Introduction
The official introduction of AI occurred during the 1956 Dartmouth Artificial Intelligence
Conference. John McCarthy coined the term “artificial intelligence” in 1955, and is
one of the “founding fathers” of artificial intelligence together with Alan Turing,
Marvin Minsky, Allen Newell, and Herbert A. Simon.[1 ] Since then, there has been widespread implementation of AI in all sectors including
healthcare[2 ]
[3 ] AI is defined as the development of computer systems to model intelligent behavior
with minimal human intervention.
The application of AI in medicine has two main branches: 1) virtual branch and 2)
physical branch.
The physical branch comprises highly repetitive work. It empowers the doctors to deliver
faster and more accurate clinical care by offering them expertise and assistance.
The virtual component is represented by Machine Learning (ML), mathematical algorithms
that improve learning through experience.
Health Data Management
The health data are nowadays available in electronic health records (EHRs). The medical
data include numerical information, laboratory test results, genetic tests, culture
results, images, treatment information, administrative data, and health research information.
Clinical data stored in EHR are both structured and unstructured[4 ]
[5 ] ([Fig. 1 ]).
Fig. 1 The cycle of medical data generation through medical health records, NPL to ML to
clinical treatment and prediction. ML, machine learning; NPL, natural language processing.
Structured Data
Structured data follow a prescribed data model and value set, constraining the users
to only be able to choose predetermined values. Computers can readily process structured
data. Data sent by medical devices to EHRs are usually structured data.
Unstructured Data
Unstructured data do not follow a predefined set of values, allowing users to instead
enter narrative information about data using their own words. This means recording
data provides the user with the most freedom for recording an entry, but because the
same clinical event could be documented in myriad ways, computers cannot easily process
unstructured data, making errors more likely. These data have to be converted to computer
readable data through the natural language processing (NLP) methods.[5 ]
[6 ]
Machine Learning Algorithms
Machine learning (ML) can be classified as follows[4 ]
[7 ]
[8 ]
[9 ]:
Unsupervised
Supervised
Deep learning
Reinforcement learning
Unsupervised Machine Learning
Unsupervised learning is a ML technique, where the model works on its own to discover
information and the outcomes of the model are not defined.
It performs more complex processing tasks compared with supervised learning but becomes
unpredictable compared with other ML techniques and is less accurate.
Clustering, association, and principal component analysis (PCA) fall into unsupervised
techniques. Clustering finds out the structure and pattern in data and identifies
different groups, whereas association establishes the relationship in the datasets
from the given database. PCA is mainly used for the dimension reduction of data.
Supervised Machine Learning
The input and output variables are provided in a supervised learning model. Thus,
specified data are used to train algorithms, and a link is established between input
and output variables in a supervised learning model. These techniques are highly precise.
The supervised learning techniques are regression and classification. Classification
separates the data, whereas regression fits the data.
The following algorithms are used in supervised ML techniques: decision tree, random
forest (RF), Naïve Bayes (NB), support vector machine (SVM), artificial neural networks
(ANN), discriminant analysis, nearest neighbor, linear regression, and logistic regression.
The SVM and ANN are frequently used in the medical field.
Support Vector Machine
This supervised learning algorithm classifies the data into two categories. The model
is built from the data already sorted in two categories ([Fig. 2 ]). This makes SVM a kind of nonbinary linear classifier. SVMs are used in text categorization,
image classification, prediction, and handwriting recognition.
Fig. 2 Depiction of SVM classifying the data into two categories and data reduction by PCA.
PCA, principal component analysis; SVM, supervised learning algorithm.
Artificial Neural Network
The neural network captures complex nonlinear relationships between input and outcome
variables by multiple hidden layers (HLs) with prior specified functions. The weights
are established through the input and outcome data; thus, the average error is reduced
and the predictions become more accurate ([Fig. 3 ]).
Fig. 3 A diagram representing ANN, showing the outcome of head injury; the input variables
are sex (M–male, F–female), age, RR, GCS, extracranial injuries, CT scan of the midline
shift, and whether surgery was performed in binary fashion. The HL is only one. ANN,
artificial neural network; CT, computed tomography; GCS, Glasgow coma scale; HL, hidden
layer; RR, respiratory rate.
Deep Learning
Deep learning is a self-teaching system in which the existing data are used to train
algorithms to find the patterns and then make predictions about new data. The ANNs
with multiple layers of nodes create deep learning algorithms that mimic the network
of neurons of the brain. This algorithm with multiple cycles defines patterns and
improves the precision of predictions with each cycle ([Fig. 4 ]).
Fig. 4 Depiction of deep learning of ANN with the following input variables: sex (M–male,
F–female), age, RR, GCS, extracranial injuries, CT scan of the midline shift, and
whether surgery was performed in the binary fashion with three hidden layers of nodes
and the head injury outcome. ANN, artificial neural network; CT, computed tomography;
GCS, Glasgow coma scale; HL, hidden layer; RR, respiratory rate.
Reinforcement Learning
Like deep learning, reinforcement learning is autonomous. But deep learning is learning
from a training set and then applying that learning to a new dataset, while reinforcement
learning is dynamically learning by adjusting actions, based on continuous feedback,
to maximize a reward.[3 ]
[6 ]
[7 ]
[8 ]
Neuroanesthesiology and neurocritical care as a discipline is rendered difficult due
to the inherent limitations in the assessment of patients with neurological injury.
As neuroanesthesiologists, we work in operating rooms and intensive care units (ICUs),
both being acute care settings, which demand vigilance, steady hands, and quick thinking.
AI can definitely assist us in making better clinical decisions and provide up-to-date
medical information from journals, textbooks, and clinical practices. This results
in early diagnosis, predicts outcome of the disease as well as treatment, provides
feedback on treatment, and reduces errors. This greatly increases patient safety and
saves costs. These traits allow AI systems to continuously monitor and treat neurocritical
care patients in real-time. Early signs of neurological deterioration could be detected
more promptly and appropriate measures taken more quickly, thereby improving patient
outcomes. AI can also help patients in areas where neurocritical care is not available,[10 ] as they take over more of the basic patient management, by analyzing the data and
titrating treatments in real-time, reducing possible delays in patient care and optimizing
the patient’s condition till he/she is transferred to a higher center with neurocritical
care facility.
In this review, we have discussed the applications of AI in neuroanesthesiology and
neurocritical care, the barriers to its implementation, and the future trends in this
field.
Applications in Neuroanesthesiology and Neurocritical Care
AI creates a potential system to manage the neuroanesthesiology and neurocritical
care patient with minimal or no supervision, freeing the clinician to focus attention
elsewhere.[11 ]
[12 ]
[13 ] Some potential parameters include anesthetics/analgesics, antiepileptic drugs (AEDs),
blood pressure, glucose, fluids/electrolytes, neuromuscular blockade, and ventilator
settings.[14 ]
[15 ]
[16 ]
[17 ]
[18 ]
[19 ]
[20 ]
[21 ]
[22 ]
[23 ]
[24 ]
[25 ] The applications of AI in our field can be broadly discussed under five categories
(1) predictive analytics, (2) imaging, (3) smart devices, (4) Administration, and
(5) research and education.
Predictive Analytics
Much of the work in critical care using AI has focused on predictive analytics. Improvement
in the prediction of adverse events such as hypotension has been shown using advanced
ML models in critical care environments.26 ML methods can predict the risk of
postinduction hypotension.[27 ]
[28 ] These models enables detection and intervention up to 15 minutes before an event
and have been generalized for use in the multicenter clinical environment.[29 ] Predictive therapeutic interventions to prevent hypotension using AI have also been
constructed for fluid resuscitation.[30 ] By combining models for early hypotension detection and therapeutic intervention,
there is potential to prevent or minimize patient deterioration and the subsequent
development of multisystem organ dysfunction.
Sepsis remains one of the largest causes of mortality in the ICU. Sepsis algorithms
become more important and these interpretable models can predict sepsis 4 to 12 hours
before clinical recognition.[31 ]
ML models can also help in the prediction of hospital-acquired infections such as
central line-associated blood stream infection and Clostridium difficile infections.[32 ]
Prediction of prolonged mechanical ventilation is useful for early tracheostomy, ventilator
weaning, and rehabilitation. Use of AI to identify patients who will require > 7 days
of mechanical ventilation has been shown to improve outcomes.33 Teams are also using
AI to aid ventilator weaning by targeting the success of extubation. Kuo et al[34 ] used neural networks to create a model with an accuracy of 80% and improved on traditional
prediction by rapid shallow breathing index.
AI has also been implemented in specialized ICUs such as in neurointensive care for
early and accurate risk assessment of seizures in critically ill patients.[35 ]
Predictive ML models for patient trajectories[36 ] and ICU readmission have been developed and have shown higher predictive values
than the conventionally used stability and workload index for transfer score or the
modified early warning score criteria for early deterioration.[37 ]
Mortality is a common outcome in medical studies, and prediction capabilities related
to it have been studied extensively using ML and NLP. Use of NLP enables inclusion
of the traditionally difficult-to-use clinical notes. Weissman et al[38 ] showed the ability to use unstructured data such as clinical notes, and terms such
as “poor prognosis,” by using various NLP techniques. Using neural networks, classification
algorithms can be constructed for identification of the most important terms in physician
notes, which then can be used to construct ML models to predict outcomes such as mortality
in the surgical ICU.[39 ] One such model, called Early Mortality Prediction for Intensive Care Unit patients,
has been shown to outperform traditional scoring systems such as acute physiology
and chronic health evaluation (APACHE) and sequential organ failure assessment (SOFA)
despite missing values within the training datasets. The area under the curve (AUC)
is 0.82 ± 0.04 compared with the traditional scoring systems which range from 0.54
to 0.65.[40 ] This model has not only focused on prediction accuracy, but has also attempted to
generate earlier prediction by hours using multimodal data. The benefits of such models
once again lie in triage, early intervention, and appropriate treatment recommendation
to minimize risk to the patient and provide for cost-effective care.
Harnessing meaningful information from EHR data and data registries is expensive,
can be of limited value, and is utilized primarily for retrospective research analysis.
As we have seen from the above examples, ML provides a more cost-effective way to
carry out retrospective research and, in constructing models, can provide real-time
or prospective guidance to clinicians.[41 ]
Machine-learning models have been developed for predicting mortality following trauma
in motorcycle riders. ANNs have been used to predict outcome following head injury.[42 ]
[43 ]
The occurrence of symptomatic cerebral vasospasm (SCV) after aneurysmal subarachnoid
hemorrhage (aSAH) is a morbid and common problem. A simple ANN model was found to
be more sensitive and specific than multiple logistic regression (MLR) models in prediction
of SCV in patients with aSAH.44 ML has been used to predict outcome in intracranial
aneurysms treated with flow diverters.[45 ]
Hollon et al[46 ] sought to build a predictive model using supervised ML to accurately predict early
outcomes of pituitary adenoma surgery. These results provide insight into how predictive
modeling using ML can be used to improve the perioperative management of pituitary
adenoma patients.
Stroke is one of the major causes of disability and death worldwide. It is estimated
that up to 80% of strokes can be prevented if one can identify or predict the occurrence
of stroke in its early stage.[47 ] AI-based methods offer several advantages in improving prediction performance for
stroke treatment, prognosis, and functional outcome prediction. This helps neurophysicians
to identify high-risk patients and guide treatment approaches, leading to decreased
morbidity. Several AI-based techniques are being investigated to develop automated
platforms for precisely predicting prognosis and the functional outcome. Park et al[48 ] have proposed a Bayesian network model for the prediction of poststroke outcomes
with the available risk factors. They also introduced an online “Yonsei stroke outcome
inference system” for predicting functional independence at 3 months and mortality
within 1 year in patients with stroke using the Bayesian network model.
The timely diagnosis of stroke is crucial for good functional recovery and minimizing
mortality. AI offers technology solutions with high-precision and accuracy for the
diagnosis of stroke, its severity, as well as prediction of functional outcomes.[49 ]
Recently, in diagnostic neuroradiology, there has been an interest in adopting AI
and ML techniques[50 ]
[51 ] and in the prediction of the outcome in patients postneurointerventional procedures.[52 ]
[53 ] Two recent studies[52 ]
[53 ] have used ANN modeling and supported vector machine algorithms in prediction of
the final Modified Rankin Score (mRS) with relatively good accuracy and precision.
The accuracy of outcome prediction, using supervised ML algorithms has shown promising
results, especially in the prediction of final outcome as good or bad as well as the
probability of requiring retreatment in future, with the potential for incorporation
of larger multicenter datasets, which will further improve predictive accuracy.[54 ]
Imaging
Point-of-care ultrasound for assessment of cardiac function, volume status, and vasopressor/inotrope
management has witnessed increasing utilization in care of critically ill patients.
Deep learning models have been developed that can enable fast and accurate classification
of cardiac anatomy on echocardiograms.[55 ] Innovations such as these are likely to propel clinicians into a newer era of enhanced
integration of various imaging techniques to generate more accurate diagnosis and
treatment methods. Automated analysis of medical imaging is a prominent area in ML
applications. ML models have been implemented in the reading of radiographic images,
including X-rays and computed tomographic scans, and have reported increasing accuracy
for clinical diagnosis.[56 ]
Within the ICU, models have been developed to provide surveillance for lines and tubes
to assess proper device positioning.57 In addition, waveform analysis from ventilator
data has been used to create models to detect patient–ventilator asynchrony that match
clinical experts.[58 ] Thus far, the primary use of ML waveform analysis has been to either automatically
screen waveforms such as electrocardiograms and electroencephalographs, which cannot
be monitored constantly by clinicians. The goal of these models is to improve time
to early intervention.[59 ]
[60 ] The tools developed for waveform detection could also be used to reduce the burden
of alarms plaguing ICUs. One of the Joint Commission International goals is to reduce
alarm fatigue among care providers, which can conceivably be achieved using modern
AI techniques.[61 ]
Smart Devices
Medication delivery and titration is a key component of patient care in the ICU and
requires a large amount of clinical resources. Smart pumps exist for medication titration,
and these devices can be further utilized for their abilities to provide closed loop
management. In future, increased utilization of closed loop infusions will hopefully
decrease manual labor while possibly enhancing consistency in steady-state drug delivery.
Models using unsupervised learning have been trialed for clinical applications, including
use in vasopressor drug delivery in the ICU.[62 ] Similarly, for tight glycemic control, AI-based artificial pancreas systems have
been developed for use in the ICU.[63 ]
Administration
Triage from emergency departments is a complicated task and includes identification
of high-risk patients who need to be promptly admitted to the ICU. AI models have
been developed that can help triage trauma patients, thereby leading to appropriate
and timely resource utilization.[64 ]
[65 ] Similarly, identification of cohorts of patients with similar clinical needs has
been postulated to provide a framework for future organizational innovations in the
ICU and provide better cost-effective care.[66 ]
Research and Education
Considerable research has been generated in all areas of AI. AI in medical education
is still in its infancy. In the future, it is likely that basic understanding of AI
and its applications will be required in clinical practice and thus will be part of
educational curricula to facilitate better understanding, interpretation, and implementation.
Anesthesiologists and intensivists work at the junction of many disciplines: surgery,
medicine, biology, pharmacology, mathematics, and physics, and are well-placed to
embrace modeling. They have access to knowledge and expertise of enormous breadth
and have experience of a huge array of induced and pathological states, and are comfortable
with biological science, physical science, numbers, technology, and medicine. Anesthesiologists
and intensivists, above all, have clinical contact, a real understanding of real-world
relevancy and empathy for the issues of importance. In addition, they have skills
in managing teams of individuals, collaborating and coordinating their efforts toward
a single goal. All that is required of the researcher who wants to use modeling is
to get an idea of what may be achieved and find a suitable question to answer. Contact
with an expert will be enormously helpful during the researcher’s early forays into
AI.
