside evaluation by novices and
Engl J Med. 376:2507-2509.
experts in echocardiography. Crit Care. 26:386.
Michard F, Futier E (2023) Predicting intraoperative hypotension:
Dhindsa K, Bhandari M, Sonnadara RR (2018) What’s holding
from hope to hype and back to reality. Br J Anaesth.131: 199-201. Wong A, Otles E, Donnely JP et al. (2021) External validation
up the big data revolution in healthcare? BMJ. 363:k5357.
of a widely implemented proprietary sepsis prediction model
Michard F, Teboul JL (2019) Predictive analytics: beyond the
Festo C, Vannevel V, Ali H, et al. (2023) Accuracy of a smartphone in hospitalized patients. JAMA Intern Med. 181:1065-70.
buzz. Ann Intensive Care. 9:46.
application for blood pressure estimation in Bangladesh,
Wong DJN, Harris S, Sahni A et al. (2020) Developing and
South Africa, and Tanzania. npj Digit. Med. Mulder MP, Harmannij-Markusse M, Donker DW et al. (2023)
validating subjective and objective risk-assessment measures
Is continuous intraoperative monitoring of mean arterial
Gaube S, Suresh H, Raue M, et al. (2021) Do as AI say: for predicting mortality after major surgery: An international
pressure as good as the hypotension prediction index algorithm?
Susceptibility in deployment of clinical decision-aids. npj prospective cohort study. PLoS Med. 17:e1003253.
Anesthesiology. 138:657-658.
Digit. Med. 4:31.
Nabi W, Bansal A, Xu B (2021) Applications of artificial
Ghamri Y, Proença M, Hofmann G et al. (2020) Automated
intelligence and machine learning approaches in
pulse oximeter waveform analysis to track changes in blood
echocardiography. Echocardiography. 38:982-92.
pressure during anesthesia induction: a proof-of-concept
study. Anesth Analg. 130:1222-33. Shimabukuro DW, Barton CW, Feldman MD et al. (2017)
Effect of machine learning-based severe sepsis prediction
Gonzalez FA, Varudo R, Leote J et al. (2022) The automation
algorithm on patient survival and hospital length of stay: a
of sub-aortic velocity time integral measurements by
randomized clinical trial. BMJ Open Respir Res. 4:e000234.
transthoracic echocardiography: clinical evaluation of an

ICU Management & Practice 4 - 2023